# AI-Native Product Surfaces Namespace Instructions This is a compiled namespace source under `pixi-vault/wikis/ai-native-product-surfaces/`. ## Rules - Follow the root `Wiki Compiler Maps/Namespace Wiki Compiler Map.md`. - Treat `Knowledge/` and `Projects/` as canonical authoring sources. - Treat Daily Notes as scratch chronology, not direct compiled content. - Keep this namespace scoped to: AI-native product/application surfaces, problem framing, product demos, PM case studies, job-search edge dashboards, J-Space-Replay, LKY Avatar, I-know-kungfu, planned program intelligence, myAbode, and product-language frameworks. - Do not widen scope silently; propose a namespace promotion/routing update first. - Update `wiki/index.md` and `wiki/log.md` whenever compiled pages are added. --- title: AI-Native Product Surfaces created: 2026-06-16 updated: 2026-07-17 type: namespace-overview status: scaffold category: product namespace: ai-native-product-surfaces confidence: medium --- # AI-Native Product Surfaces > **Definition:** AI is any device that perceives its environment and constraints to take actions that maximize its chance of successfully achieving its goals. > Scaffold namespace for the Pixi Vault AgentWikis compiler. ## Scope ### Covers AI-native product/application surfaces, problem framing, product demos, PM case studies, product-management system-steering frameworks, side-quest validation, AI-era judgment and taste, role-aligned deployed portfolio proof, side-door opportunity and public-proof synthesis, J-Space-Replay, LKY Avatar, I-know-kungfu, planned program intelligence, myAbode, agent output decision artifacts, and product/capability-language frameworks. ### Not Covered Generic product management notes that do not involve AI-native surfaces; local model infrastructure unless relevant to product behavior. ### Current As 2026-07-17 — active namespace. Includes Jamie's top-level AI definition, Product Management as System Steering, Side-Quest Validation Loop, Interaction Mode Routing, Material Loop / Glass Interfaces, Taste Requires Contact, World Model Control Surfaces, Agent Output Decision Artifacts, Role-Aligned Deployed Project Proof, Side Doors: Make Useful Work Legible, Anthropic J-space / Jacobian lens as a product-surface concept, video retrieve-then-verify / verified video answer concepts, compiled product-surface entities and concepts, Job Edge as a live job-search edge/dashboard prototype, Shifu as a local-first searchable video knowledge prototype, J-Space-Replay as a public demo for replaying VLM lens readouts over video, and LKY Avatar as a local fictional-interview stack with tuned voice, animated portrait, executed interaction gates, and a merged audited fact-grounding layer awaiting live operator proof. Cross-format attention and distribution guidance lives in `content-distribution`. ## Canonical Source Roots - `Projects/Job Edge/Index.md` - `Projects/J-Space-Replay/Index.md` - `Projects/LKY Avatar/Index.md` - `Projects/Shifu/Index.md` - `Projects/I-know-kungfu/Index.md` - `Projects/Planned Program Intel/Index.md` - `Projects/Planned/PRD.md` - `Projects/myAbode/Index.md` - `Knowledge/concepts/ai-native-problem-framing-framework.md` - `Knowledge/concepts/interaction-mode-routing.md` - `Knowledge/concepts/j-space-global-workspace.md` - `Knowledge/concepts/material-loop-and-glass-interfaces.md` - `Knowledge/concepts/taste-requires-contact.md` - `Knowledge/concepts/world-model-control-surfaces.md` - `Knowledge/concepts/agent-output-decision-artifacts.md` - `Knowledge/concepts/role-aligned-deployed-project-proof.md` - `Knowledge/concepts/product-management-as-system-steering.md` - `Knowledge/concepts/side-quest-validation-loop.md` - `Knowledge/concepts/video-retrieve-then-verify-loop.md` - `Knowledge/concepts/verified-video-answer-surfaces.md` - `Knowledge/raw/articles/tonbistudio-mini-vss.md` - `Knowledge/raw/articles/nvidia-vss-docs.md` - `Knowledge/raw/transcripts/yann-lecun-world-models-next-ai-revolution.md` - `Knowledge/raw/transcripts/if-you-want-good-taste-you-have-to-eat.md` - `Knowledge/concepts/verb-first-product-positioning.md` - `Knowledge/concepts/find-the-lock-problem-first.md` - `Knowledge/concepts/side-door-opportunity-search.md` - `Knowledge/raw/articles/how-to-enter-side-doors-maja.md` ## Crosslinks - [[../agent-workflows/README|agent-workflows]] - [[../eval-trace/README|eval-trace]] - [[../content-distribution/README|content-distribution]] ## Public Output Contract When published to `pixi-wiki`, this namespace should expose: ```text /raw/ai-native-product-surfaces/README.md /raw/ai-native-product-surfaces/wiki/index.md /wiki/ai-native-product-surfaces/README.md /wiki/ai-native-product-surfaces/wiki/index.md ``` ## Maintenance - Edit canonical source notes first. - Use `Wiki Compiler Maps/Namespace Wiki Compiler Map.md` for routing decisions. - Do not compile Daily Notes directly unless promoted or verified. --- source_url: https://youtu.be/72Xj8k5WQX4?si=tFQOgcbG-xzmz7WI ingested: 2026-06-26 sha256: b9eb1d14ab4133de9902cdc46cbdc6b3410b601d35175b57dab3244dc17f7fd5 source_type: transcript speaker: Yann LeCun --- # World Models: Possibly the Enabler for the Next AI Revolution **Format:** Cleaned Markdown transcript **Source:** User-provided transcript **Note:** Light cleanup applied for punctuation, paragraphing, and obvious speech disfluencies. Meaning preserved. ## Talk **Speaker:** Yeah, I’ll talk about world models, possibly the enabler for the next AI revolution. There are a lot of machine learning people in the room, perhaps. I have bad news for you: machine learning sucks. When we compare the learning abilities of machines with humans and animals, clearly there is a big gap. People and animals can learn new tasks extremely quickly and with very few trials, very few samples. People have common sense. Animals too, physical common sense. There are a lot of tasks that we can accomplish zero-shot, even if we have never faced them before. How do we do this with machines? We have very powerful AI techniques that everybody is using, but they do not really handle the real world. They do not handle continuous, high-dimensional, noisy data. Language is easy by comparison. The real world is messy. Language is simple. This connects with what Vladlen said earlier, and Jitendra as well: Moravec’s paradox. Things that are simple for humans are difficult for computers, and things that are complicated for humans turn out not to be that difficult for computers, like playing chess, computing integrals symbolically, solving equations, proving math theorems, and so on. How is it that a 10-year-old can basically do what you would like a domestic robot to do, and do most of those tasks without actually being trained to do them? The first time you ask them, they can do it. They may not want to do it, but they can. How come any teenager can learn to drive a car in a few hours of practice, yet self-driving car companies have literally millions of hours of training data? Despite that, they cannot use those millions of hours of training data to get a machine, just imitating humans, to drive at the same level of reliability. Otherwise, we would have Level 5 self-driving cars, and we do not. At best in the consumer car business, we have Level 2 or 3. Robo-taxis are heavily engineered with various sensors and other things. So we keep bumping into Moravec’s paradox, and we really have to go beyond this. If you believe that intelligence requires grounding, and some philosophers and certainly some language people do not believe that is necessary, but I think it is, then this matters. Like Vladlen, we are in Switzerland, outside Jean Piaget. He was a big influence on me. Piaget had a debate with Noam Chomsky in France in the late 1970s. They were debating whether language was innate or learned. There were transcriptions of that debate, with people participating in it. One of them was a guy who had worked with Piaget and was a professor at MIT. He was talking about the perceptron, saying that these simple machine learning models were capable of learning surprisingly complex tasks, and that this might be evidence that learning is possible, contrary to what Chomsky was saying. This guy was Seymour Papert. He was a professor at MIT, and 10 years before that he had written a book that basically killed the entire field of neural nets by pointing out the limitations of the perceptron. But here he was 10 years later arguing that those things were actually interesting to study. Piaget is often quoted as saying: “Intelligence is not what you know, it is what you do when you do not know.” In fact, he never actually said this. It is apocryphal. But other psychologists distilled his thinking into this sentence, which he never said. Intelligence is not an accumulation of declarative knowledge. LLMs are an accumulation of declarative knowledge. Not just that, but the main reason they are useful is because they can accumulate a lot of declarative knowledge. Intelligence is not a collection of skills. You can probably build a machine to accomplish almost any task if you spend enough resources on it, including things like self-driving. But that is not really what intelligence is. Intelligence is the ability to learn to drive in about 20 hours, to learn any new task with very little training, or to accomplish new tasks. That is really what intelligence is. That is really what Piaget means. That means we are not going to have any simple measure of intelligence, because any particular task can always be cracked if you spend enough effort and time. It is more about how adaptive you are. This connects to something Vladlen said: the notion of AGI is complete nonsense. Human intelligence is specialized. The characterization of human intelligence is that it is very quickly adaptive, and we can learn new tasks. All of us know different sets of knowledge and have different skills, because we have been exposed to different environments and have had to solve different problems. We are adaptive. That is really what intelligence is. ## How Humans and Animals Learn How do humans learn, and animals for that matter? A lot of learning takes place in the early months of life, mostly by observation. A two-month-old baby can gesticulate and can develop a dynamical model of its own limbs, but basically cannot affect the world. It cannot move an object or anything. But it can learn a lot of things about the world. One thing a baby can learn really quickly is that the world is three-dimensional. Why? Because the fact that every point in the world has a distance from us is the best way to explain how our view of the world changes when we move our head. Babies do not necessarily move their heads, but they are being moved. They see parallax and derive from this the fact that the world is three-dimensional. We can do this with learning machines today. They learn that the world is three-dimensional only by being exposed passively to videos. That is interesting. Basic concepts like object permanence are learned really quickly. Notions of stability, rigidity, and things like that. But what we would consider intuitive physics, things like inertia and gravity, take nine months for human infants. It is shorter for most animals. If you put an eight-month-old or nine-month-old on a high chair and put a bunch of toys in front of them, the child will most likely systematically take all the toys, throw them on the floor, and watch the result. They are doing the experiment that gravity applies to everything. That takes a long time. How does that happen? What type of learning is taking place? They are doing experiments, but they can also learn about gravity by observation. If you show the scenario at the bottom, where a car is on a platform and you push it off the platform, but it appears to float in the air, a six-month-old will barely pay attention. They have not learned about gravity yet. A 10-month-old will be very surprised, like the little girl in the slide. That is how psychologists measure whether a baby has learned a particular concept about the world, through violation of expectation. We can use those techniques to test whether machine learning systems have acquired some notion of common sense. There is a lot that can be said about this. Jitendra and I collaborated on a paper here, mostly written by Emmanuel Dupoux, and Jitendra had very little contribution to it, on this whole set of questions. ## What Intelligence Is What is intelligence really, if it is not an accumulation of skills or declarative knowledge? It is the ability to accomplish new tasks and solve new problems without prior training. Again, AGI makes no sense as a phrase. Human intelligence is specialized. The question is not whether you know how to do everything. The question is whether you can learn quickly how to do anything, or a wide spectrum of things. This is a somewhat philosophical paper at the bottom, written by some of my young colleagues. Here is a simple calculation. There are still a lot of people, particularly on the west coast of the US, who believe that we are going to reach what they call AGI by scaling up LLMs, maybe training them on synthetic data, maybe using a few tricks in post-training and reinforcement learning. I think that is impossible. I am a believer in grounded intelligence. You can do this simple calculation. A typical LLM today is trained on something like 20 trillion words. That corresponds to about 30 trillion tokens. Each token is about three bytes, so the data volume is about 10^14 bytes. This would take about 400,000 years for any human to read. Compare this with what a four-year-old has seen during his or her life. That is about 16 hours of wake time per day, which is a small amount of video, about 30 minutes of YouTube uploads. We have two million optic nerve fibers carrying about one byte per second each. So the data volume that a four-year-old has seen through vision, and probably through touch as well, is about 10^14 bytes. A four-year-old has seen the world through vision with the same amount of data as 400,000 years of text, with all the human-produced text publicly available on the internet. We are not going to get to anything like human-like intelligence by just training on text. It is just not going to happen. Of course, you might say video is much more redundant than text. But that is a feature, not a bug. If you want to train a system, particularly using self-supervised learning, you need redundancy in the data. If you do not have redundancy, you cannot learn anything. Redundancy is a good thing. You do not want too much of it, though. ## Inference by Optimization There is another question about the right properties of intelligent systems. In my opinion, an important property is the mode of inference. Does the system compute its output by propagating through a fixed number of layers of some neural net? Or consider the alternative: computing the output by searching for an output that is most compatible with the input. You observe a situation. That runs through some perception module that produces some representation of the current state of the world as you observe it. You can directly produce an action. That is a reactive system. Or you could imagine an action and have the intelligent system figure out whether it is a good action for this observation. Is this something that will accomplish the task I want? The objective here characterizes whether the task the system wants to accomplish has been accomplished. Think of it as a cost function. It is not used for learning. It is used for inference. Think of it as negative likelihood in a probabilistic inference model, or as I prefer to think of it, an energy function. The inference process is a process by which you search for an output that minimizes some energy function at inference time. That is intrinsically more powerful computationally than just propagation through a fixed number of layers. Contrast the model on the left, which is LLM-like. You take a window of inputs, run it through a fixed number of layers of a big neural net with a few hundred billion parameters, and produce one token. Then you shift that token into the input and produce the second token, and so on. That is autoregressive prediction. Every token involves a fixed amount of computation: running through a fixed number of layers of some neural net. This is not a good model of reasoning. The way you coerce an LLM to do reasoning is that you trick it into generating more tokens. But that is not the way we reason. We reason internally. We do not reason in token space, or even in language. Compare this with the model on the right, which is a slight specialization of the previous one. You perceive the world or your environment. You get some idea of the current state of the world. Then you imagine a sequence of actions, a proposal for an action. You feed it to an internal world model, and the world model predicts the outcome. Then it feeds this outcome to an objective that measures to what extent a task has been accomplished. Then, by optimization, you search for an action sequence that optimizes this objective, or minimizes this energy, at inference time. In my opinion, that is a much more powerful model. But you need a world model. ## A World Model Architecture I settled on this idea or architecture about five years ago. I wrote a long paper about it and put it online in 2022 with some general architecture. If you want to take pictures, here are QR codes. It is relatively easy to read, but long. It is based on the idea that reasoning and planning are essential, and they proceed by energy minimization rather than forward propagation. For this to work, you need a world model. It is the same process I described before, with a few additional tricks. You observe the environment. A perception module produces a representation of the initial state of the world, but only a representation of what you currently perceive. You may have to combine this with the content of a memory to get a complete idea of the state of the world, or at least what you know about it. Then you feed this to your world model, together with a proposal for an action sequence. Your world model predicts the outcome of that action sequence. You feed this to an objective, an energy function, that measures to what extent a particular task has been accomplished. This function outputs zero if the task is accomplished, and some positive number if the task is not accomplished. Perhaps it measures some distance to the task being accomplished. You can have another set of objectives that are guardrails. They ensure that whatever state sequences the system is going to take the world through will not kill anyone, hurt anyone, or have any deleterious effect. A system constructed this way can be made intrinsically safe because it has to obey and optimize the guardrail objective with every output it produces. This is not the case for an LLM. The only way an LLM can be made safe or non-toxic is by fine-tuning it. There is always a way to break the conditioning, or jailbreak the system. Here, you cannot jailbreak a system like this. It can do nothing but optimize the guardrail objectives and the task objective. If you have a world model, there are certainly a lot of roboticists and optimal control people in the room, you can apply this world model multiple time steps. Each action sequence can be decomposed into a sequence. The guardrails can be applied to all the steps in the sequence. That is the way you would use a world model. The way you plan by optimization is akin to model predictive control, MPC, very classical stuff in optimal control going back to the 1960s. ## Hierarchical Planning Ultimately, what you want is something that can do hierarchical planning. All of us do hierarchical planning. Animals do hierarchical planning. What is hierarchical planning? Suppose I am sitting in my office at NYU and I want to be in Paris tomorrow. There is no way I can plan my entire trip to Paris in terms of muscle actions 10 milliseconds by 10 milliseconds, which are the elementary actions that humans can do. I cannot do that because, first of all, it is too long. Second, I do not have the information. I do not know how long I will have to wait on the street before a taxi stops. There is no way I can plan the entire thing. I have to do hierarchical planning. At a high level, I can say: I do not know how long it will take me to go to the airport, but maybe roughly an hour or an hour and a half. So I need to get to the airport and catch a plane. That is a two-step high-level plan. I do not need to know many details to make that plan. Now I have a subgoal: being at the airport. I am in New York, so going to the airport involves going down to the street, hailing a taxi, and going to the airport. Now I need to go down to the street. I am in an NYU building, so that involves walking to the elevator, pushing the button, getting down, and walking out the door. Now I have a subgoal: getting to the elevator. You can go down this hierarchy. At some point, you get to a point where the action you need to take is very simple. It is something you are familiar with. You may not have to use your full mental power to plan the action. You can probably stand up from your chair without having to think about it. That could just be a policy. Ultimately, we want systems to do hierarchical planning. How do we solve that? This is an unsolved problem. If you are a roboticist, or an AI for robotics person, or an agentic AI person starting a PhD on this topic, this is a great topic. It is completely open. Nobody knows how to do this, or nobody has proved that they know how to do this. ## Training World Models Now the big question is: how are we going to train those models? Hierarchical or not, let us start with non-hierarchical. First we have to figure out what architecture to give them. A natural instinct these days is to train a generative model. In fact, I have been working on trying to train world-model-like things for about 15 years, mostly failing for the first 10, because I was trying to train generative models. What is a generative model? Self-supervised learning has been incredibly successful in the context of language. You take a string of words, remove some of the words, corrupt the input, run the corrupted input through a big neural net, and train it to recover the missing parts. That works amazingly well for text. The original models like BERT used to do this. An LLM is a special case where the only word you remove is the last one, so the entire system is trying to produce the next word in a sequence. It works amazingly well and it scales if you do it right. It does not work if you apply it to video. If you take a video and show the initial segment of the video to the system, then ask it to predict what will happen next at a pixel level, it does not really work. The representations you get from the system for your video are not particularly good. The reason is that you simply cannot predict everything that takes place in a video. There is an infinite number of plausible things. In text, it is easy because there is only a finite number of words. You can get the system to produce a probability distribution over all possible words or tokens in your dictionary. You cannot do this with video. There is an incredibly large number of possible video frames. Take an example. If I take a video of this room, start here, slowly rotate the camera, stop here, and ask the system to continue the video, it is probably going to predict that we are in some sort of classroom or auditorium, that the room has a finite size, that there might be windows on this side, and things like that. There is absolutely no way the system can predict what all of you look like, or which chairs are unoccupied. It is impossible. You just do not have the information. So when you train a system to make this kind of prediction, you kill it. Of course, you are going to tell me: but we can train systems to produce cute videos. Video generation, yes. But this prediction is usually done in representation space, not pixel space. It is only a second stage that turns the predictions into high-resolution, high-frame-rate videos. The system only needs to produce one cute-looking video. It does not need to actually represent all plausible videos. That is a much simpler problem. As I said, I have been working on this for the better part of the last 15 years. Here is a 10-year-old paper where we tried to train a neural net to predict short video clips, two frames from four frames of context. You get blurry predictions. Why? Because the system predicts the average of everything that can happen. Of course, you can correct that with latent variable models like diffusion models, which we did not know at the time. We tried to use GANs and things like that, but were not too successful. Perhaps using latent variable models would help, diffusion models in particular, which of course produce cute videos. Do they actually understand the world? The evidence is no. ## Joint Embedding Predictive Architecture Here is my solution: an architecture called joint embedding, or more precisely Joint Embedding Predictive Architecture, JEPA, shown on the right. On the left you have a generative architecture. You observe X, maybe you observe A, an action that is taking place, and you observe the result Y. The system is trying to reconstruct Y in its most minute details. With JEPA, you observe X, Y, and A, but you encode both X and Y, and prediction takes place in that representation space. That is a major difference. The system can essentially eliminate from the input, by constructing a representation of Y, all the information about Y that is simply not predictable. That makes the prediction more abstract, with fewer details, but more accurate in a way. How do you train a generative model? It is easy because the cost is just a reconstruction cost. You train it to reconstruct. You can train it as an autoencoder, but then you need to restrict the information content in the code, or as a denoising autoencoder, which is what a lot of techniques like masked autoencoders have attempted to do. That means taking an input, corrupting it in some way, and training an autoencoder to recover the initial one. Diffusion models are a special case of this general idea of denoising. The bad news is that when you train systems of this type to learn representations of images, you do not get good representations. If you use the representations of images obtained this way and feed them to a downstream task that you train supervised, the results are not great. To get good results, you have to use joint embedding architectures. All the best systems that use self-supervised learning to train image or video representation systems use joint embedding. None of them uses reconstruction. For images, either you have two views of the same scene and you train a neural net to produce representations, telling the system you want those two representations to be identical, or you use the corruption technique. You take an input, corrupt it or transform it in some way, and train this JEPA architecture to predict the representation of the original image from the representation of the corrupted version. There is a big issue: the system can collapse. Generative models can also collapse to some extent. If you train an autoencoder without a restriction on the information content of the code, your autoencoder learns the identity function. That is a collapse. It is not going to learn anything useful. Similarly, a system like this can collapse by completely ignoring the inputs and producing constant representations. Then the prediction problem is trivial. If you just train a system of this type to minimize prediction error, it is going to collapse. It is not going to do anything useful. The whole trick of self-supervised learning for joint embedding systems is how you prevent collapse. My favorite concept for preventing collapse is information maximization. You come up with an objective function that measures some sort of information content of the representation that comes out of your encoders, and you try to maximize that information content. Your cost function is minus the information, or something like that. There are a bunch of techniques for this from the last six or seven years, with names like MCR, MCR^2, VICReg, VICRegL, and Barlow Twins. Barlow Twins and VICReg come from people working with me. The other ones are from other groups. MCR comes from Berkeley, and MCR^2 from a colleague at NYU in neuroscience. This idea of JEPA is gaining popularity. There are about 1,700 papers that mention “joint embedding predictive architecture” spelled out on Google Scholar. ## Measuring Information Content There is an issue with this type of method: how do you measure information content? We need a cost function that is a differentiable measure of information content so we can backpropagate gradients and maximize it. The bad news is that, first, we do not have objective measures of information content, because all the proper definitions are based on knowing the distribution of the vectors, or whatever you want to measure the information content of. We do not know the distribution. We only have samples coming out of an encoder. How do you compute information content from a finite number of samples? That is the first problem. The second problem is that, to maximize something, you would need a lower bound on information content, so that when you maximize, you push the actual information content up. The problem is that every empirical measure we have is an upper bound. So what do we do? We come up with a good upper bound, cross our fingers, show some theorems, and so on. ## Energy-Based Models The way to properly explain how you can train self-supervised learning systems, and every learning system really, is a framework I call energy-based models. I have been advocating for this for 20 years or so. The basic idea is this: if you want to capture the dependency between two variables X and Y, and there is no real functional relationship between X and Y, meaning there is no single Y for a given X, only a dependency, then you cannot run a function that computes Y from X. This is indicated by the diagram on the right. You have a bunch of data points, the black dots, and they indicate some sort of dependency between X and Y. How do you capture this dependency, given that you cannot run a function that computes Y from X? One way is to learn or build a contrast function, an energy function, that tells you whether a point in this XY space is near the training data or not. Think of it as a landscape where the black dots are in the valley. In Switzerland, there would be a lake. Then you get level curves. As you move outside of those regions, the altitude goes up. The energy goes up. Now, if I give you a value for X, you can infer a bunch of values for Y that are compatible with X. There are values of Y that minimize the energy. It is the kind of inference I was talking about earlier: inference by optimization, not by forward propagation. You can possibly do it the other way around. If I give you Y, you can infer X from Y and give me multiple answers. In situations like video prediction, where there is basically an infinite number of possible answers, the proper way to train a system of this type is to think of it in terms of energy-based models. Probabilistic models are a special case where your energy has a particular form and the way you train it has a particular loss function. This is a slightly more general framework than probabilistic inference and learning. To train an energy-based model, you have to prevent collapse. The collapse problem I was telling you about before would be manifested by the energy function being flat everywhere. You train the system to minimize energy for a bunch of training samples, and the system gives you an energy function that is zero everywhere. That is what an autoencoder that learns the identity function does. That is what a JEPA that ignores the input and produces constant representation does. It has zero prediction error for everything. So it is a collapse. To prevent collapse, you need one of two things. One is contrastive methods. You generate points outside the region of data and push the energy up. You come up with a cost function that makes sure the energy of the data points comes down and the energy of other points is higher. There is another set of methods, which I have come to prefer: regularized methods. They work by minimizing the volume of space that can take low energy. If you push down the energy of certain regions, the rest has to go up because there is only a small volume of low energy to go around. In practice, this reduces to one of those two methods. ## Information Maximization Let us go back to this idea of information maximization. Suppose I run a batch of samples through one of the encoders. I get a matrix where each row is the representation for one sample, and each column is the value of one variable in the representation for all samples. There are two ways to make that matrix informative. One way is to make sure all the rows are different. Another way is to make sure all the columns are different. You want to make sure the columns are different because if all the columns are the same, every variable in the representation carries the same information. That is not very informative. You want each variable in the representation to be maximally disentangled from the others, to give independent information from the other variables. That is an example of what we can call dimension-contrastive methods, which are a form of regularized method. At the bottom, the type of criterion that makes the rows all different corresponds to contrastive methods, or sample-contrastive methods. Sample-contrastive methods are very popular for certain applications. A lot of the perceptual pipelines in LLMs are trained with a technique called CLIP, which is basically a contrastive method that does joint embedding between images and text. But I prefer the other one. ## Abstraction and Prediction The idea that you need to find an abstract representation of an input to make prediction is very natural. We do this all the time as humans. We do this all the time as scientists and engineers. Animals do it too. In principle, I could explain or simulate everything taking place in this room at the level of quantum field theory or particle physics. I could simulate the trajectory of every particle in this room, going all the way down to simulating all of our brain processes and everything. In principle, running the simulation, I could figure out whether any of you actually understand the words I am saying, whether you are sleeping, or whether you are totally bored. Of course, that is completely impractical. What we do in science is invent abstractions that allow us to make predictions. Those abstractions ignore a lot of details about the underlying state of the system. We invent abstractions from quantum fields to particles, atoms, molecules, proteins, organelles, cells, organisms, individuals, societies, and ecosystems. Every level in this hierarchy is a particular level of abstraction with which we describe the world. It allows us to make longer-range predictions than the levels below by ignoring many details of the lower level. That is why the way to understand what is going on in this room at the moment is more at the level of psychology than at the level of particle physics. Of course, physicists always make fun of everyone by saying that everything is just applied physics. Even psychology is applied physics to some extent. But in fact, there is specific knowledge about chemistry that does not derive directly from physics. This abstraction contains new knowledge, information, or structure that was not apparent at the level below. This idea of JEPA constructs on the concept that you need to find an abstraction to be able to make predictions. Suppose you want to design an airplane. You need to design the airfoils for the airplane, so you do computational fluid dynamics. You simulate the flow of air around the wing. You model the state of the air in every little cube around the wing by velocity, density, and so on, and then you solve Navier-Stokes partial differential equations. That simulates the flow of air. But it ignores a huge amount of detail in the underlying mechanism. The underlying mechanism is molecules of air bumping into each other and bumping into the plane. You never simulate fluids at that level. It is too complicated, and it would diverge from reality really quickly because it has too many details. You have to ignore details to be able to make accurate long-term predictions. We do this in science all the time. World models should not be simulators. They should work in abstract space. They should not be digital twins, which is a buzz phrase. They should definitely not be generative models, as I just explained. They should not be video generation systems. A lot of people are working on video generation and calling this world models. They are not world models. They are video generation systems. One big message from my talk is: if you want to use world models, do not work on video generation. That is a different problem. If you want to produce cute videos, work on video generation. But if you want to control robots or industrial processes or understand the world, do not work on generation. You want models to control complex systems where you cannot model the dynamics by writing a bunch of equations. If you have a humanoid robot, or any kind of robot, you can write down the dynamical equations and simulate the dynamics of the robot. You can get your humanoid robot to do somersaults and kung fu and whatever. That is simple. As soon as a robot starts to interact with the real world, it becomes much more complicated. That is more difficult to reduce to simple equations. Think about a complex system like a turbojet, a chemical plant, a patient, or a robot that interacts with the real world in complex ways. You cannot reduce this to a small number of equations. You have to learn a phenomenological model of the whole system, the system you control and its interaction with the environment, so you can make predictions and plan a sequence of actions to arrive at a particular outcome. That is a world model. The concept is very old. It goes back to the 1960s and is the root of optimal control. ## SIGReg Now I come to a particular technique that I am very fond of, and that I think we will expand over the next few months and years to do this information maximization. It is called SIGReg: Sliced Isotropic Gaussian Regularization. The trick is the following. You run a batch of samples through your encoders, and you get a bunch of points in a vector space, with dimension equal to the dimension of your representation space. We are going to try to make the distribution of those points isotropic Gaussian, with the same variance in all dimensions. Why? Because an isotropic Gaussian is a distribution where all variables are independent. They are maximally informative individually. It is also the distribution that has maximum entropy for a given variance, but we do not really care about that. What is interesting is that it makes the variables independent of each other. How do we do this? Of course, we do not have the distribution. We just have a bunch of points. It may be a high-dimensional space, like 2,000 dimensions, and we may have a few hundred or a few thousand points. How can we make sure this is a Gaussian? Here is the trick. You project the individual points along a single direction, and what you get is a marginal distribution. Of course, you still have discrete points. You do not have a continuous density. You have discrete points. One trick is to compute the cumulative distribution that those points give you. It is a staircase because you have discrete points in one dimension. Then you can ask: what is the distance between the staircase, the empirical cumulative distribution of my points, and the cumulative distribution of a Gaussian? You can do that because you know what the Gaussian looks like. For every point on the staircase, you can tell whether it is to the left or to the right of the ideal Gaussian. That gives you a gradient: do I move the point this way or that way in that projection? It gives you a gradient for every training sample in your batch. If you make the distribution Gaussian along that projection by gradient descent, it makes the marginal distribution Gaussian along this projection. There is a theorem that says if you do this along lots and lots of directions, then in the limit your joint distribution is actually isotropic Gaussian. What we need to do is many projections. For all of those projections, compute those gradients, move the points, or backpropagate through the network and change the weights, so that the points move and the overall distribution becomes more Gaussian. If you apply this to a distribution like the one on the top left, an X shape in two dimensions among 1024, and do gradient descent, you move the points. You do not train a neural net in that example. The technique I am advocating for gives you something that is sort of Gaussian-ish. This really works in practice. We have applied it to training world models that are action-conditioned, and we have used them for planning. It works decently. The source code is available. It is very simple. You can train it on one GPU. What we need to do with this technique is scale it up. There are a few other things we need to do, but that is the main one. In simple cases, you can train this world model and use it to plan simple actions, as in Push-T or simple robotic situations in simulated environments. That needs to be scaled up, but it is good work. There is a theoretical paper that we put out just a few days ago. If you make the hypothesis that the underlying distribution of your data is actually an isotropic Gaussian, and assume that the observations you get from the world are some complicated nonlinear transformation of those points, like a spiral transformation, then if you train a neural net with SIGReg on it, it will recover the original Gaussian in representation space. It is not a general proof that it works in every case. But it is a proof that if your original explanatory variables are Gaussian, the system will recover those variables up to a rotation. ## Distillation Methods: I-JEPA, V-JEPA, and DINO We can use these techniques in the context of self-supervised learning to train an image recognition system. There is another set of techniques that I should mention because they work really well and are the ones that have been scaled up so far. SIGReg is conceptually my favorite method, but it is very recent and we have not scaled it up yet. These other methods are based on distillation, and we have scaled them up and obtained really good results for both images and video, with techniques like I-JEPA and V-JEPA. What is the basic idea of those distillation methods? You still have two encoders. So it is a JEPA architecture. You take an input, transform it, corrupt it, or mask it, and then train the system to predict in representation space. But you do not propagate gradients through the encoder on the right. Those are two encoders with identical architectures, and they kind of share the weights. The encoder on the right uses an exponential moving average over time of the weights of the encoder on the left. The encoder on the left gets gradients and gets updated all the time. The encoder on the right gets updated more slowly and shares the weights. This is derived from intuitive ideas by people at Google DeepMind who were using techniques like this to stabilize the variance in reinforcement learning. They realized you could apply this to self-supervised learning from images. They called this BYOL, Bootstrap Your Own Latent. There are a whole bunch of methods from Meta in particular, such as SwAV and others, that use this exponential moving average idea. A particular method called I-JEPA, which I show here, produces really good results. With I-JEPA, we were able to compare results against a generative approach called MAE, masked autoencoder. I-JEPA is not only better, it is much faster to train. Another technique is called DINO. Many of you, I am sure, have heard of it. I know some of you have used it because there were projects in the robot demos that used DINO. This was done by some of our former colleagues at Meta in Paris. It is completely self-supervised. It is a joint embedding architecture. It uses distillation, with various tricks I am not going to explain. There is a lot of engineering behind it. These systems basically, at this time, produce the best generic representations of images. If you have any type of vision task you want to do, this is probably the best encoder for images. Among other things, what we have done is use DINO as an encoder, train a world model, and do planning. ## Planning Demo Let me show you a cute video. You have an initial state of a simulated environment with pretty complex dynamics. You have goals at the top. At the bottom, what you see is the sequence of actions of a planner that uses this trained world model to get the world to a configuration as close as possible to the original one in less than 25 steps. This has been applied to a number of different scenarios, like double pendulum, Push-T, and others. It now works really well. ## V-JEPA and Common Sense More recently, we applied this to video. You take a video, mask a big chunk of it, and train the JEPA to produce good representations, so that it can predict the representation of a full video from the representation of a partially masked one. Once the system is trained, you use the encoder as a way to extract features from the video, and you train a head on top of it to accomplish some task. It works really well, state of the art for many traditional vision tasks, particularly from video: action recognition, action prediction, and so on. Instead of boring you with a table of results, one interesting thing I want to mention is that V-JEPA has learned some level of common sense. Because we train it to predict what will happen next in a video, we can train a predictor to do that and measure its internal prediction error. We can show it a video and monitor the internal prediction error at every time step. This system takes a window of 16 frames. We slide those frames over a video and measure the prediction error for the next 15 or 16 frames. The cool thing is that if you show it a video where something impossible happens, something unphysical, the prediction error shoots to the roof. It is like the little girl in one of the early slides looking at the scene of the car not falling. If you have a video of a ball being thrown and the ball disappears, the prediction error shoots through the roof. That is interesting because it is the first time, at least from my point of view, that I have seen a completely self-supervised system acquire some level of common sense, telling you what is possible and what is not possible. Let me skip this. It is cute, but no. It just says V-JEPA can be used for planning. New versions of this do a better job at planning and everything. Here is an interesting thing. Remember I told you that the way babies learn the world is three-dimensional is because it is the best way to explain how your view of the world changes when you move your head. We took the representation learned by some version of V-JEPA called V-JEPA 2.1, and then trained a head on top of it to predict depth from a single image. It does a really good job. It produces really good results, in fact better than DINOv3. That shows that this system, by just being trained to fill in the blanks in videos at a representation level, basically understands, in double quotes, that the world is three-dimensional. It understands the notion of object. If you use the representation as input to a segmentation system, it works decently well, and for various other things. ## Conclusion Let me conclude. Abandon generative models. I mean if you work on LLMs, of course, but you should not work on LLMs. At least if you are in academia, you should absolutely not work on LLMs. There is nothing you can bring to the table. Abandon generative models in favor of joint embedding architectures if you are interested in intelligence. Abandon probabilistic models in favor of energy-based models. I did not have time to really explain why. I made an argument in favor of regularized methods, or information maximization through variables instead of samples. So, dimension-contrastive methods rather than sample-contrastive methods, though sample-contrastive methods have many practical applications. I have been saying forever to abandon reinforcement learning. I do not really mean abandon it. I mean minimize its use because it is so horribly inefficient in terms of sample efficiency. I know there are people here who work on this, but reinforcement learning is what you do when you are desperate and there is nothing else you can do. You have to do most of the learning by observation, learning world models, and so on. Once you have good representations, you can use reinforcement learning on top of them because you already have good representations and will not require too many samples. Sometimes you cannot avoid it. If you are interested in making real progress in AI, in grounded AI, AI for the real world, or physical AI, do not work on LLMs. Do not work on generative models either. As you can probably guess, this does not make me very popular in Silicon Valley. As many of you probably know, I left Meta at the end of last year and formed a new company, heard in the transcript as “AMI Labs.” Its purpose is AI for the real world, physical AI. Robotics is a use case, but it is not just that. It is control of industrial processes, anything that is high-dimensional, continuous, and noisy, for which LLMs are completely helpless. That is the kind of problem we are working on. That is it. Thank you very much. *[Applause]* ## Q&A **Moderator:** Okay. I know there are many questions. Maybe we will take one or two, but then we have to wrap up. Please, quick questions and quick answers. **Audience question:** Thanks for the talk. I wanted to ask about the guardrails that you mentioned on one of the earlier slides, where you also talked about MPC. Engineers love MPC because they can put in their constraints and describe them in state space, like 3D space. But from what I understand, in your system everything works in representation space. How do I even get a constraint like “do not bump into the wall” into this representation space? Do you envision the system learning the constraints by itself, or can engineers really put them in? **Speaker:** No, you would have to learn a very small head on top of your representation that maps that to the constraint you are interested in. That part has to be trained, but you can train it with a very small number of samples because it is tiny, basically just a projection. **Audience question:** But you need a different encoder for each kind of constraint that you might want to put in? **Speaker:** Well, you need a different projector for each constraint. If your task is to open a door, I am not talking about a constraint, I am talking about a task objective. You need some cost function to tell you: is the door open or not? That might have to be trained when you train the system to accomplish the task. But basically, that requires two samples. **Audience:** All right. Thanks. **Moderator:** Okay. I think we will have to leave it here. Thank you, Yann, very much. **Speaker:** All right. Thank you. *[Applause]* --- title: Agent Output Decision Artifacts created: 2026-06-27 updated: 2026-06-27 type: concept status: compiled namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, ai-native, product-framing, interaction-design, agent-workflows] sources: - Knowledge/concepts/agent-output-decision-artifacts.md - Knowledge/concepts/interaction-mode-routing.md - Knowledge/concepts/material-loop-and-glass-interfaces.md - Knowledge/concepts/visual-plan-review-surfaces.md confidence: medium --- # Agent Output Decision Artifacts **Agent Output Decision Artifacts** are concise, visual, interactive surfaces that convert verbose AI-agent output into an inspectable decision. The artifact is not a generic summary. It is a control surface: preserve what matters for the decision, expose evidence where trust is needed, and make the next action obvious. ## Core pattern 1. **Compress** the agent run into decision-critical meaning. 2. **Structure** it around three visible elements: options, risks, actions, evidence cards, or tradeoffs. 3. **Close the loop** with approve, reject, annotate, choose, or steer-back controls. ## Artifact contract A good decision artifact: - fits on one screen by default; - removes repetition, caveats, filler, and process narration; - uses simple clear sentences; - uses cards, tables, hierarchy, or diagrams when they carry structure faster than prose; - keeps details behind progressive disclosure; - preserves source anchors or expandable evidence; - makes the next action visible within 5-10 seconds. ## Why it matters Verbose chat is useful while an agent is working. It is a weak final surface when the user needs to decide. Decision artifacts let chat remain the command channel while the review/approval work moves into a surface that is easier to scan, compare, and steer. ## Relationship to existing lenses - [[interaction-mode-routing|Interaction Mode Routing]] decides when chat should give way to generated review/control surfaces. - [[material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] explains why judgment-bearing work needs to stay visible, steerable, interruptible, and traceable. - [[../../agent-workflows/wiki/concepts/visual-plan-review-surfaces|Visual Plan Review Surfaces]] is a workflow-specific subtype for turning PRDs and implementation plans into inspectable MDX/HTML review artifacts. - [[world-model-control-surfaces|World Model Control Surfaces]] gives the control-loop shape: show state, actions, predicted outcomes, objectives, guardrails, evidence, and next safe step. ## Example surfaces | Surface | Compresses | User action | |---|---|---| | Ideation artifact | long brainstorm or agent proposal | choose, reject, or validate one direction | | PR review artifact | code-review run and verification output | merge, block, or request changes | | Research artifact | source-heavy investigation | accept answer, ask follow-up, or inspect evidence | | Compile/release artifact | source changes and generated outputs | approve publish or fix route gaps | ## Boundaries - Not every agent output deserves a generated artifact; sometimes a table, diff, or short answer is enough. - The artifact should point back to stable truth rather than becoming the only source of truth. - Do not hide uncertainty; make it visible through confidence, evidence, open questions, or expandable detail. ## Source Compiled from `Knowledge/concepts/agent-output-decision-artifacts.md` and adjacent Pixi Wiki concepts. --- title: AI-Native Problem Framing Framework created: 2026-06-16 updated: 2026-07-12 type: concept status: compiled namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, ai-native, product-framing, product-management] sources: - Knowledge/concepts/ai-native-problem-framing-framework.md - Knowledge/concepts/material-loop-and-glass-interfaces.md - Knowledge/concepts/world-model-control-surfaces.md confidence: high --- # AI-Native Problem Framing Framework The **AI-Native Problem Framing Framework** is the reusable lens for deciding whether a product surface is genuinely AI-native or merely has AI attached. ## Core frame Define the system before picking models: - **Environment** — what data describes the world? - **Actions** — what can the system do? - **Goal** — what is success or what is optimized? - **Constraints** — what must never be violated? - **Agency constraints** — what must remain visible, steerable, inspectable, interruptible, or user-owned? Bad framing creates bad AI. Environment, action space, objective, constraints, and agency boundaries define the intelligence problem. [[world-model-control-surfaces|World Model Control Surfaces]] extend the frame into a planning/review loop: observed state -> candidate actions -> predicted outcomes -> objective/guardrail score -> recommended next safe step. Use it when an AI-native surface needs to expose what the system expects to happen before a human or agent acts. Agency constraints come from [[material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]]: AI can shorten the path from idea to artifact, but the interface should not hide the judgment-bearing parts of the loop. ## Interface mode after framing After the environment/actions/goal/constraints frame is clear, use [[interaction-mode-routing|Interaction Mode Routing]] to choose which parts should be direct UI, agentic delegation, generative UI, or stable truth/routing. This keeps AI-native product work from collapsing into either chatbot theatre or agentic overreach. The interface should preserve provenance, constraints, and human control where the domain requires them. ## Product-surface use For `ai-native-product-surfaces`, this framework prevents vague “add AI” product thinking. It asks whether the surface perceives a domain, chooses or prepares actions, improves the chance of achieving a goal, and respects hard constraints. It is especially useful for comparing: - Planned Program Intel: event-program decision routing and institutional memory; - myAbode: real-estate prepared next actions under compliance and adoption constraints; - future surfaces that need prediction, optimization, and execution separated rather than collapsed into a black box. After framing the AI system, use [[product-management-as-system-steering|Product Management as System Steering]] to make the human product calls around decision tempo, ownership seams, stakeholder dynamics, scope, and ecosystem incentives. ## Boundary Do not blindly copy game/RL patterns into product domains. Real-world operational products have partial visibility, noisy outcomes, multiple stakeholders, and constraints that must be represented explicitly. ## Related pages - [[interaction-mode-routing|Interaction Mode Routing]] - [[material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] - [[world-model-control-surfaces|World Model Control Surfaces]] - [[product-management-as-system-steering|Product Management as System Steering]] ## Source Compiled from `Knowledge/concepts/ai-native-problem-framing-framework.md`. --- title: Interaction Mode Routing created: 2026-06-23 updated: 2026-07-13 type: concept status: compiled namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, ai-native, product-framing, interaction-design, generative-ui] sources: - Knowledge/concepts/interaction-mode-routing.md - Knowledge/concepts/material-loop-and-glass-interfaces.md - Knowledge/concepts/taste-requires-contact.md - Knowledge/concepts/ai-native-problem-framing-framework.md - Projects/I-know-kungfu/Index.md - Projects/Hermes Mission Control/Index.md confidence: medium --- # Interaction Mode Routing **Interaction Mode Routing** is a product refactor lens for choosing whether a task should use direct UI, agentic delegation, generative UI, or stable truth/routing surfaces. The question is not "should this use AI?" The question is: **what interaction mode best fits the user's need for speed, control, exploration, inspection, and execution?** ## Four modes | Mode | Use when | Failure smell | |---|---|---| | Direct UI | The human can act faster by manipulating visible objects than by describing the action. | Replacing a faster button, slider, table, or visual control with a slower chatbot. | | Agentic delegation | The user wants an outcome across repetitive or cross-tool work, not every step. | Hiding judgment, provenance, or approval behind autonomous action. | | Generative UI | The user needs to compare, inspect, tune, approve, or understand a middle-complexity task. | A prompt box that only feeds a fixed grid, or a generated surface that hides source/constraints. | | Stable truth/routing | The surface is durable source truth, provenance, or routing. | Generated UI becoming the only place a decision, source, or constraint exists. | ## Product-surface use Use [[ai-native-problem-framing-framework|AI-Native Problem Framing Framework]] first to define environment, actions, goal, constraints, and agency constraints. Then choose the interaction mode: - direct manipulation for fast, visual, precise work; - agents for boring multi-step execution; - generative UI for review/control surfaces; - stable truth for PRDs, project hubs, GitHub issues, handoffs, skills, MOCs, `llms.txt`, `index.json`, raw Markdown, and MCP entrypoints. Use [[material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] as the material closeness test: if the work carries taste, trust, scope, provenance, architecture, or release responsibility, keep it visible and steerable instead of collapsing it into black-box output. Use [[taste-requires-contact|Taste Requires Contact]] as the learning-friction test: when firsthand use, diagnosis, imitation, preference formation, or final selection is how the user develops judgment, assist that act without delegating it away. ## I-know-kungfu refactor For [[../entities/i-know-kungfu|I-know-kungfu]], the strongest next surface is a generated **Fit Check Surface** over stable wiki truth: - source coverage; - local overlap; - recommended serving entrypoint; - refusal boundaries; - trust/eval state; - clearly labeled evidence provenance. This keeps the product out of two traps: pure chatbot and static card pile. ## Pixi Wiki / vault refactor Pixi Wiki should keep source/navigation stable and generate temporary review surfaces above the corpus: - compile-review report; - source-to-output trace map; - namespace coverage dashboard; - MCP/live-route visibility report; - broken/stale route repair panel. ## Agent workflow connection For Hermes Mission Control, chat is the command channel, not the whole interface. Review and approval should be shaped as small generated control surfaces while durable truth remains in GitHub, Obsidian, handoffs, skills, and knowledge entrypoints. For verbose agent output, use [[agent-output-decision-artifacts|Agent Output Decision Artifacts]] when the user needs a decision, approval, comparison, or steering surface instead of a wall of chat prose. ## Source Compiled from `Knowledge/concepts/interaction-mode-routing.md` plus project applications in I-know-kungfu, Pixi Wiki, and Hermes Mission Control. --- title: J-Space as Global Workspace created: 2026-07-08 updated: 2026-07-08 type: concept status: active namespace: ai-native-product-surfaces source: Knowledge/concepts/j-space-global-workspace.md confidence: medium --- # J-Space as Global Workspace J-space is Anthropic's name for the model-internal subspace spanned by Jacobian lens vectors: directions in a language model's residual stream that make a token more likely to be verbalized now or later. The paper argues that this verbalizable subspace behaves like a **functional global workspace**: content written there can be reported, modulated, used in flexible reasoning, and broadcast to downstream computations. This is a functional claim about accessible representations, not a claim about subjective consciousness. ## Product translation ```text model activation at a position/layer -> project onto J-lens directions -> read token/concept directions with high activation -> inspect what is available for verbal report or downstream use -> optionally intervene to test whether the representation is causal ``` The useful product idea is not "read model thoughts." The useful idea is: **replay what a model is poised to verbalize so humans can inspect intermediate, silent, or weakly expressed content before trusting the answer.** ## What the paper claims Anthropic reports that J-space has the cluster of properties associated with global workspace theory: - **Verbal report:** J-lens readouts track what the model will say, and interventions can change reports. - **Directed modulation:** instructed mental content can appear in J-lens readouts even while the surface output is unrelated. - **Internal reasoning:** unspoken intermediates can appear in J-space and be causally load-bearing. - **Flexible generalization / broadcast:** one J-space vector can feed many downstream functions. - **Selective mediation:** flexible explicit tasks depend on J-space more than routine automatic processing. - **Structural support:** model components appear arranged to read, write, amplify, and broadcast J-space content. ## Alignment-auditing relevance The paper applies J-lens readouts to cases where silent strategic or situational assessments appear in J-space before or without appearing in the model output. This makes J-space a promising review surface for evaluation awareness, prompt-injection recognition, self-monitoring, or hidden objective signatures. The boundary is just as important: the authors do **not** claim J-space monitoring is sufficient for alignment. Automatic or highly practiced behavior can route outside J-space, and single-token limitations can hide multi-token or diffuse concepts. ## Relationship to J-Space-Replay J-Space-Replay borrows the interaction shape: it lets users replay Qwen2.5-VL logit-lens and fitted J-lens readouts over a video-answer timeline. Keep the public claim honest: - the Anthropic paper validates text-model J-space behavior, not VLM video behavior; - J-Space-Replay is a demo-quality glass-box surface, not a full replication; - the app's preset traces and UI are useful for inspecting readouts, but upload/local GPU use is required for new videos; - the honest claim is "inspect VLM readouts," not "read model thoughts." ## Source handles - Anthropic Transformer Circuits paper: https://transformer-circuits.pub/2026/workspace/index.html - Canonical Knowledge page: `Knowledge/concepts/j-space-global-workspace.md` - Source digest: `Knowledge/raw/papers/anthropic-verbalizable-representations-global-workspace-2026.md` - Related entity: [[../entities/j-space-replay|J-Space-Replay]] --- title: Material Loop and Glass Interfaces created: 2026-06-24 updated: 2026-07-13 type: concept status: compiled namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, ai-native, product-framing, interaction-design, generative-ui] sources: - Knowledge/concepts/material-loop-and-glass-interfaces.md - Knowledge/concepts/interaction-mode-routing.md - Knowledge/concepts/ai-native-problem-framing-framework.md - Knowledge/concepts/taste-requires-contact.md - https://youtu.be/az6OEZV8iHw confidence: medium --- # Material Loop and Glass Interfaces **Material Loop and Glass Interfaces** is the agency/authorship lens for AI-native product surfaces. The **Material Loop** is the cycle where a person turns an idea into a visible artifact, inspects what feels wrong, changes it, and develops judgment through contact with the work. [[taste-requires-contact|Taste Requires Contact]] adds the input side: judgment also depends on firsthand contact with the work a person consumes. Reference gathering is not enough; the person must use, notice, name, imitate, compare, curate, and sometimes depart from what they encountered. A **Glass Interface** keeps AI-shaped work inspectable, steerable, interruptible, and traceable. It exposes enough plan, state, sources, tools, diffs, commands, constraints, and intermediate artifacts for the user to stay close to the material without micromanaging every step. Black-box AI optimizes for clean output. Glass AI optimizes for human agency over shapeable material. ## Output vs material | Output | Material | |---|---| | Finished-looking answer | Shapeable intermediate artifact | | Accept/reject loop | Inspect/steer/edit loop | | Hides process | Shows relevant state and provenance | | User becomes approver | User remains author | | Can weaken judgment | Can build judgment | ## Material closeness test Before choosing an interaction mode, ask: 1. Is this a taste-bearing or judgment-bearing decision? 2. Would hiding the process make the user less able to learn, steer, or verify? 3. Does the user need to see the plan, sources, diff, trace, or intermediate state? 4. Is the artifact still shapeable, or has the system collapsed it into final output? 5. Can the user interrupt, edit, branch, or take over? If the work is boring execution, delegate it to an agent with verifiable output. If the work carries judgment, taste, trust, scope, provenance, architecture, or release responsibility, keep the user close to the material through direct UI, generative UI, stable truth surfaces, or visible agent traces. ## Relationship to Interaction Mode Routing [[interaction-mode-routing|Interaction Mode Routing]] chooses the tactical surface: direct UI, agentic delegation, generative UI, or stable truth/routing. Material Loop and Glass Interfaces explain why the route matters: the interface should preserve the user's ability to inspect, steer, learn, and care. ## Pixi Wiki implication Pixi Wiki should keep source/navigation stable while generating temporary review surfaces above the corpus: - source-to-output trace maps; - namespace coverage dashboards; - compile-review reports; - MCP/live-route visibility reports; - stale route repair panels. The goal is not only to publish knowledge, but to keep the knowledge system close enough to the material that humans and agents can inspect, route, correct, and trust it. ## I-know-kungfu implication [[../entities/i-know-kungfu|I-know-kungfu]] should not become a magic import button. Its strongest surface is a glass fit-check interface over stable wiki truth: source coverage, local overlap, proposed harmonization, serving entrypoint, refusal boundaries, trust/eval state, and provenance. ## Hermes Mission Control implication Hermes review surfaces should show enough plan, diff, evidence, risk, and next-action state for Jamie to remain the author of the decision instead of merely approving a plausible agent summary. ## Source Compiled from `Knowledge/concepts/material-loop-and-glass-interfaces.md`, inspired by Ryo Lu's "Closer to the Material" talk for Cursor Compile 26. --- title: Product Management as System Steering created: 2026-07-12 updated: 2026-07-12 type: concept status: compiled namespace: ai-native-product-surfaces source: Knowledge/concepts/product-management-as-system-steering.md resource: https://roachcap.com/memos/secrets-of-the-best-pms.html confidence: medium --- # Product Management as System Steering A PM steers a system of decisions, ownership seams, people, constraints, and participant incentives. The practical review formula is: ```text Decide under ambiguity. Own the gaps. Route through people. Trade scope first. Align ecosystem incentives. ``` This lens is adapted from Fahd Ananta's [“Secrets of the Best PMs”](https://roachcap.com/memos/secrets-of-the-best-pms.html). Treat the article as practitioner judgment, not empirical proof. ## Five steering responsibilities ### 1. Match decision tempo to reversibility Make timely calls with incomplete information, including explicit decisions not to pursue something. Move quickly on reversible choices; slow down when downside is hard to undo. Record the evidence, assumptions, rejected options, owner, and review trigger. ### 2. Close ownership gaps Extreme ownership means no unowned seam. It does not mean the PM personally executes every task. Keep the path from problem and design through build, launch, support, and learning connected with clear handoffs and escalation. ### 3. Adapt influence to people Understand strengths, limits, incentives, working styles, and feedback needs. Use data for empirical disagreement, prototypes for hard-to-verbalize choices, memos for durable reasoning, coaching for learning, and escalation only when its trust cost is justified. ### 4. Trade scope first For a bounded delivery window, assume team and time are fixed until proven otherwise. Cut, sequence, or defer scope to preserve the smallest coherent user outcome. Team and time can change, but usually more slowly and expensively than scope. ### 5. Align ecosystem incentives Map users, buyers, operators, employees, vendors, partners, investors, regulators, and other constituents. Durable products create mutual value that makes continued participation rational. The healthy target is incentive alignment and embedded value, not coercive lock-in. ## Product review questions 1. What call are we making, what are we declining, and when will we revisit it? 2. Which seam could fall between roles, and who closes it? 3. Which stakeholder dynamics or incentives could block execution? 4. With team and time fixed, what is the smallest coherent scope? 5. Which participants must gain durable value for the product to work? 6. What evidence will change the next decision? Use [[ai-native-problem-framing-framework|AI-Native Problem Framing Framework]] to define environment, actions, goals, constraints, and agency boundaries. Use [[agent-output-decision-artifacts|Agent Output Decision Artifacts]] to make the call and its evidence reviewable. Use [[role-aligned-deployed-project-proof|Role-Aligned Deployed Project Proof]] to show this product judgment in a portfolio case study. ## Guardrails - Decisiveness is not impulsiveness. - Ownership is not micromanagement. - Prioritization must preserve the core user outcome. - Influence tools lose force when overused; trust is a renewable coordination asset. - Incentive alignment must not become exploitative retention or dark patterns. ## Source - Fahd Ananta, [“Secrets of the Best PMs”](https://roachcap.com/memos/secrets-of-the-best-pms.html), published 2026-02-26. --- title: Role-Aligned Deployed Project Proof created: 2026-07-09 updated: 2026-07-12 type: concept status: active namespace: ai-native-product-surfaces source: Knowledge/concepts/role-aligned-deployed-project-proof.md confidence: medium --- # Role-Aligned Deployed Project Proof A strong portfolio project is not merely “an AI app.” It is a live, owned artifact built around a real responsibility, workflow, or pain from a target role. The selection rule is: ```text target role or JD → repeated job verb → painful or decision-critical workflow → smallest useful live artifact → evidence and limits → tailored case study ``` The intended hiring signal is simple: > This person already understands the work we do. This is a useful heuristic, not universal proof about interview outcomes. The quality of the artifact, role fit, communication, and the rest of the application still matter. ## Four hard gates 1. **Role-specific** — maps to a responsibility, toolchain, decision, or pain visible in the job description. 2. **Deployed** — a reviewer can use a live URL without installing the repo. 3. **Yours** — the builder can explain the product choices, implementation, tradeoffs, limits, and feedback loop. 4. **Truthful** — claims are supported by real behavior, tests, data, or user evidence. AI API wiring alone is not the proof. The role-specific decisions around the system are the proof. For PM roles, use [Product Management as System Steering](product-management-as-system-steering.md) to show the ambiguous call, ownership seam, stakeholder/influence choice, scope cut, ecosystem incentives, and evidence that changed the next decision. ## Build-selection questions - Which exact role or job description is this for? - Which verbs repeat: review, investigate, forecast, analyze, summarize, route, monitor, optimize, communicate? - Who owns that workflow, and what delay, error, risk, or decision cost do they experience? - What is the smallest end-to-end artifact that performs one useful transformation? - Can it be deployed cheaply and reviewed in under two minutes? - What evidence will show that it works? - Which failure mode or boundary should be visible? - What demonstrates judgment from the target role rather than only API integration? ## Starter idea catalog Tailor these against a real job description. Do not treat them as default builds. ### Software engineering - AI code review bot: diff in, bounded risks and review questions out. - Smart bug explainer: logs and environment context in, likely causes and next checks out. - Resume parser API: resume text in, validated structured fields and confidence out. - PR summarizer: pull request in, changes, risks, tests, and reviewer questions out. - Semantic documentation search: question in, cited answer or explicit refusal out. ### Data science and ML - Job-market trend analyzer: public postings in, sourced skill and role trends out. - Churn predictor: public dataset in, evaluated classification and explainable review surface out. - Sentiment dashboard: public posts in, time-based trends with coverage and bias notes out. - Sales forecasting tool: historical series in, backtested forecasts and scenario comparisons out. ### Business, marketing, and operations - Content brief generator: keyword and audience in, editable sourced brief out. - Email campaign analyzer: campaign copy in, prioritized CTA, tone, and subject-line experiments out. - Competitive intelligence bot: public company sources in, cited positioning comparison out. - AI meeting summarizer: transcript in, decisions, owners, deadlines, and unresolved questions out. ### Communications and information systems - Press release analyzer: release in, strategy, claims, risks, and journalist angles out. - Internal knowledge-base Q&A: documents in, permission-aware cited answers or “not found” out. ## Translate the proof by role - **Software engineering:** integration quality, reliability, tests, observability, security, and deployment. - **Data science / ML:** data provenance, baselines, evaluation, uncertainty, drift, and model limits. - **Business / marketing / operations:** decision impact, workflow adoption, experiment design, and measurable action. - **Communications / information systems:** source fidelity, information architecture, editing judgment, permissions, and retrieval quality. - **Product management:** problem selection, user/workflow evidence, scope tradeoffs, success measures, launch choice, and learning. ## Guardrails - Start from the role, not the project list. - A GitHub repo alone is not deployed proof. - Do not hide empty states, weak data, model uncertainty, or unsupported claims. - Prefer one complete role-shaped workflow over a broad platform. - Tailor the artifact itself, not only the application copy. - Build the smallest missing signal; do not duplicate a capability already proven by another project. ## Related concepts - [AI-Native Problem Framing Framework](ai-native-problem-framing-framework.md) - [Interaction Mode Routing](interaction-mode-routing.md) - [Agent Output Decision Artifacts](agent-output-decision-artifacts.md) - [Product Management as System Steering](product-management-as-system-steering.md) --- title: Side-Quest Validation Loop created: 2026-07-13 updated: 2026-07-13 type: concept status: active namespace: ai-native-product-surfaces confidence: medium source_url: https://www.youtube.com/watch?v=SE401zf_fgM --- # Side-Quest Validation Loop ## Definition A side quest is a **small, reversible, useful, and enjoyable experiment** that puts an idea in contact with real people before it earns serious product or business commitment. > Shrink the idea, ship a useful probe, expose it to reality, observe commitment, then kill, adjust, continue, or promote it. Deya's video frames this through creator and business anecdotes and demonstrates prototypes using sponsored tool Lovable. The transferable method is tool-independent, and the examples should be treated as practitioner evidence rather than universal proof. ## Design criteria 1. **Tiny enough** — test the important uncertainty without triggering a large project. 2. **Useful enough** — solve one recognizable problem for one specific person well enough to produce a real reaction. 3. **Fun enough** — sustain a short ambiguous exploration. Fun helps execution; it does not validate demand. ## The loop 1. Name the person, problem, and uncertain assumption. 2. Choose one learning question. 3. Build the smallest valid artifact: offer, manual service, workshop, calculator, waitlist, or prototype. 4. Put it in front of the intended users. 5. Observe behavior rather than collecting compliments. 6. At the timebox, kill, adjust, continue, or promote it. ## Evidence ladder From weaker to stronger: 1. compliments; 2. use or sharing; 3. reply, signup, booking, or another intentional commitment; 4. payment; 5. repeat use or continued payment; 6. unsolicited referral. A generated prototype or live link is an artifact, not validation. Validation begins when the intended user behaves differently. ## Promotion gate Define this before launch: ```text Person and problem: Smallest artifact: Distribution path: Timebox: Evidence target: Promote if: Kill or revise if: ``` Promote only when there is both **external pull** from user behavior and **builder pull** after doing the real work. Increase commitment one step at a time: a waitlist may justify interviews or a manual pilot, not a full product. ## Common failure modes - branding, planning, or funnel work without user contact; - treating publication or compliments as proof; - choosing only for fun while ignoring problem severity and willingness to pay; - running too many quests without timeboxes or decisions; - letting a prototyping vendor define the method. ## Product-surface connection This loop complements [[product-management-as-system-steering]]: move quickly on reversible tests and slow down as commitment becomes expensive. It also complements [[role-aligned-deployed-project-proof]]: use tiny public artifacts to test whether a workflow is useful before investing in a full portfolio or product build. ## Compact formula > Tiny enough to start. Useful enough to test. Fun enough to finish. Real enough to produce evidence. ## Source - Deya, ["your $1M business starts as a side quest, here's how."](https://www.youtube.com/watch?v=SE401zf_fgM), YouTube. The full user-supplied transcript remains in the private source vault and is not compiled into this public namespace. --- title: "Taste Requires Contact: Building Judgment in the AI Era" created: 2026-07-13 updated: 2026-07-13 type: concept status: compiled namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, ai, judgment, taste, curation, learning, interaction-design] sources: - Knowledge/concepts/taste-requires-contact.md - Knowledge/raw/transcripts/if-you-want-good-taste-you-have-to-eat.md - Knowledge/concepts/material-loop-and-glass-interfaces.md resource: https://youtu.be/F4igbiu9eR8 confidence: medium --- # Taste Requires Contact: Building Judgment in the AI Era AI can make almost anything for you. It cannot decide what deserves to exist. That judgment begins long before the prompt, in the things you have used, heard, worn, tasted, copied, compared, and learned to name. Jason Liu captures this with a simple line: **if you want good taste, you have to eat**. A menu can show what a restaurant serves, but it cannot teach how the food tastes. In the same way, screenshots, moodboards, launch videos, summaries, and AI-generated references can point toward quality without giving someone the firsthand experience required to understand it. ## Taste is a practice **Taste requires contact** means that judgment develops through repeated, attentive encounters with real work. It is not downloaded as a reference collection or produced by generating more options. The practice loop is: ```text Contact → notice → name → imitate → compare → curate → risk ``` - **Contact:** use, hear, wear, eat, play, or handle the thing itself. - **Notice:** detect specific choices and reactions. - **Name:** gain language that turns impressions into distinctions. - **Imitate:** test whether you actually perceived the structure. - **Compare:** inspect gaps across the original, your attempt, and alternatives. - **Curate:** select, reject, combine, sequence, and protect. - **Risk:** depart from consensus with a choice that expresses conviction. ## Taste has three jobs Liu's account combines: 1. **Aesthetic judgment** — sensing what is beautiful, coherent, expressive, or well made. 2. **Audience judgment** — anticipating what other people will understand or value. 3. **Personal conviction** — choosing something that may not already be validated by consensus. Good taste is not simply predicting popularity. Audience awareness without conviction produces safe consensus. Conviction without communication can become private expression that reaches no one. The useful tension is understanding the audience without becoming ruled by it. ## Vocabulary makes perception actionable Contact creates impressions. Language turns them into diagnoses. Art has composition, depth, form, colour, and line confidence. Engineering has abstraction, contracts, coupling, and architecture. Animation has timing, easing, and sound design. Cooking has acidity, salting, searing, and marination. Fashion has silhouette, proportion, and drape. Without vocabulary, judgment stops at “I don't like it.” With vocabulary, it can move: ```text vague reaction → named distinction → testable change ``` Words do not replace experience. They make experience inspectable enough to compare, explain, and refine. ## AI flips the traditional gap Beginners have often developed taste faster than execution. They could see that their work was wrong but could not yet repair it. AI can reverse that relationship. It produces polished artifacts before the user has developed the judgment required to evaluate them. Output leaps ahead while perception, vocabulary, and standards remain unchanged. Experienced practitioners often gain more from AI because they bring references, causal models, diagnostic language, failure patterns, and the ability to distinguish a plausible result from a fitting one. Their advantage is not merely prompting. It is judgment. ## Preserve the friction that teaches Not all friction is valuable. Repetitive formatting, boilerplate, file transfer, and deterministic checks are good automation targets. Other activities may carry the learning: - firsthand observation; - careful comparison; - attempted imitation; - diagnosis and error correction; - preference formation; - final selection. Liu's music-transcription example makes the distinction clear. The value is not possessing completed notation. The value is listening closely enough to produce it, making an attempt, detecting the mismatch, and correcting it. Before automating a learning task, ask: > **Is performing this activity how the person develops the judgment they are trying to acquire?** If yes, assist the loop without replacing its learning-bearing centre. ## Copying and wandering Copying is active perception. A failed imitation reveals the gap between what someone thought they noticed and what the original actually does. Once the structure is understood, variation can become deliberate rather than accidental. Wandering expands the reference field. Browsing a book, trying clothes you will not buy, testing unfamiliar software, listening outside a familiar genre, or following a strange reference creates encounters that destination-only systems filter out. Efficiency is useful when the destination is known. It is a poor default when the purpose is discovery. ## Product and agent implication [[material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] explains how people develop judgment by staying close to shapeable work. Taste Requires Contact adds the input side: people must also stay close to the material they consume. Together they form four loops: 1. **Reference loop:** encounter, notice, compare, and name. 2. **Making loop:** imitate, generate, inspect, and revise. 3. **Curation loop:** select, reject, sequence, and protect. 4. **Conviction loop:** depart from consensus and take a creative risk. [[interaction-mode-routing|Interaction Mode Routing]] should preserve firsthand use, diagnosis, preference formation, and final selection when those activities carry the learning. Agents can retrieve comparisons, widen the reference set, handle repetitive execution, and expose alternatives. They should not silently replace the perceptual act the user needs to practise. Contact develops judgment; artifacts make that judgment legible. [[syntheses/side-doors-make-useful-work-legible|Side Doors: Make Useful Work Legible]] shows the applied bridge: public work can expose what someone notices, selects, rejects, and protects, allowing other people to inspect the judgment instead of trusting a claim to “have taste.” An artifact is not proof of good taste merely because it exists; it makes the underlying decisions available for evaluation. ## Guardrails - Taste is domain-specific and socially situated, not a universal score. - Vocabulary sharpens attention, but jargon without contact becomes performance. - Friction is not virtuous by itself; preserve only the friction that carries learning or responsibility. - Audience judgment and personal conviction should constrain each other, not erase each other. ## Related pages - [[material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] - [[interaction-mode-routing|Interaction Mode Routing]] - [[syntheses/side-doors-make-useful-work-legible|Side Doors: Make Useful Work Legible]] - [[../../agent-workflows/wiki/concepts/creative-ideation-routing|Creative Ideation Routing]] ## Source - Jason Liu, [“if you want good taste, you have to eat”](https://youtu.be/F4igbiu9eR8), YouTube. This page synthesizes a user-supplied transcript as a practitioner framework; it does not reproduce the transcript publicly. --- title: Verified Video Answer Surfaces created: 2026-06-27 updated: 2026-06-30 type: concept status: compiled namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, video-ai, product-framing, interaction-design] sources: - Knowledge/concepts/verified-video-answer-surfaces.md - Knowledge/concepts/video-retrieve-then-verify-loop.md - Knowledge/concepts/agent-output-decision-artifacts.md confidence: high --- # Verified Video Answer Surfaces **Verified Video Answer Surfaces** turn video AI from a ranked search box into an evidence-backed answer workflow: ```text question -> verified moments -> evidence cards -> clips/report -> next action ``` The answer should show the moment, timestamp, confidence, evidence, and coverage boundary. ## Shifu pivot note Shifu's current best framing is not “AI understands video.” It is a local video workbench that helps users find, verify, save, compare, and export moments with visible timestamps and evidence. AI is an assistant for transcript search, suggested tags, candidate expansion, and verifier notes only where it earns trust. For gameplay VODs, the primary surface may be workflow plus structured evidence: timeline marks, round tags, HUD/OCR state, killfeed/spike/phase metadata, manual corrections, collections, notes, and exports. ## Surface contract A useful verified-video answer includes: 1. interpreted question; 2. searched videos / time range / camera or source scope; 3. verified moments with clip, timestamp, confidence, and evidence sentence; 4. evidence trail: frame, transcript/caption excerpt, modality signal, and source video handle; 5. explicit no-match state when weak candidates are rejected; 6. recall warning for "every X" queries; 7. next action: save clip, export report, refine search, mark false positive/negative, or ask a follow-up. ## Product wedge formula ```text For [person who scrubs video], find [repeated high-value moment], return [timestamped clips + evidence + confidence], so they can [make a decision / create an artifact / coach / report / publish]. ``` Examples: - coach -> fast breaks, press breaks, missed rotations -> film-review clips; - creator -> beat, quote, visual action -> exportable clips; - training lead -> correct/incorrect procedure -> teaching examples; - operations lead -> candidate incidents -> verified report with rejected false alarms separated. ## UI mode guidance - **Direct UI:** video library, timeline, filters, saved clips. - **Agentic delegation:** long ingest, batch indexing, recurring scans, report generation. - **Generative UI:** answer cards, evidence strips, comparison views, recall/coverage warnings. - **Stable truth/routing:** source IDs, timestamps, captions, transcripts, verdicts, and eval logs. ## Quality bar Measure precision of verified results, recall for "every X" claims, time saved versus manual scrubbing, evidence-card trust, false-positive correction, false-negative discovery, and whether a returned report can be shared without redoing the review. ## Boundaries - Do not promise "find every" without recall measurement. - Do not treat top-k ranking as a verified answer. - Do not hide transcript/commentary dependence when target footage may be silent. - Do not build real-time/multi-camera/alerting infrastructure before the verified answer loop works. - Do not make AI the source of truth for precise gameplay state. - Do not judge a Tonbi-style implementation before captions/transcripts/fusion/verifier are wired. ## Related pages - [[video-retrieve-then-verify-loop|Video Retrieve-Then-Verify Loop]] - [[agent-output-decision-artifacts|Agent Output Decision Artifacts]] - [[interaction-mode-routing|Interaction Mode Routing]] - [[material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] - [[world-model-control-surfaces|World Model Control Surfaces]] ## Source Compiled from `Knowledge/concepts/verified-video-answer-surfaces.md` and adjacent AI-native product-surface concepts. --- title: Video Retrieve-Then-Verify Loop created: 2026-06-27 updated: 2026-06-30 type: concept status: compiled namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, video-ai, retrieval, verification, product-framing] sources: - Knowledge/concepts/video-retrieve-then-verify-loop.md - Knowledge/raw/articles/tonbistudio-mini-vss.md - Knowledge/raw/articles/nvidia-vss-docs.md confidence: high --- # Video Retrieve-Then-Verify Loop The **Video Retrieve-Then-Verify Loop** is the reusable pattern for answering natural-language questions over video: ```text ingest video -> segment -> caption/transcribe/embed -> retrieve candidates -> verify with visual/audio evidence -> return timestamped answer ``` Retrieval maximizes candidate recall. Verification filters false positives, cites evidence, sets confidence, and refines the answer. ## Why it matters NVIDIA VSS shows the industrial version: video IO/storage, agent routing, model endpoints, search, summarization, alert verification, stream handling, and telemetry. Tonbi's `mini-vss` shows the desk-scale version: segment one video, embed frames/captions/transcripts, fuse retrieval, then judge candidates with evidence. Its reported "fast break" test moved from 0.20 precision@5 on retrieval alone to 1.00 verified precision with 75% recall. The verifier also found a full-court press the human label pass had missed. The product lesson: **video search quality is candidate recall plus evidence-bearing verification, not just vector ranking.** ## Shifu / gameplay correction Jamie's private Valorant/CS smoke sharpened the boundary: OpenCLIP visual-only retrieval can work technically and still return wrong tactical moments. A fair Tonbi-style test needs segment captions, transcript search, fusion, deeper candidate pools, and verifier judgment. For precise gameplay, structured signals matter more than generic frame similarity: OCR/HUD, round phase, spike state, killfeed, economy, map location, utility usage, VOD metadata, before/after context, manual tags, and game APIs/demos where available. ## Design rules 1. Segment is the product unit; frames are evidence inside a segment. 2. Visual, caption, transcript, object, and metadata signals propose candidates. 3. Fusion should preserve recall before optimizing rank aesthetics. 4. A verifier should output match/no-match, confidence, evidence, and refined timestamp. 5. "Find every X" is a recall claim and needs deeper candidate pools. 6. A trustworthy no-result state can be better than noisy ranked guesses. 7. For gameplay, structured state should outrank generic visual embeddings until evidence says otherwise. ## Query tiers | Tier | Example | Answer path | |---|---|---| | Appearance | "player shooting", "forklift in aisle" | visual embeddings / detections | | Event | "made layup", "person enters restricted zone" | captions, transcript, object/action signals | | Tactical / semantic | "fast break", "unsafe loading pattern" | retrieval pool + verifier over clip context | ## Product implication For Jamie's app exploration, do not start with "video search app" as the promise. Start with one painful video-review job where verified timestamped answers save obvious scrubbing time. Strong first-user candidates include coaches, creators, course/community operators, and small teams reviewing support, sales, training, or operations footage. ## Related pages - [[verified-video-answer-surfaces|Verified Video Answer Surfaces]] - [[ai-native-problem-framing-framework|AI-Native Problem Framing Framework]] - [[world-model-control-surfaces|World Model Control Surfaces]] - [[../../local-ai-infrastructure/wiki/concepts/rag-over-agent-wikis|RAG over Agent Wikis]] ## Source Compiled from `Knowledge/concepts/video-retrieve-then-verify-loop.md`, Tonbi `mini-vss`, and NVIDIA VSS docs. --- title: World Model Control Surfaces created: 2026-06-26 updated: 2026-06-26 type: concept status: compiled namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, ai-native, product-framing, interaction-design, grounded-ai] sources: - Knowledge/concepts/world-model-control-surfaces.md - Knowledge/raw/transcripts/yann-lecun-world-models-next-ai-revolution.md - https://youtu.be/72Xj8k5WQX4?si=tFQOgcbG-xzmz7WI confidence: medium --- # World Model Control Surfaces **World Model Control Surfaces** translate Yann LeCun's world-model argument into a product/interface lens: expose state, available actions, predicted outcomes, objectives, guardrails, and evidence before recommending action. Core loop: ```text observed state -> candidate actions -> predicted outcomes -> objective / guardrail score -> recommended next safe step ``` ## Why it matters The transcript argues that grounded intelligence is not just declarative knowledge or next-token generation. It requires abstract predictive models of the world, planning by optimization, and guardrails that score imagined action/state sequences before execution. For product surfaces, the useful lesson is not “build a full world model now.” It is: **make the system's action model visible**. ## Product-surface use Use this lens after [[ai-native-problem-framing-framework|AI-Native Problem Framing Framework]] and before choosing the final interface mode with [[interaction-mode-routing|Interaction Mode Routing]]. A good AI-native control/review surface should show: 1. current state; 2. possible actions; 3. predicted outcomes; 4. task objective; 5. guardrails/constraints; 6. evidence and uncertainty; 7. recommended next safe step. ## Application targets - **Shifu / I-know-kungfu:** show source coverage, local overlap, route effects, refusal boundaries, provenance risk, and recommended import/serve action. - **Hermes Mission Control:** show task state, candidate next slices, likely side effects, verification evidence, and approval guardrails instead of only “agent says done.” - **Pixi Wiki:** show source-to-output trace, namespace coverage, MCP/raw/HTML visibility, stale route risk, and suggested repair or promotion action. - **RL Sim Labs:** separate environment state, allowed actions, dynamics model, objective/reward, evidence gates, and policy/planner output. ## Boundary This is a concept, not a standalone namespace and not a claim that current LLM agents already have robust learned world models. Do not create separate entity pages for Yann LeCun, JEPA, V-JEPA, SIGReg, or AMI Labs until those entities recur across more Pixi Wiki sources or become project-critical. ## Related pages - [[ai-native-problem-framing-framework|AI-Native Problem Framing Framework]] - [[interaction-mode-routing|Interaction Mode Routing]] - [[material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] --- title: I-know-kungfu created: 2026-06-19 updated: 2026-06-24 type: entity status: active namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, i-know-kungfu, knowledge-base-wikis, cookbook, local-first, agent-context] sources: - Projects/I-know-kungfu/Index.md - Knowledge/concepts/material-loop-and-glass-interfaces.md - https://github.com/pixiiidust/I-know-kungfu - https://github.com/pixiiidust/I-know-kungfu/issues/1 confidence: medium --- # I-know-kungfu I-know-kungfu is an AI-native product surface for growing a user's own knowledge base by importing, adapting, and serving bounded knowledge base wikis that agents can search, cite, and refuse against. ## Product shape The current flow is: ```text Find useful knowledge base wiki → check local fit → choose serving entry point → harmonize overlap → inspect scope / proof / refusal ``` The product is not primarily a generic "pack" marketplace. Each imported serving unit is conceptually a wiki: source-backed pages, scope/non-scope, provenance, freshness, and agent-friendly entry points such as MCP, `llms.txt`, raw Markdown, and `index.json`. A Knowledge Pack is the portable package/install format for a knowledge base wiki. The wiki is the thing users grow and agents use; the pack is how that wiki moves between local storage, Cookbook listings, and agent harnesses. ## Why it matters The goal is to let users evolve their knowledge base without reinventing the wheel. In the ideal case, a user can adapt proven wikis with known quality or track record to improve their own coverage, fill gaps, and avoid duplicating or polluting what they already know. For agents, bounded wiki entry points should be faster and more token-efficient than broad web search. Agents can search a specific source, cite exact pages, and refuse when a task falls outside the wiki's scope. ## Current status The first static Cookbook serving prototype passed Jamie's initial smell test with Variant C. The accepted direction is table/list-first and decision-first: check fit before trust, choose one serving entry point, then make overlap harmonization explicit. The repo now has a PRD, README, glossary, and ADR that center knowledge base wiki as the product object and demote Knowledge Pack to package/install format. ## Namespace role I-know-kungfu belongs in `ai-native-product-surfaces` because it is primarily about the user-facing and agent-facing product surface for trusted context routing and knowledge-base growth. It crosslinks to: - `agent-workflows` for agent consumption, MCP, `llms.txt`, and bounded source routing mechanics; - `pixi-vault` for compiled wiki / namespace publishing patterns; - `local-ai-infrastructure` for local-first serving and future retrieval infrastructure. ## Interaction mode refactor Use [[../concepts/interaction-mode-routing|Interaction Mode Routing]] as the product-surface lens for the next slice. I-know-kungfu should not become a pure chatbot or a static card pile. Its strongest wedge is a generated fit-check surface around stable wiki truth. Use [[../concepts/material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] as the authorship lens: importing a wiki should not feel like a magic install button. The user should see what will enter their knowledge system, where it overlaps, what it will route to agents, what it refuses to cover, and what evidence supports trust before it becomes durable context. | Mode | I-know-kungfu surface | |---|---| | Direct UI | Browse candidate wikis, search/filter metadata, choose install/serve target, inspect source links. | | Agentic delegation | Fetch candidate wiki metadata, summarize scope, detect local overlap, propose harmonization, run quality/eval/provenance checks. | | Generative UI | Fit-check report, overlap map, scope/non-scope checklist, provenance coverage grid, serving-entrypoint decision table, trust panel. | | Stable truth/routing | Wiki contracts, source pages, install manifests, `llms.txt`, `index.json`, MCP, PRD, ADRs. | Next product slice to spec: **Generated Fit Check Surface** — source coverage, local overlap, recommended serving entrypoint, refusal boundaries, trust/eval state, and clearly labeled evidence provenance. ## Boundaries The first slice should remain local-first and endpoint-ready. Do not treat the current project as a hosted marketplace, vector database, hosted RAG layer, payments system, public upload-moderation system, or cloud MCP service. ## Source Compiled from `Projects/I-know-kungfu/Index.md`, the public repo README/PRD, and GitHub issue #1. --- title: J-Space-Replay created: 2026-07-08 updated: 2026-07-08 type: entity status: working-demo namespace: ai-native-product-surfaces source: Projects/J-Space-Replay/Index.md confidence: high --- # J-Space-Replay J-Space-Replay is a public working side project for replaying a vision-language model's decoded internal readouts on a video timeline while it answers a question. The public lite version lets users browse preset traces; upload/new-video generation requires local install/GPU. The product surface is intentionally glass-box and cautious: it helps a user inspect Qwen2.5-VL logit-lens or fitted J-lens readouts, but it does **not** claim to reveal model thoughts or validate mechanistic claims about VLMs. ## Product frame ```text short video + question -> one offline Qwen2.5-VL pass on a local NVIDIA GPU -> per-layer residual capture -> logit-lens-v1 or j-lens-v1 decode -> schema-v1 trace -> replay UI: answer, video, word grid, patch/box overlays, unspoken/adversarial cues ``` The UI is a technical replay dashboard: honesty banner, query/answer console, video transport, readout-strength bars, answer-token × layer workspace grid, raw top-token drilldown, lens selector, and trace library. ## What the demo shows - **Answer precursors:** answer-token readouts can surface words such as `gravity` before the model emits them. - **Adversarial checks:** when an answer rules something out, cells that still decode the ruled-out word pulse as a warning surface. - **Unspoken readouts:** the UI lists words read across many cells that appear in neither prompt nor answer. - **Computed but unsaid content:** the ball-drop example shows late-mid J-lens readouts such as `level`, `horizontal`, `move`, `off`, and `right` while the answer remains a generic gravity explanation. ## Architecture - Backend: Python/FastAPI package `src/jsr/`. - Model path: Qwen2.5-VL-7B-Instruct, NF4 + SDPA + fp16 stack. - Trace pipeline: video frame sampling, prompt preparation, residual capture, lens decode, label extraction, grounding queries, validated `trace.json` schema v1. - Frontend: React + Vite screens for upload, progress, replay, and library. - Demo mode: no-GPU pre-baked traces for instant browsing. - J-lens: ships as a fitting recipe, not committed weights. ## Evidence and caveats The repo reports a verified J-lens port against component-level autograd checks, with the paper-faithful identity seed making final-layer J-lens readouts exactly equal to the logit lens. On synthetic clips, J-lens improves concept recall by about 31% and reveals a late-mid visual content band around layers 22-26. The honest boundary is central: - demo-quality interpretability only; - the Anthropic workspace paper validated text models, not VLMs; - single-token readouts only; - adversarial/unspoken surfaces are mechanical string analysis; - natural-video baselines and interventions are not done; - raw-token grid remains primary because concept-label recall is still limited. ## Source handles - Project hub: `Projects/J-Space-Replay/Index.md` - Repo: https://github.com/pixiiidust/j-space-replay - Lite preset library: https://pixiiidust.github.io/j-space-replay/ - Inspiration concept: [[../concepts/j-space-global-workspace|J-Space as Global Workspace]] - Screenshot/demo: `docs/screenshot.png`, `docs/demo.mp4` - Evidence: `reports/jlens_evidence.md`, `reports/m2_quality_gate.md` - Related concepts: [[../concepts/video-retrieve-then-verify-loop|Video Retrieve-Then-Verify Loop]], [[../concepts/verified-video-answer-surfaces|Verified Video Answer Surfaces]], [[../concepts/material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] --- title: Job Edge created: 2026-06-27 updated: 2026-06-30 type: entity status: live-public-prototype namespace: ai-native-product-surfaces source: Projects/Job Edge/Index.md confidence: high --- # Job Edge Job Edge is a live public prototype for finding **edge in crowded job searches**: freshness, geography, fit, and distribution signals that make a role more worth applying to now. The first use case is **Ashby PM Radar**, a PM/product-role radar that turns public Ashby job-board data into an apply-action queue. ## Thesis Most job-search tools optimize for more listings. Job Edge optimizes for better timing and lower competition. Core question: > Which roles are worth applying to today because timing, fit, and crowding signals create an edge? ## Current prototype The public dashboard is live at: ```text https://pixiiidust.github.io/job-edge/ ``` The public `pixiiidust/job-edge` repo contains: ```text job_edge/ashby_pm_radar.py # Ashby use-case CLI/scorer scripts/refresh_ashby_pm_radar.py # refresh wrapper for local/GitHub Actions use .github/workflows/refresh-ashby-pm-radar.yml dashboard.html # static interactive dashboard data/ # saved discovery, scoring, dashboard, provenance artifacts docs/ # product/PRD notes tests/ # unit tests ``` `ashby-pm-radar` is the first use case, not the full product boundary. ## Ashby PM Radar flow Current automatic flow: ```text saved Ashby company slugs → fetch public boards → score jobs → commit static JSON → publish GitHub Pages ``` Ashby exposes company-scoped public boards, not a global search endpoint: ```text https://api.ashbyhq.com/posting-api/job-board/{companySlug}?includeCompensation=true ``` GitHub Actions now runs the refresh every 6 hours and on manual workflow dispatch. The browser does not run Python; the dashboard only re-fetches the latest published static JSON. ## Discovery boundary Job Edge currently refreshes **known Ashby boards** from `data/discovered_slugs.txt`. It does **not yet automatically search the public web for brand-new Ashby companies**. This means a future run can still show `150` jobs and be healthy. The success signal is a fresh generated timestamp plus current board data, not a changed count. Next discovery layer: ```text automated search queries → extract jobs.ashbyhq.com/{slug} → update slug list with provenance → refresh boards ``` ## Edge signals The current scorer combines: - **Freshness** — recently posted roles are more actionable. - **Geographic narrowing** — Toronto/GTA/Canada roles shrink the applicant pool. - **Role specificity** — niche PM roles can be less crowded than generic product listings. - **Personal/product fit** — AI, agents, workflow, developer tools, design/product overlap, integrations, platform, and B2B SaaS. - **Distribution crowding** — public LinkedIn posting presence implies a larger applicant pool. Competition is inferred. Ashby does not expose applicant counts, and LinkedIn evidence only detects public posting presence; it does not scrape applicant counts. ## Triage buckets ```text apply_now high score, fresh/local, no LinkedIn posting found apply_fast_crowded high score, found on LinkedIn, likely larger applicant pool maybe plausible but weaker timing/fit/competition profile low_priority stale or low-score roles ``` Freshness buckets: ```text new_0_3d fresh_4_7d recent_8_14d aging_15_30d old_31_90d stale_90d_plus ``` ## Dashboard contract The dashboard is an action queue, not a generic job board. It supports: - Apply and source-job links. - Mark-applied state in browser local storage. - Copyable job notes. - Search by title, company, and location. - Filters for triage, freshness, and LinkedIn presence. - Sorting by best triage, freshness, low crowd risk, Canada/Toronto fit, role fit, or score. - Manual `Run Ashby refresh` link to the GitHub Actions workflow. - `Reload latest published data` button for re-fetching the current static snapshot. Primary workflow: ```text 1. Apply now — fresh/local roles with no LinkedIn evidence. 2. Apply fast — strong roles already visible on LinkedIn. 3. Review maybes — backup queue after the top targets. ``` ## Verification snapshot 2026-06-30 auto-refresh milestone verified: - PR #10 merged on `pixiiidust/job-edge`. - GitHub Action run succeeded. - Pages status returned `built`. - Live dashboard rendered 150 jobs from the auto-refreshed JSON snapshot. - Latest verified generated timestamp: `2026-06-30T22:08:53.138975+00:00`. - Browser smoke confirmed the refresh link, reload button, job rows, and no JavaScript console errors. ## Next slice Add automated discovery for new Ashby company slugs and a “new / removed / changed since last refresh” diff layer so users can distinguish a healthy refresh from a stable job count. --- title: LKY Avatar / Voice Persona Stack created: 2026-07-15 updated: 2026-07-17 type: entity status: working-demo namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, voice-ai, avatar, local-ai, evaluation] sources: - Projects/LKY Avatar/Index.md - Projects/LKY Archive/Index.md - https://github.com/pixiiidust/lky-brain confidence: high --- # LKY Avatar / Voice Persona Stack LKY Avatar is a working local fictional-interview product that turns the `lky-brain` reasoning-style adapter into a voice-first, interruptible experience with an animated elderly-statesman portrait, a tuned watermarked voice, and a transcript designed as the record of proceedings. It is a separate applied product from the completed [[../../curated-tuning-datasets/wiki/entities/lky-archive|LKY Brain / LKY Archive]] corpus-and-adapter case study. The brain supplies the reasoning-style model; the avatar product owns the interaction, serving, disclosure, failure handling, and public-experience gates. The source avatar and voice-training repos are private. This page records a public-safe system summary, not private launch instructions. ## Product frame ```text speak or type a question -> streaming transcription -> local LKY-style language model -> fine-tuned elder voice -> animated portrait + spoken reply -> transcript/export as inspectable record ``` The web surface uses a broadcast-interview hierarchy rather than a generic chat layout: - portrait stage and session-state lamp; - persistent fictional-AI disclosure; - `TRANSCRIPT OF RECORD` for spoken and written turns; - voice input, typed-note fallback, and free interruption; - transcript export, visible reset, and explicit busy/rate-limit states; - text-only continuation when voice synthesis is unavailable. The current default avatar is a bundled illustrated elderly-statesman portrait sprite with state-mapped expressions, mouth motion, blink, and breathing. It replaced an earlier anime placeholder. A separate Live2D path remains for a future full custom rig. ## Architecture - LiveKit Cloud carries realtime rooms and WebRTC media. - Deepgram provides streaming speech-to-text. - A Python voice agent owns persona prompting, history, interruption, TTS state, degraded text delivery, per-turn fact retrieval, and the uncertainty guardrail. - A small audited fact sheet is split into retrievable sections. The best matches are inserted immediately before the latest question behind a source-over-memory instruction. - Deepgram Nova-3 receives a Singapore proper-noun `keyterm` list; the same vocabulary extends the TTS pronunciation seam. - The brain is a merged epoch-2 `lky-brain` LoRA served as Q4_K_M GGUF through llama.cpp behind an OpenAI-compatible seam. - The voice is a Chatterbox t3 model with a merged rank-16 LoRA overlay, served through a loopback-only HTTP adapter. - A Vite/TypeScript client renders the interview surface and keeps transcript export client-side. - A small token server mints short-lived room tokens and enforces unique rooms, one active session, and per-IP limits. ## Evidence gates reached ### Brain and interaction - Warm brain decode: 80.5 tok/s p50; warm time to first token about 0.05 s. - End-of-speech to first audio: 3.96 s p50, 5.95 s worst observed over 10 live turns. - Agent-side playback stops about 18–21 ms after interruption detection; about 270 ms from raw speech onset including the deliberate detection window. - Thirty-minute soak: 37/37 turns with no failures. ### Fine-tuned voice The voice-training project compared two arms against a frozen Chatterbox baseline: | arm | blind preference | similarity | WER | outcome | |---|---:|---:|---:|---| | baseline | 2/20 in final LoRA pack | 0.8693 | 0.0324 | retained rollback | | GPT-SoVITS | 13/20 tuned | 0.9049 | 0.1274 | failed intelligibility gate | | **Chatterbox LoRA epoch 14** | **18/20 tuned** | **0.8900** | **0.0390** | **integrated** | The integrated voice kept the existing HTTP contract, passed a ten-turn same-GPU placement run at RTF mean 0.369 / max 0.397 with zero failures, and preserved Chatterbox's PerTh watermark at confidence 1.0000 on served output. ### Factual-grounding implementation - The audited sheet covers constituencies and offices, independence and merger, HDB, water, selected policies, family, and a critical Tanjong Pagar correction with institutional source notes. - Deterministic keyword retrieval selects only relevant sections per turn and leaves the brain server unchanged. - A 12-question eval records factual accuracy, persona quality, and fabrication as independent signals and supports matched grounding-on/off runs. - Repository verification passed across 148 root tests, 199 voice-agent tests with 3 live-service skips, 91 web tests, and a production build. - The implementation is merged; live factual lift and real-microphone proper-noun transcription are still operator-side proofs. ## Trust and honesty boundary - This is a fictional simulation. Generated answers are not authentic quotations and the portrait is illustrative, not an authentic photo or endorsement. - The style adapter is not a factual database. Live testing found invented constituencies, dates, and historical events. The application now has an audited retrieval layer, but model-quality lift has not yet been measured live. - Voice data and weights remain local. The generated audio stays watermarked. - Singapore proper-noun input/output seams are implemented. Real-microphone STT and broader TTS pronunciation remain acceptance checks rather than proven coverage. - The demo is not publicly deployed as of 2026-07-17. Home-GPU hosting is a measured recommendation, not a verified live service. ## Next gate Run the 12 fact questions with grounding on and off against the local brain, score factual accuracy separately from persona quality and fabrication, and live-check the Singapore keyterms through a real microphone. Reopen the grounding gate if either fails. Only after those proofs should the project choose between the full Live2D rig and public hosting. ## Related - [[../../curated-tuning-datasets/wiki/entities/lky-archive|LKY Brain / LKY Archive]] - [[../../local-ai-infrastructure/wiki/summaries/lky-brain-consumer-gpu-qlora|LKY Brain Consumer-GPU QLoRA Case Study]] - [[concepts/material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] - [[concepts/interaction-mode-routing|Interaction Mode Routing]] --- title: myAbode created: 2026-06-16 updated: 2026-06-16 type: entity status: parked namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, myabode, real-estate, ai-crm, product-strategy] sources: - Projects/myAbode/Index.md confidence: medium --- # myAbode **myAbode** is a parked AI-native real-estate CRM case study. It belongs under `ai-native-product-surfaces` because the learning value is product-surface design: how to turn messy constraints into prepared next actions that agents actually choose to use. ## Current posture myAbode is not an active build. It should be treated as a problem-first product-strategy case study, with the Command Grid prototype as the strongest artifact. The key adoption constraint is that real-estate agents are independent contractors. The product cannot rely on mandated workflow change; it must make the better action obvious enough that time-poor agents adopt it voluntarily. ## Product-surface lessons The durable lessons for this namespace are: - deterministic compliance boundaries matter in regulated workflows; - AI should reduce cognitive load without surveilling or nagging users; - prepared next actions are stronger than generic dashboards; - a precedent library and scratchpad-then-reconcile flow can keep humans in authority; - adoption incentives are part of the product architecture, not a go-to-market afterthought. ## Promotion boundary Do not promote myAbode into its own namespace while parked. Reconsider only if it becomes active again with its own source corpus, recurring update lifecycle, and public-facing audience. ## Source Compiled from `Projects/myAbode/Index.md`. --- title: Planned Program Intel created: 2026-06-16 updated: 2026-06-16 type: entity status: compiled namespace: ai-native-product-surfaces tags: [ai-native-product-surfaces, planned-program-intel, product-management, decision-routing] sources: - Projects/Planned Program Intel/Index.md - Projects/Planned/PRD.md confidence: high --- # Planned Program Intel **Planned Program Intel** is an AI-native decision layer for enterprise event programs. It is the strongest current product-surface proof inside the `ai-native-product-surfaces` namespace. ## Product shape The product routes customer-side decisions by answering: 1. What needs attention? 2. Who should decide? 3. What happened in similar past events? 4. What is different this time? 5. What action is recommended? 6. Should the human accept, change, override, or escalate? The core loop is not a chatbot. It is a decision-routing and institutional-memory surface: signals create decisions, decisions route to owners, evidence packages prior cases/exceptions/precedents, and resolutions become program memory. ## Status The Vite prototype is done and live. The source hub records the build queue as complete and treats future work as optional portfolio packaging: deck, Loom walkthrough, application submission, or case-study narrative. ## Namespace role Planned Program Intel anchors the product-surface namespace because it shows: - decision moments as the product primitive; - inspectable evidence rather than opaque confidence; - human authority through accept/change/override/escalate verbs; - program memory compounding through resolutions; - a PM-portfolio-friendly artifact that connects product reasoning to a working demo. ## Source Compiled from `Projects/Planned Program Intel/Index.md` with supporting historical PRD context in `Projects/Planned/PRD.md`. --- title: Shifu created: 2026-06-27 updated: 2026-06-30 type: entity status: pivot-thesis-under-review namespace: ai-native-product-surfaces source: Projects/Shifu/Index.md confidence: high --- # Shifu Shifu is a local-first video knowledge/workbench prototype: private videos stay close to the user while the app helps users find, verify, save, compare, export, and reuse timestamped moments. The current corpus direction uses Valorant/CS VOD-style material as an evaluation corpus, not as the full product identity. The gameplay thesis is under review because generic visual embeddings are weak at precise tactical state. ## Product frame ```text source video -> light indexing -> local heavy worker -> evidence/workflow surface ``` Shifu should expose candidate moments, keyframes, transcript/caption/structured evidence when available, modality state, verifier decisions, manual tags, and clear boundaries. It should not pretend a query was answered when evidence is missing; AI assists only where it earns trust. ## Local-first architecture The current architecture decision is: ```text VPS app/orchestrator -> upload, source registration, browser UI -> light MiniVSS smoke: segmentation + keyframes -> search/save/export surfaces -> verifier manifests and reports Local desktop GPU worker -> heavy visual embeddings -> transcription/caption artifacts when available -> structured verifier verdicts -> worker_artifacts//status.json ``` This follows the same product logic as games and creator tools: use the user's desktop GPU for heavy local media work, then make cloud GPU a later optional accelerator. Short contract: > Local first. Cloud when useful. Evidence always. ## Current milestone state The implementation chain through parent issue #3 is complete: PR #26 merged, issue #21 closed, parent #3 closed after post-merge verification on `main`, and PR #27 added the Windows/D: local-video guide. The app can run on an actual local video today for upload/light-processing/search smoke. It includes browser/API upload, source registration, light segmentation/keyframes, search/save/export surfaces, local worker status seam, verifier manifests, structured verdict import, deeper T3 verifier pools, verified-only T3 recall, baseline-vs-verified deltas, and negative/refusal reporting. The product thesis is now under review. Jamie's private Valorant/CS `mini-vss` smoke showed the GTX 1070 CUDA/OpenCLIP path and LanceDB visual index can work technically, but manual inspection found the visual-only `post plant throw` hits wrong. Transcript/commentary looked more useful, and the full Tonbi loop still needs a clean reproduction with captions/transcripts/fusion/deeper pools/verifier before judging `mini-vss` itself. Post-merge verification on `main`: focused verifier/eval tests 18 passed; full suite 80 passed with one warning; compileall clean; `git diff --check` clean; fixture evaluator passed 5/5 while honestly reporting `Private VOD detection proof: no`, `Verified hits: 0`, and T3 verified recall `0.00` until structured verdicts are imported. ## Boundaries - Jamie's previous local proof target was **NVIDIA GTX 1070**; next heavy experiments are expected after the repaired 5070 Ti desktop is available. - Cloud GPU rental is deferred until the local-worker seam proves useful. - Private media, generated frames, transcripts, embeddings, worker artifacts, verifier verdicts, and private reports should not be committed. - Fixture reports are plumbing smoke only. Real private-VOD detection proof requires private media, private hand labels, non-placeholder modality artifacts, and structured verifier verdict imports. - Shifu can run on an actual video today for upload/light-processing/search smoke; production-grade proof is not established. - Visual embeddings are low-trust recall for tactical gameplay until structured signals and verifier evidence support them. - Do not judge Tonbi `mini-vss` from OpenCLIP visual-only output; its intended loop includes captions/transcripts, fused retrieval, candidate pools, and verification. ## Source handles - Project hub: `Projects/Shifu/Index.md` - Repo: https://github.com/pixiiidust/shifu-app - Parent issue: https://github.com/pixiiidust/shifu-app/issues/3 - Final PR: https://github.com/pixiiidust/shifu-app/pull/26 - Final child issue: https://github.com/pixiiidust/shifu-app/issues/21 - Windows/D: guide PR: https://github.com/pixiiidust/shifu-app/pull/27 - Related concepts: [[../concepts/video-retrieve-then-verify-loop|Video Retrieve-Then-Verify Loop]], [[../concepts/verified-video-answer-surfaces|Verified Video Answer Surfaces]] --- title: AI-Native Product Surfaces — Master Index created: 2026-06-16 updated: 2026-07-17 type: index status: active namespace: ai-native-product-surfaces --- # AI-Native Product Surfaces — Master Index > **Definition:** AI is any device that perceives its environment and constraints to take actions that maximize its chance of successfully achieving its goals. > Scaffold index for `ai-native-product-surfaces`. Add compiled pages here as they are created. ## Concepts - [[concepts/agent-output-decision-artifacts|Agent Output Decision Artifacts]] — Compress verbose agent output into concise, visual, source-backed decision artifacts with explicit next actions and feedback controls. - [[concepts/ai-native-problem-framing-framework|AI-Native Problem Framing Framework]] — Defines environment/actions/goal/constraints for AI-native product surfaces. - [[concepts/interaction-mode-routing|Interaction Mode Routing]] — Refactor lens for choosing direct UI, agentic delegation, generative UI, or stable truth/routing surfaces. - [[concepts/j-space-global-workspace|J-Space as Global Workspace]] — Anthropic's J-space / Jacobian lens concept: a verbalizable representation subspace that behaves like a functional global workspace for report, modulation, flexible reasoning, broadcast, and alignment-auditing caveats. - [[concepts/material-loop-and-glass-interfaces|Material Loop and Glass Interfaces]] — Agency/authorship lens for keeping AI-shaped work inspectable, steerable, interruptible, and traceable. - [[concepts/taste-requires-contact|Taste Requires Contact]] — Judgment-building lens for firsthand experience, precise vocabulary, active imitation, comparison, curation, and creative risk; paired with Side Doors as the public-legibility layer. - [[concepts/product-management-as-system-steering|Product Management as System Steering]] — PM operating lens for timely decisions, gapless ownership, adaptive influence, scope-first tradeoffs, and ecosystem incentive alignment. - [[concepts/role-aligned-deployed-project-proof|Role-Aligned Deployed Project Proof]] — Project-selection heuristic for turning a target role or job description into a live, owned, evidence-backed artifact that demonstrates role understanding. - [[concepts/side-quest-validation-loop|Side-Quest Validation Loop]] — Low-pressure, timeboxed loop for turning an idea into a tiny useful public experiment and letting behavioral evidence earn further commitment. - [[concepts/video-retrieve-then-verify-loop|Video Retrieve-Then-Verify Loop]] — Video AI architecture pattern: retrieve high-recall candidate moments, then verify with timestamped evidence; for gameplay, visual-only is low-trust and structured signals matter. - [[concepts/verified-video-answer-surfaces|Verified Video Answer Surfaces]] — Product-surface pattern for video apps/workbenches that return clips, confidence, evidence, and recall boundaries with AI as assistant, not source of truth. - [[concepts/world-model-control-surfaces|World Model Control Surfaces]] — Grounded-AI control/review lens for exposing state, actions, predictions, objectives, guardrails, evidence, and the recommended next safe step. ## Entities - [[entities/i-know-kungfu|I-know-kungfu]] — Active local-first Cookbook wiki serving project for growing a user's knowledge base with bounded, agent-readable wikis. - [[entities/job-edge|Job Edge]] — Live job-search edge dashboard prototype using scheduled Ashby refresh, freshness, geography, fit, and LinkedIn-crowding signals to prioritize applications. - [[entities/j-space-replay|J-Space-Replay]] — Working public demo for replaying Qwen2.5-VL logit/J-lens readouts over video with explicit demo-quality interpretability boundaries. - [[entities/lky-avatar|LKY Avatar / Voice Persona Stack]] — Working local fictional-interview product combining the LKY reasoning adapter, tuned watermarked voice, animated portrait, audited per-turn fact retrieval, transcript/eval surfaces, and explicit live-proof boundaries. - [[entities/shifu-app|Shifu]] — Local-first video workbench/proof question with implementation chain complete; visual gameplay thesis under review after visual-only misses, next heavy tests wait for the repaired 5070 Ti desktop. - [[entities/myabode|myAbode]] — Parked real-estate AI CRM case study focused on prepared next actions and adoption constraints. - [[entities/planned-program-intel|Planned Program Intel]] — Done decision-routing and institutional-memory prototype for enterprise event programs. ## Summaries ## Syntheses - [[syntheses/side-doors-make-useful-work-legible|Side Doors: Make Useful Work Legible]] — Illustrated synthesis of problem-first opportunity search, public proof, five story examples, and verb-first taste/distribution; linked to Taste Requires Contact as the judgment-formation layer. ## Source Roots - `Projects/Job Edge/Index.md` - `Projects/J-Space-Replay/Index.md` - `Projects/LKY Avatar/Index.md` - `Projects/Shifu/Index.md` - `Projects/I-know-kungfu/Index.md` - `Projects/Planned Program Intel/Index.md` - `Projects/Planned/PRD.md` - `Projects/myAbode/Index.md` - `Knowledge/concepts/ai-native-problem-framing-framework.md` - `Knowledge/concepts/interaction-mode-routing.md` - `Knowledge/concepts/j-space-global-workspace.md` - `Knowledge/concepts/material-loop-and-glass-interfaces.md` - `Knowledge/concepts/taste-requires-contact.md` - `Knowledge/concepts/world-model-control-surfaces.md` - `Knowledge/concepts/agent-output-decision-artifacts.md` - `Knowledge/concepts/role-aligned-deployed-project-proof.md` - `Knowledge/concepts/product-management-as-system-steering.md` - `Knowledge/concepts/side-quest-validation-loop.md` - `Knowledge/concepts/video-retrieve-then-verify-loop.md` - `Knowledge/concepts/verified-video-answer-surfaces.md` - `Knowledge/raw/articles/tonbistudio-mini-vss.md` - `Knowledge/raw/articles/nvidia-vss-docs.md` - `Knowledge/raw/transcripts/yann-lecun-world-models-next-ai-revolution.md` - `Knowledge/raw/transcripts/if-you-want-good-taste-you-have-to-eat.md` - `Knowledge/concepts/verb-first-product-positioning.md` - `Knowledge/concepts/find-the-lock-problem-first.md` - `Knowledge/concepts/side-door-opportunity-search.md` - `Knowledge/raw/articles/how-to-enter-side-doors-maja.md` --- title: AI-Native Product Surfaces — Activity Log created: 2026-06-16 updated: 2026-07-17 type: log status: scaffold namespace: ai-native-product-surfaces --- # AI-Native Product Surfaces — Activity Log > Append-only namespace log. ## 2026-07-17 update | LKY Avatar fact-grounding layer merged - Refreshed `wiki/entities/lky-avatar.md` after issue #45 / PR #47 added an audited sectioned fact sheet, per-turn retrieval, source-over-memory and uncertainty instructions, Singapore STT keyterms, and a 12-question factuality eval. - Recorded the implementation verification and the new signal boundary: factual accuracy, persona quality, and fabrication must be judged separately. - Kept the live-proof boundary explicit: real-microphone STT and grounding-on/off local-brain results remain pending; the local demo is still not presented as a public live service. ## 2026-07-15 create | LKY Avatar / Voice Persona Stack - Added `wiki/entities/lky-avatar.md` from the canonical `Projects/LKY Avatar/Index.md` hub after reconciling current private avatar/voice repo docs, merged PRs, open issues, and eval reports with the public `lky-brain` foundation. - Kept the applied interview product separate from the LKY Brain corpus-and-adapter entity. - Captured the tuned Chatterbox LoRA win, watermarked same-GPU integration, interaction/stability gates, portrait-sprite milestone, and the factual-grounding / Singapore proper-noun frontier. - Omitted private weights, secrets, machine-only launch details, and any claim that the local demo is already publicly deployed. ## 2026-07-14 migrate | Move long-form attention guide to content-distribution - Removed the earlier misconception-first article and its four public figure assets from this product namespace. - The rewritten cross-format guide now lives at `content-distribution/wiki/syntheses/attention-architecture-for-long-form-content.md`. - Kept product demos in scope here while routing general video, essay, Substack, and X attention/distribution systems to the dedicated namespace. ## 2026-07-14 update | Add source-video framework figures - Added four Jamie-supplied frames to `assets/misconception-first-explanation-loop/` and embedded them beside the misconception, question-to-explanation, A-plot/B-plot, and combined-summary passages. - Added descriptive alt text, numbered captions, source attribution, and a rights note. - Kept the formula framed as a mnemonic for the video's structure, not a validated quantitative model. ## 2026-07-14 create/update | Misconception-First Explanation Loop - Added `wiki/concepts/misconception-first-explanation-loop.md` from the canonical Knowledge concept and Jamie's supplied “Veritasium - What you don't see” transcript. - Preserved the reusable sequence: misconception → question → prediction → explanation → revised model, plus connected concrete A-plot / technical B-plot switching. - Routed it narrowly to technical explainers, product demos, and case studies; kept indexes, `llms.txt`, runbooks, and reference docs retrieval-first. - Kept the full supplied transcript private and separated practitioner learning evidence from retrospective virality claims. ## 2026-07-13 update | Connect taste formation to public legibility - Cross-linked `Taste Requires Contact` and the Side Doors `Verb-first taste` section in both directions. - Clarified the relationship as formation → evidence → distribution: contact develops judgment, verb-first choices expose it, and public artifacts let it travel. - Preserved the boundary that an artifact does not prove good taste by existing; it makes decisions inspectable. ## 2026-07-13 create/update | Side-Quest Validation Loop - Added `wiki/concepts/side-quest-validation-loop.md` from the canonical `Knowledge/concepts/side-quest-validation-loop.md` synthesis and Deya's YouTube video. - Preserved the tiny/useful/fun criteria, behavior-weighted evidence ladder, and predeclared kill/adjust/continue/promote gate. - Kept the framework tool-independent and separated the sponsored Lovable demonstration and anecdotal business claims from universal proof. - Kept the full user-supplied transcript private; the namespace contains a public-safe synthesis and source attribution only. ## 2026-07-13 create/update | Taste Requires Contact - Added `wiki/concepts/taste-requires-contact.md` from the canonical `Knowledge/concepts/taste-requires-contact.md` article and Jason Liu's “if you want good taste, you have to eat.” - Preserved the contact → notice → name → imitate → compare → curate → risk loop and the distinction between wasteful friction and learning-bearing friction. - Cross-linked Material Loop and Interaction Mode Routing so AI can remove boring execution without replacing the perception tasks that develop judgment. - Kept the supplied transcript private; the namespace contains a public-safe synthesis and source attribution only. ## 2026-07-13 create/update | Side Doors: Make Useful Work Legible - Added the illustrated synthesis `wiki/syntheses/side-doors-make-useful-work-legible.md` from the canonical `Knowledge/concepts/side-door-opportunity-search.md` page and Maja's “How to Enter Side Doors.” - Mirrored six Jamie-supplied framework and story screenshots under `assets/side-door-opportunity-search/` and used normal Markdown image syntax with local Pixi Wiki asset routes. - Preserved the default/outbound/inbound model, five story patterns, verb-first taste/distribution extension, and survivorship, attention, privacy, and unpaid-labor guardrails. - Updated namespace README and index; generated public output remains deployment-gated. ## 2026-07-12 create/update | Product Management as System Steering - Added `wiki/concepts/product-management-as-system-steering.md` from the canonical Knowledge concept and Fahd Ananta's “Secrets of the Best PMs.” - Preserved the operating formula: decide under ambiguity, own the gaps, route through people, trade scope first, and align ecosystem incentives. - Linked the framework to AI-native problem framing, decision artifacts, and role-aligned deployed project proof; retained guardrails against impulsiveness, micromanagement, and coercive lock-in. - Updated namespace README and index; no Daily Notes were copied or compiled. ## 2026-07-09 create/update | Role-Aligned Deployed Project Proof - Added `wiki/concepts/role-aligned-deployed-project-proof.md` from the canonical `Knowledge/concepts/role-aligned-deployed-project-proof.md` page. - Preserved the role/JD → repeated job verb → painful workflow → smallest live artifact → evidence → tailored case-study selection rule. - Kept the multi-role idea list as a starter catalog and made live deployment, ownership, truthfulness, and role-specific judgment explicit hard gates. - Updated namespace README and index; no Daily Notes were copied or compiled. ## 2026-07-08 create/update | J-space as Global Workspace concept - Added `wiki/concepts/j-space-global-workspace.md` from the canonical `Knowledge/concepts/j-space-global-workspace.md` page and Anthropic Transformer Circuits paper. - Routed the paper as a product-surface concept: replay verbalizable readouts to keep model work inspectable, while preserving the caveat that this is not direct access to model thoughts. - Cross-linked it to J-Space-Replay as the app's inspiration and honesty boundary. ## 2026-07-08 create/update | J-Space-Replay entity - Added `wiki/entities/j-space-replay.md` from the canonical `Projects/J-Space-Replay/Index.md` hub and public `pixiiidust/j-space-replay` repo docs. - Routed J-Space-Replay as an AI-native product surface: a glass-box replay dashboard for VLM logit/J-lens readouts over video. - Preserved the public honesty boundary: demo-quality interpretability, not validated VLM mechanistic evidence or model-thought claims. - Updated namespace README, index, and compiler map; no Daily Notes were copied or compiled. ## 2026-06-30 update | Job Edge public dashboard and Ashby auto-refresh - Updated `wiki/entities/job-edge.md` from the canonical `Projects/Job Edge/Index.md` hub after PR #10 merged on `pixiiidust/job-edge`. - Captured the live-public milestone: GitHub Pages dashboard, scheduled/manual GitHub Action refresh, successful workflow run, built Pages status, 150-job live snapshot, and no-browser-Python boundary. - Preserved the discovery boundary: current automation refreshes known Ashby company boards from `data/discovered_slugs.txt`; automated discovery of brand-new Ashby slugs and new/removed/changed diffs remain future slices. - Updated namespace README and index; no Daily Notes were copied or compiled. ## 2026-06-30 update | Shifu pivot/proof boundary - Updated `wiki/entities/shifu-app.md`, `wiki/concepts/video-retrieve-then-verify-loop.md`, and `wiki/concepts/verified-video-answer-surfaces.md` from the canonical project/Knowledge updates. - Captured the public-safe pivot posture: keep the local-first video workbench artifact, demote the old AI-core visual gameplay bet, and treat visual embeddings as low-trust recall until structured signals and verifier evidence prove value. - Recorded that Jamie's local `mini-vss` smoke likely has not exercised the full Tonbi loop yet; future proof should reproduce captions/transcripts/fusion/deeper pools/verifier before judging the implementation. - Updated namespace index; no Daily Notes were copied or compiled. ## 2026-06-29 update | Shifu implementation chain closed - Updated `wiki/entities/shifu-app.md` from the canonical `Projects/Shifu/Index.md` hub after PR #26 merged and parent #3 closed. - Captured post-merge verification: focused verifier/eval tests 18 passed, full suite 80 passed with one warning, compileall clean, `git diff --check` clean, and fixture evaluator passed 5/5 while preserving `Private VOD detection proof: no` and verified hits 0. - Reframed current state from PR-open gate to implementation-chain complete; next frontier is the actual private VOD proof run with GTX 1070 artifacts, labels, and structured verifier verdicts. - Updated namespace index; no Daily Notes were copied or compiled. ## 2026-06-28 update | Shifu verifier report gate - Updated `wiki/entities/shifu-app.md` from the canonical `Projects/Shifu/Index.md` hub after issue #21 / PR #26. - Captured the verifier-report milestone: manifests, structured verdict import, deeper T3 verifier pools, verified-only T3 recall, baseline-vs-verified deltas, negative/refusal reporting, and actual-video guide. - Preserved the proof boundary: fixture smoke may pass baseline retrieval, but real private-VOD proof still requires Jamie's GTX 1070 run, private labels, non-placeholder modality artifacts, and structured verifier verdict imports. - Updated namespace index; no Daily Notes were copied or compiled. ## 2026-06-27 create/update | Shifu local-first video knowledge entity - Added `wiki/entities/shifu-app.md` from the canonical `Projects/Shifu/Index.md` hub. - Routed Shifu as an AI-native product surface: local-first searchable video knowledge with VPS upload/light smoke and a local GTX 1070 worker artifact seam. - Preserved the boundary that cloud GPUs are optional future adapters and private media/model artifacts stay out of git. - Updated namespace README and index; no Daily Notes were copied or compiled. ## 2026-06-27 create/update | VSS video retrieve-then-verify concepts - Added `wiki/concepts/video-retrieve-then-verify-loop.md` and `wiki/concepts/verified-video-answer-surfaces.md` from the canonical Knowledge pages. - Routed Tonbi `mini-vss` and NVIDIA VSS as product-surface source material: the durable concept is candidate recall plus evidence-bearing verification, not a new namespace or infrastructure commitment. - Updated namespace README and index; no Daily Notes were copied or compiled. ## 2026-06-27 create/update | Agent Output Decision Artifacts - Added `wiki/concepts/agent-output-decision-artifacts.md` from the canonical Knowledge page. - Routed it as an AI-native product-surface concept: verbose agent output should become one-screen, visual, source-backed decision artifacts when the user needs to decide, approve, compare, or steer. - Kept pricing, revenue path, and product wedge details in the private brainstorm namespace; public Pixi Wiki gets the reusable surface principle only. ## 2026-06-16 create | Namespace scaffold initialized - Created README, CLAUDE instructions, raw folder, index/log, and typed wiki folders. - Source routing comes from `Wiki Compiler Maps/Namespace Wiki Compiler Map.md`. - No Daily Notes were copied or compiled. ## 2026-06-16 update | Add first product-surface compiled pages - Added `wiki/entities/planned-program-intel.md`. - Added `wiki/entities/myabode.md`. - Added `wiki/concepts/ai-native-problem-framing-framework.md`. - Updated `wiki/index.md` from scaffold to active content index. - No Daily Notes were copied or compiled. ## 2026-06-19 update | Clarify I-know-kungfu wiki-first framing - Updated `wiki/entities/i-know-kungfu.md` and `Projects/I-know-kungfu/Index.md` after product-language clarification. - Set **knowledge base wiki** as the core product object. - Demoted **Knowledge Pack** to the portable package/install format for a wiki. - Preserved Variant C: find useful wiki → check local fit → choose serving entry point → harmonize overlap → inspect proof/refusal. - No Daily Notes were copied or compiled. ## 2026-06-23 update | Add AI definition to namespace top - Added Jamie's AI definition to the top of `README.md` and `wiki/index.md`. - Preserved existing namespace scope and source roots. ## 2026-06-23 create/update | Interaction Mode Routing - Added `wiki/concepts/interaction-mode-routing.md` from the canonical Knowledge page. - Updated the AI-native framing page with interface-mode selection after E/A/G/C framing. - Updated I-know-kungfu with the Generated Fit Check Surface direction. - No Daily Notes were copied or compiled. ## 2026-06-24 create/update | Material Loop and Glass Interfaces - Added `wiki/concepts/material-loop-and-glass-interfaces.md` from the canonical Knowledge page. - Updated Interaction Mode Routing and AI-Native Problem Framing with agency/material-closeness framing. - Updated I-know-kungfu with the glass fit-check authorship lens. - Updated namespace README and index; no public Pixi Wiki deploy. ## 2026-06-26 create/update | World Model Control Surfaces - Added `raw/transcripts/yann-lecun-world-models-next-ai-revolution.md` from the user-provided transcript. - Added `wiki/concepts/world-model-control-surfaces.md` from the canonical Knowledge page. - Updated AI-Native Problem Framing, namespace README, and index with the state/action/prediction/objective/guardrail control-loop lens. - Classified this as a concept, not a standalone entity or namespace; no public Pixi Wiki deploy was pushed from this source update. ## 2026-06-27 create/update | Job Edge entity - Added `wiki/entities/job-edge.md` from the canonical `Projects/Job Edge/Index.md` hub. - Routed Job Edge as a job-search edge/dashboard product surface under `ai-native-product-surfaces`; `ashby-pm-radar` remains the first use case, not the whole product boundary. - Updated namespace README and index with the new source root and entity listing. - No Daily Notes were copied or compiled. --- title: "Side Doors: Make Useful Work Legible" created: 2026-07-13 updated: 2026-07-13 type: synthesis status: compiled namespace: ai-native-product-surfaces source: Knowledge/concepts/side-door-opportunity-search.md confidence: medium --- # Side Doors: Make Useful Work Legible A job is not fundamentally a title. It is a bundle of problems someone wants solved badly enough to spend money, trust, or attention on another person. That framing changes opportunity search from: ```text find posted role → submit credentials → wait to be interpreted ``` into a larger search problem: ```text find meaningful work or a live problem → understand it with specificity → make useful capability visible → place the signal where a relevant person can recognize it → let a conversation reveal the opportunity ``` This synthesis draws from Maja's ["How to Enter Side Doors"](https://velvetnoise.substack.com/p/how-to-enter-side-doors), Jamie-supplied framework and story screenshots, the public [Traveler's Guide to the Latent Space](https://sweet-hall-e72.notion.site/A-Traveler-s-Guide-to-the-Latent-Space-85efba7e5e6a40e5bd3cae980f30235f), and a supplied transcript of the Shopify internship video. ## Companies contain problems before they contain jobs  *Figure 1. A company is a set of problems and desired outcomes. Screenshot supplied by Jamie from Maja's article.* Organizations are groups of people trying to make things happen under constraints. They need to understand markets, find customers, launch things, support users, explain their work, improve operations, hire, stay compliant, and remove bottlenecks. Those needs exist before a hiring team writes a job description.  *Figure 2. Some problems get packaged into formal jobs. Screenshot supplied by Jamie from Maja's article.* A formal role is a **packaged problem bundle**. Some needs have become legible enough to receive a title, budget, manager, and evaluation process. Other needs remain unscoped, cross functional, newly noticed, or ownerless. The front door begins after the packaging. Side doors operate around the gap between **problems that already exist** and **roles that have already been formalized**. ## Three routes  *Figure 3. The default, outbound, and inbound routes. Screenshot supplied by Jamie from Maja's article.* ### Default: apply to the package ```text person → job board → posted role → application queue → company ``` The organization defines the problem and the evaluation frame. The candidate enters after the role is legible. This route remains useful, but it is crowded and lossy. ### Outbound: go toward the live problem ```text person → specific company/person/problem → useful proof or observation → conversation → possible advocate or opportunity ``` Strong outbound is not generic networking. It notices a particular body of work, studies the context, and makes a small unit of relevant capability visible. The recipient should not have to invent the connection. ### Inbound: make the signal findable ```text real work → public artifact → discovery by a tuned-in person → conversation → possible opportunity ``` Essays, guides, tools, analyses, prototypes, experiments, videos, events, and communities can act as ambassadors. They let another person inspect how the creator thinks before deciding whether to make contact. ### Hybrid: the pitch is also the proof Some artifacts address one organization while remaining public and discoverable. The application itself performs the capability being offered. ## What creates signal The recurring ingredients are: 1. **Specificity:** attention aimed at a real person, problem, company, or discourse. 2. **Proof:** a prior action, useful artifact, or inspectable decision trail. 3. **Legibility:** another person can infer what the creator notices and can do. 4. **Placement:** the proof reaches a channel or community where its value can be recognized. 5. **Invitation:** the next step is clear, small, and optional. ```text specific attention + observable proof + useful placement + bounded invitation ``` ## Story 1: the Calm cold email  *Figure 4. Excerpt from Maja's "Two Thank You's + One Offer" email. Screenshot supplied by Jamie from Maja's article.* At eighteen, Maja sent Calm's CEO a long, earnest email after encountering his public work. She connected that trigger to relevant ideas and evidence from growing large Instagram pages. He replied, asked her to prepare a pitch, and they worked together. ```text specific person → genuine trigger → relevant proof → concrete offer ``` The lesson is not to imitate the email's length or intensity. The useful mechanism is that the message contained more than admiration: it gave the recipient enough specific evidence to imagine a working relationship. ## Story 2: the Blackbird internship that did not exist  *Figure 5. Maja's "Random moonshot" LinkedIn message. Screenshot supplied by Jamie from Maja's article.* Maja wanted to learn from Joel at the Blackbird Foundation, but no internship application was open. She explained the Startmate internship context, named why Blackbird and Joel specifically mattered, and connected her startup experience to the investor side she wanted to understand. An internship was later created. ```text specific person → specific context → credible fit → request beyond the published taxonomy ``` From far away, this can look like luck. From inside, it begins with a person paying enough attention to try the handle on a door that has no official label. ## Story 3: Jae's essay on taste  *Figure 6. LinkedIn post sharing Jae's essay on taste. Screenshot supplied by Jamie from Maja's article.* Jae published a strong critique of contemporary taste discourse on Substack and LinkedIn. The essay exposed how he selected evidence, rejected a prevailing frame, and constructed an argument. Eucalyptus CEO Tim Doyle saw it, commented, met him for coffee, and Jae was eventually hired. ```text independent point of view → public essay → recognition → relationship ``` The essay was not a disguised application. It was real thinking directed at a real discourse. That is why it could travel ahead of its author. ## Story 4: A Traveler's Guide to the Latent Space Ethan Smith created [A Traveler's Guide to the Latent Space](https://sweet-hall-e72.notion.site/A-Traveler-s-Guide-to-the-Latent-Space-85efba7e5e6a40e5bd3cae980f30235f) in response to repeated questions from the early AI-art community: "What's the prompt?" and "What are the settings?" It is not a casual post. The chaptered guide covers Disco Diffusion setup, prompt engineering, init images, model settings, GPU/runtime troubleshooting, experiments, and comparisons. It organizes scattered frontier learning so other people can begin farther along. Someone who recognized that capability contacted Ethan through Discord about becoming a technical cofounder of Leonardo.ai, later acquired by Canva. ```text frontier obsession → repeated community questions → useful guide → discovery by someone carrying a matching problem ``` The guide demonstrates experimentation, synthesis, technical curiosity, teaching, and community awareness without listing them as identity claims. ## Story 5: Shopify's proposed first marketing internship The supplied video transcript begins: > "Shopify doesn't seem to have any marketing internships. What if I became the first?" The creator does not stop at saying she is creative. She narrates repeated actions: building a YouTube community of more than 10,000 people, collaborating with brands, launching a podcast, organizing Socratica events, creating belonging, and continuing to publish without guaranteed attention. It ends: > "If you're all about bold ideas, then here's mine. Let me be your first." ```text missing role → propose role → show repeated actions → perform the capability in the pitch → direct ask ``` The video is a hybrid side door. It addresses Shopify directly while also functioning as a public artifact. Its storytelling, emotional structure, and ability to attract attention are part of the evidence. ## One framework, five different objects | Example | Route | What was made legible | Proof object | |---|---|---|---| | Calm email | Outbound | Internet attention, ideas, initiative | Specific email plus prior audience-building evidence | | Blackbird message | Outbound | Person-specific fit and learning intent | Context-rich LinkedIn message | | Jae's taste essay | Inbound | Judgment and argument | Public essay | | Traveler's Guide | Inbound | Technical experimentation and teaching | Useful frontier guide | | Shopify video | Hybrid | Story, connection, community, and audacity | Public application video | The reusable move is not a particular medium. It is: > Make the work inspectable enough that one relevant person can already imagine what engaging with you would feel like. ## Verb-first taste Technology discourse often treats **taste** as a noun someone possesses: a refined aesthetic, the right references, or membership in a fashionable scene. Verb-first taste asks what a person repeatedly: - notices that others miss; - selects and rejects; - edits or removes; - sequences under constraints; - protects when tradeoffs appear; - explains as better for the intended experience. A moodboard is an input. Taste is the pattern of choosing, rejecting, combining, and protecting. [[concepts/taste-requires-contact|Taste Requires Contact]] explains the upstream learning loop: firsthand encounter, precise noticing, vocabulary, imitation, comparison, curation, and creative risk develop judgment. Verb-first taste describes the downstream evidence of that judgment. Public work can then make those choices inspectable without reducing taste to a self-applied label. Jae's essay demonstrates taste because the artifact exposes selections and refusals. It does not merely claim the identity. ## Verb-first distribution Distribution is also commonly treated as a possession: a channel, audience, community, platform, or growth hire. A verb-first distribution model describes the behavioral chain: ```text person encounters → recognizes relevance → tries → experiences value → returns or shares ``` "We use LinkedIn" names a channel. It does not explain why someone notices, acts, continues, or tells another person. The examples show several distribution verbs: - the Calm email **reaches** one person with specific relevance; - the Blackbird message **routes** around a nonexistent application form; - Jae's essay **travels** through public discourse; - Ethan's guide **answers** repeated community questions and **circulates** among frontier practitioners; - the Shopify video **attracts**, **demonstrates**, and **invites** simultaneously. ## Community, brand, and strategy are verbs too - **Community:** Who gathers, contributes, returns, helps, and develops belonging? - **Brand:** What expectations are repeatedly created and kept? - **Strategy:** What is chosen, refused, sequenced, and protected under constraints? Nouns describe the identity someone wants credited. Verbs reveal the value they can repeatedly produce. ## How to do this without becoming annoying - Study before contacting. - Begin from the recipient's world, not a long biography. - Make the proof small enough to inspect quickly. - Do not require the recipient to invent the relevance. - Ask for one bounded next step. - Make "no" easy. - Do not confuse audacity with entitlement. - Do not spray identical messages at many people. - Do not donate weeks of speculative labor to prove seriousness. - Respect privacy, confidential information, and organizational boundaries. The useful unit is usually not a complete unpaid solution. It is the smallest artifact or observation that makes the quality of attention and capability visible. For the message layer, see [Reader-Centered Outreach Asks](/pixi-wiki/wiki/agent-workflows/wiki/concepts/reader-centered-outreach-asks.md.html). ## Limits These are memorable success stories, so they contain survivorship bias. They do not prove that cold outreach, public writing, or viral artifacts reliably cause jobs. Luck, privilege, timing, network position, taste, and platform distribution remain part of the outcome. Most attempts may receive no response. The defensible claim is narrower: > Specific proof can create routes and signals that a formal application alone cannot create. Side doors expand the move set. They do not make an unfair market fair, and they should not become a moral test that blames people when the market remains unresponsive. ## Working checklist Before using a side door, ask: 1. What real person, work, or problem am I paying attention to? 2. What have I noticed that is specific rather than generic? 3. What is the smallest useful proof I can make or point to? 4. Which verbs does the proof demonstrate? 5. Where can someone tuned to this problem encounter it? 6. Is the invitation bounded and easy to decline? 7. Am I respecting attention, privacy, and unpaid-labor boundaries? 8. What will I learn even if the door stays closed? ## Sources and image note - Maja, ["How to Enter Side Doors"](https://velvetnoise.substack.com/p/how-to-enter-side-doors), Velvet Noise, 2026-05-14. - Ethan Smith, ["A Traveler's Guide to the Latent Space"](https://sweet-hall-e72.notion.site/A-Traveler-s-Guide-to-the-Latent-Space-85efba7e5e6a40e5bd3cae980f30235f). - Framework and story screenshots supplied by Jamie from Maja's article for commentary and analysis. Copyright remains with the respective creators and publications. - Shopify transcript supplied by Jamie from the public internship video. Only short excerpts are reproduced here.