Quantizing AI into Executable Skills as Math Operators.

Hello community. So great that you are back. Let's talk about the next level of artificial intelligence. No, we're not talking about autonomous AI system. Come on. We are going to talk about the next step, a proactive AI, a continuous monitoring AI that actively sets its own action. And we have two brand new study here. One is from the Chinese University of Hong Kong and the other is from Google. So, let's dive in. Now, kind of sure you the next token prediction will not completely disappear, but we do not have time for this anymore in our classical transform architect, you know. It is now relegated here to the, if you want, cognitive control layer, while the action space where you decide which action to take here is now, if you want, coarse-grained into a higher-order symbolic operator field. And these are, of course, our skills. Now, you know, in theoretical physics, you don't do actually for each macromolecule a complete quantum electrodynamic analysis or a simulation, yeah? For every single quark or electron or gluon, you know, you need use approximation. So, you decouple now, let's say, the fast light degrees of the freedom, let's say, the electrons from the slow heavy degrees of the nuclei. So, you try to make your life easy if you code here your solution. Guess what? We're going to do maybe the same thing here. Now, let's say in modern AI, LLM next token prediction is here the fast probabilistic electron cloud, yeah? It's brilliant at fuzzy reasoning, some intent interpretation, and navigating unstructured state spaces. However, if you force this electron cloud to compute a real heavy, rigid mechanism here of a nuclear of executing here multi-step API protocol or whatever, the system becomes highly entropic. It will start to hallucinate, and finally, it will collapse. So, this is not our goal. Now, let's talk about agent skills, because agent skills, you will see in our study here, act like the heavy nuclei. So, you know skills. No skill MD files, beautiful localized table, and the beauty is deterministic procedural bundles of whatever we have. So, the LLM no longer wastes the token hallucinating how to write an API call character by character, now. Now, the LLM next token prediction is elevated, if you want to an orchestration level, now. Because now, its only job is to compute the transition amplitudes between the different microscopic skills, between the deterministic procedural bundles, and not anymore between each next token. This is just here not intelligent enough anymore, now. If you want to learn more about skills and how I see skills here, yeah, I am working as quantum error codes industrial AI. Skill MD is not enough, and how QML handles the agent skill MD outperforms Intropack. I have a lot of videos. But, let's come back to the basics. So, in classical LLM agents, like a React agent, the agent follows here a microscopic perception, reasoning, and an action loop, now. Simple. So, when observation leads here to reasoning state, generated here token by token, mapping to a low-level atomic action, like a raw tool call. This is our old system. And the authors of the first paper, you will see, identified that doing this at a micro level creates a massive procedural gap, because, think about it, the phase space of all possible generation is simply too large to have here a fast computation. And the probability of an unbroken chain of flawless execution, you know it, it will decay exponentially given here a particular sequence length. So, this is maybe not the way to go. So, therefore, what about this idea? We introduce here skill tuple here where the autos perform now a mathematical projection. So, what they do, and to be absolutely clear, they now say think about my example here with the electron cloud and here the nuclei, you know. See, they pull now the procedural knowledge out of the LLM's implicit parametric memory from its neural weights and externalize it into an explicit modular vector state. This is a step to significant become at first almost deterministic system and second extreme fast. We don't have any more the bottleneck of predict the next token, yeah? So, skill tuple, beautiful. So, when the agent operates, it evaluates here the applicability of some constraints to see against the current observation over the time T, and if a match is found, let's say like a rag-like retrieval methodology, and yes, this is the reason why my last two

videos were on rag because we will use something like rag here maybe also here for this particular skill manifold. The agent executes now all the procedure that are defined by M, our main document here, and the code or our auxiliary resources we call R. So, what this means mathematically, the action space transitions now from a continuous-like distribution over the vocabulary of size, let's say 65,000 text tokens, to a discrete Markov decision process or MDP over a finite basis set of skills. You see how simple mathematics can make the simple things in life. So, this means the policy pi theta now optimizes here over the skills S, and instead of computing at a microscopic path to mall acts as a high-level scheduler, binding it available to the operational inputs of our operator skills. This is here the first study. May 8, 2026, Chinese University of Hong Kong, Shenzhen. Real nice. And they start with a comprehensive survey of agent skills. Everything that we know up until today, you will find in this paper. Highly recommended. So, basic ideas, agents handle the high-level reasoning and the planning. No, this is also here where you have here this monster LLM, CRG B5. 5, or our Opus 54. 7. While the skills then, on the other hand, form now the operational layer that enables a reliable, reusable, and composable execution. And maybe you run this on your local LLMs, like, I don't know, you look here maybe at a gamma 4. So, let's go to the definition because a lot of questions are received from my viewers. Hey, skills and tools and API calls and MCP, is this not all the same? No, it is not all the same. Let's have a look. Tools, such as search engines, code interpreters, databases, domain-specific APIs that you have, extend now the agents far beyond their parametric memory alone, no? This is why we invented rag and then our tools. And standards that we have established, such as the mall context protocols or MCP, further reduce here the integration friction by providing a unified mechanism for the discovery and the invocation across heterogeneous provided that you find on the internet. A tool exposes here, if you want, an atomic capability. It specifies what can be done, but it does not specify how it should be used. Careful. So, a search tool does not say when search is preferable to the memory retrieval if you have a multi-agent system. An API tool does not say what to do when the schema changes. And a code interpreter does not say how the output should be validated. So, you have clearly defined boundaries. MCP and similar infrastructure standards solve also only an interoperability problem, not the procedural problem that we going to have a look at of turning multiple tool calls into a robust, validatable, and stable workflow. The authors of this first paper give us here a definition that is interesting exactly in the idea that I started this video. So, they rather than treating your tool calls as an isolated operation, system store and reuse successful procedures that your authors now define explicitly as skills. You might know them as kill MD file or soul MD files or whatever you have, yeah? But what is the exact mathematical definition? A skill is a reusable procedural artifact with bounded scope that externalizes task-focused know-how. Not only what can be done, but when to act, how to execute, what heuristic or failure modes matter, and how to judge completion. Formally, skill is modeled as I showed you as a tuple S, and we have M as our root instruction document that is here what the agent can load and follow. R is of course here a set of auxiliary resources, our reference documents, our reusable templates, executable scripts, or domain artifacts that it scan that extend what a skill accomplishes beyond M alone. And C, C encodes now the application conditions that govern when the skill should be retrieved and applied expressed as metadata, natural language descriptions, or simply vector embeddings in high-dimensional vector spaces. So, this is a very general representation that we can start with. Agent skills are now increasingly managed through some dedicated internet platforms. You can find Skillnet, Glo Up, Skill Up, whatever you found, and maybe this list is already out of date because new resources are available for you. Now, let's have a look at agent skill, and they give us here in this paper

some illustrative example, yeah? You have here prepare grounded literature review. Step one, a tool call to a literature database searching. Step two, the reasoning, the topic clustering. Step three, another tool call. Step four, summarization with the tool call to the paper sightings or whatever, or you go with pure coding here. You have the same tool call, reasoning, tool call, fix. Investigating a data anomaly, you got the idea. Now, skills differ from our tools and MCP servers, because they encode situated procedural knowledge, triggers, sequencing, fallbacks, pitfalls, and appears bounded reusable artifacts that can be loaded, inspected, shared, revised without becoming yet undifferentiated handbook. Skills need not to be tool centric. You have cognitive skills like a review checklist, no? Or an analysis workflow definition, mainly uses you the internal knowledge of an LLM, but still supplies structure and reuse beyond some ad hoc prompting, prompt optimization, yeah? And tools expose operation, skills package you the know-how for using them in context. Beautiful. Now, the paper gives us here a real deep definition of skill representation. We classify our agent skills are packaged here. Each skill consists of an instruction based main document M. Already have seen this tuple, now, then the auxiliary resources are and the trigger conditions C. Semantically, skill externalized procedural knowledge include operational structure, branching heuristic, normative constraints with M serving as the primary human-readable representation of our skill MD file. Now, they have here some representation. Of course, you can have text-based, you can be code-based, and you can be a hybrid-based and you skill representation. And they give you here all the different methodologies and other publication they found here for this particular sequence. So, if you want to have a look at a literature, I highly recommend this and look at what they just found at hybrid base from Jarvis here. Unbelievable. The next is the skill acquisition. This is the process of constructing or generating new skills. And here you all just tell us, "Listen, we can be human-derived like those. We have an experience-derived methodology like 1 2 3 my goodness 10 20 here. We can be task-derived like here Meta-GPTs self-discover or SWE agent. And we have a corpus-derived here like AutoGuide or Hugging GPT if you remember this years ago. So, they really give you here a beautiful overview of all the possibilities. If you want to see this here more here in an image-oriented presentation, the skill acquisition methods. So, we have the human-derived skill acquisition, the experience-derived skill acquisition, the task-derived skill acquisition, and the corpus-derived skill acquisition. And please have a look here at all those images. The next one is the retrieval of our skills. Now, as I've shown you here SkillHub and whatever you prefer, but careful, they can be heavily infested here with malware and whatever. So, yeah, this is really not here a secure operation at all. Skill retrieval and skill selection. Let's clearly define what is what. Skill retrieval concerns how a large skill pool is reduced to a manageable candidate set for your particular task. For example, through semantic matching, lexical lookup, generative access, or structure ever search. You have your keyword search, you have all the rag possibilities you can find you have here to find here from a large skill pool your manageable candidate set. And this reminds you perfectly here after complete retrieval augmented generation rag methodology from the old times when we had a vector lookup. And then the skill selection in contrast concerns which candidate skills should ultimately now be invoked, whether multiple skills should be composed, and how such choices should adapt to the current observation, to the sub goal, to the budget that we have, to the resources that are available for us. And you understand exactly what we're talking about. So you see, rag comes back to bite us here also. The more, I think, how many skills we have? 100,000, 200,000 skills available on the internet. This is just going crazy. So, just cross-reference, rag is going to come beautifully. Skill retrieval and skill selection here the screenshot from the image to what is provided here. Dense retrieval, sparse retrieval, generative retrieval, structural

retrieval. This is here what we would call a graph rag search, no? Then the skill selection for your particular job, you have your costs, your utility, all the feedbacks you have maybe from other users, the context of our selection, and then a skill composition if you need four or five different skill elements here for your particular complexity. Here the authors give us also here beautiful here literature recommendation. Here you go from dancing betting sports and keywords retrieval to generative retrieval like tool LLM or skill weaver auto skill. Here is skill net structure aware with hierarchical dependency graph here. Like skill net here and context aware skill composition, mem skill, momento skill, skill bench, and my goodness. Look at all this literature that is available for you. Now, what is absolutely fascinating skill evolution? How does it evolve? So, evolution ask how an already formed skilled artifact, and let's go simple with a skill markdown file is revised, validated, optimized, shared, and governed after it has been formed, no? But it can also be a skill folder, a program API, toolbox function. Skill graph structure of whatever you're working. And here the authors give us here this division here into skill revision, skill validation, policy coupling, repository evolution, and runtime governance. Beautiful. You have here now, if you want to see the skill revision, skill validation, everything here also as a screenshot here in a visualization. Beautiful. So, this survey is really providing here a great overview over examined LLM-based agent systems with the lens of agent skills. Remember, reusable procedural artifacts that coordinate our tools that we have, our memory that is available, the runtime context, here the resources, your budget, whatever are here the boundary conditions for your systems. And yeah, while the agents handle here the high-level reasoning and the real complex planning how to execute a particular job, how to reduce the complexity of a single task into a lower complexity five, six, 10 subtasks, skills from the operational layer now enable a reliable and composable execution. As I told you, maybe you go with a proprietary huge LLM here for the complex reasoning and the planning task, and maybe for the skill execution you go with a local cheaper smaller language model. And I have a particular video that shows you which are the best small language model for this particular task. Interesting. But now let's take really a look into the future with our second preprint. And now we want to have here an AI system that is always on. So, not open claw, let's go the next step. What is the next category here? The level three of AI that we can now find. Not autonomous driving, what is the next level after autonomous driving? So, this is an interesting question, and you are not going to believe that Google presented here paper May 7th, 2026. A gigantic coding, and of course we focus now here on the simplest sector that we have, code. Code we can verify, falsify, execute different languages, everything is available in code. Beautiful. So, a gigantic coding needs proactivity, not just autonomy. Google Labs. So, what is the idea by Google? They say the next shift that we expect in artificial intelligence is not here that you have your cloud codes, I don't know, whatever systems you use. We move now to proactive agents. Why should the AI wait till the user comes with a prompt, with a task for the AI to fulfill? The AI can be always on. scanning all your communication, all your action, all your environment, all your competitor, all your industry, continuously. And the AI can understanding here completely what you as a human, what is your intention, what is your job, professional or maybe private, what is your environment, what is your industrial professional level of coding. How can the I help you to become, I don't know, to the go to the next step. Let's formulate it in this way. So, this is what we're talking about. Proactive agents. You don't interact with the I, the I now interacts with the human whenever the I sees it fit. So, Google tells us, "Continuously observe repository, tool chain, workflow context. " Remember, we only here focus on coding. "Infer what matters from a high-level developer needs, identify emerging problems that you have maybe in your pipeline or opportunities you simply did not notice or you forget or

whatever happened to you. " And then the I decides what to do next before you as a human type a narrowly specific prompt. And this arrives then at the I. So, Google tells us, "You know what? We have here open claw always on. We have the system that are completely scanning as and let's say an AI assistant, your professional life, your private life, whatever you do, whatever is in your environment. " So, they say, "Okay, we do have some products and they that already moved from an initiation away here from the explicit prompt here. Cross automations or cloud code routines or tools schedule task. Let coding agents run here from schedules, you know, do we have the web hooks, GitHub events, the integration, or monitored repo state, whatever is the trigger. It doesn't matter, yeah? Whatever is here an important trigger, but they tell us here, "But this is what we have today. But if you look in the future, we want to they are not yet situation aware proactivities. So, we want here to go to the next step. And Google calls it a unit of proactive behavior here with an AI system. They go here with the term of an insight. So, you see really becomes important that we really understand exactly what we're talking about here in the definition of our terms. An insight, a context grounded time sensitive hypothesis about what matters next for the human developer paired with a decision to notify the human, question the human, draft something, or stay silent. So, this is interesting and interesting if you read the paper, Google has a particular focus on stay silent. Which I had to smile a little bit because they understand that sometimes you do not want your AI companion to do everything for you, no? You want as a human also to be still in the loop, make decisions, maybe meta decisions, but yeah. So, interestingly Google gives us here this level of agency. Level one, reactive agents. Run only when a human ask them to do something when prompted by a human. Level two, what you have in Open Claw, you know, your scheduler. Scheduler agents. Monday at 9:00 do A B C D. Scheduler agents run from schedules or predefined triggers and may filter, batch, rank outputs, do not learn across context per developer interruption policy, nothing. It just has a fixed schedule, it doesn't learn. But it knows what is happening. And the third level is the situation aware agent. The agent that knows everything about you as a human, as a person, as a professional coder, whatever. So, this situation aware agent monitors a continuous event stream that is happening live. Compare expected benefits with interruption cost. Treats silence as an explicit action, smile. And update a pair developer model from feedback. So, this is now interesting. Think about this. Suddenly, you are not anymore a member of the group of software developer or coder or whatever you'd like to call you when you work with the AI. Now, this is an individual AI-specific, targeted on your abilities, on your knowledge, what you can do, how you are connected, what you're currently working on as a human. The AI knows this because the AI is always on. And when you go to sleep, the AI might have a look at all your documents, all your emails, and whatever you have done in the last 5 years. So, great. Yeah. People who bought here a Mac Mini said, "Oh, maybe it was a good idea to have here an isolated system. " But, this is not the point. Google tells us that this is coming. So, what is here a prototype of a proactive agent in coding engine, huh? You have all your resources that you use during the day, huh? The calendar, GitHub, Slack, Docs, whatever app or system you use, doesn't matter, huh? And you have a continuous stream that goes now into the AI because the AI you have given access the AI to your complete computer, huh? Real-time data ingestion from diverse external sources, your personal sources, privacy, forget about it. Intelligence. And then we have our proactive engine, huh? So, there we have the development state that is exactly understanding what is happening right now, where you are as a human right now, where is are you with your task, with your coding, with your particular procedures. And then we have, and this is, I think, absolutely fascinating, a developer through your personal human profile, your mental model. Who the developer is and what the developer cares about. So, your name is Alex. You have a particular focus on back-end API and

security. You have your topics, Rust, GraphQL, distributed systems. You have your personal preferences, like low interruption mode and everything, but you also have, let's say, areas where you are not the perfect coder. So, it knows your strengths, your weaknesses, everything, because this system is continuously monitoring whatever you do. Each single question that you ask any AI system, and it integrates this in its knowledge and its insight. And now it understands, "Okay, I have a coder that is uh not really an expert in the area A. " So, whenever we have a new development here in area A, it will notify me, or it will question me. It will test me. It will give me here maybe here a puzzle to do to see exactly where I am here with my knowledge. It will draft here. Maybe if I wake up at 7:00 in the morning and I say, "Hello, my little uh proactive developer mental model, how are you? " And the AI will respond to me, "Hey, I noticed that you wanted to do today this and this coding, so I already have done it overnight. And now let's talk how we can further develop you as a human coder, because I have noticed you have the following 17 weaknesses. " So, life is going to be beautiful with this always-on AI. And then you have your interaction, and hopefully Apple will make Siri here really here the beautiful interaction here ways, so you have it wherever you go, wherever room you enter in your house. Now, you will have here the your conversation. It's a little bit strange, this paper, but it is absolutely fascinating if you look at the mathematical side. Because how do they imagine to train this system. So, let S denote here the develop state at a particular time T. This means the open bar for the branch, the recent commits, the calendar, the sprint deadlines, the ticket status, the communication context, everything that you have, yeah? Let E denote here the event stream, the cross context event stream, since ever, since the last decision. And then, AI groups event into code, project, communication, infrastructure, developer behavior. Yes, I forgot to mention that your emotional and professional behavior is also recorded, and therefore your part here of this decision process. And then, let A for an action have here the inside, okay? The action space that we have normally in optimization here of any AI model of an LLM. And let's make it easy, we just have the following: notify, question, draft, and stay silent. And again, you see stay silent. This is, yeah, where I smile. So, the first three actions show here a classical inside that we define. And Google wants to make sure that there is also here a particular inside when the AI learned from your behavior with the AI, this companion that you have right on your side every little second of your complete life, that it should stay silent. Mathematically, it is, of course, here that we work with expectation here. And in the simplest case, the artists of Google tell us, "Hey, look, this is your optimization that we have, yeah? Like a classical LLM optimization when the LLM is learning. " And it is kind of frightening that you can apply for this human mental behavior model exactly the AI guidelines where the AI learned copying here the complete internet. But think about it. Yes, of course, whatever it found on the internet is from human, yeah? So, okay, we use here the exact same a little bit further developed mathematical statements here, like you see here, where O here is, of course, as you already guessed it here, an outcome and developer observable outcome of a particular action A. Can be a response, a downstream code change, a recovered context, whatever. U is, of course, how useful that outcome is now to the human developer, while our theta captures here what the agent has learned about this particular developer, you, and the costs in dollars here is, of course, here the cost of breaking here the flow. Interestingly, they try to integrate here breaking the human flow. You know, if you are focused here on a particular task, should the AI come now and interrupt you and say, "Hey, excuse me, I just found that you made a mistake. I just found or I already did the job for you, you know? " And yeah, you are not here really the using the optimized mathematical structure, but you are a little bit here non-professional, no? So, try to find here a code statement here to really get here the perfect companionship running, no? So, you know, we have here a limited discrete action space. Notify, so this means a state change, the question, the clarification, the draft, the pull

request, comment, the patch, or the review thread, and then stay silent. Choose not to interrupt the human. So, I think this is interesting, but you know what? We go up a complexity level because now the policy optimization, as I told you, is a classical LLM training, no? Is now with an added layer, and this is here your human behavior. Because now it is not that you are the a group member where you can be averaged over your behavior, but now it is your individual daily second-per-second behavior. An inside is defined context grounded time-sensitive hypothesis that the AI has about what matters next to you in your work, in your profession. What is your human reaction? What should the AI provide for you if you have a particular goal, whatever. Paired with an action that the AI is going to take. And the AI is always on. So, maybe you're surprised when you wake up one day and the AI has a surprise for you. As I told you, we have to add a complexity level because now we don't have a PaLM 2 or classical policy optimization, but now we have to deal with an inside policy. The Google calls it. Selects now the action, chooses the evidence, frames a message, and updates the future decision here from the human feedback. So, now it is not that you have just the coding as a topic to optimize coding, but now the inside is it is not a multi-agent system. You as a human with your AI assistant, you are not multi-agent. You are in a different cooperation pattern, yeah? And the AI has to learn, according to Google, the inside policy here, the interplay between the AI and you at individual you as a human. So, this will be highly personal data. So, I think Anthropic or OpenAI or Google or whatever will provide us here with some highly specific individual personal let's call it memory or knowledge or an inside policy. How an AI should deal with you. What is your preferred way to have an AI companion right next to your side. Yeah, mathematics. Level 3 engine inside produces here an inside stream that can be monitored, acted on, ignored, never shown, whatever. But we have this stream. Every global corporation can touch this data stream, yeah? And then it tell us we score whether that stream keeps an agent on the right path. So, they want to steer here every little second here what is happening. So, this is a little T being index here of the decision point, capital T the number of decision point in a scenario, ST the current developer in the project state, pi T to the surface policy being evaluated, and they go now with three new metrics. The first is an inside decision quality IDQ, and this scores whether the agent choose the right action at the right time. Did it provide you with the right code piece before you ask it? Did the AI really understand looking at what you are typing what your job is? That now is the perfect time to provide you with the solution to the next step. Yes, no. Inside decision, you see beautiful, we can optimize this, we can learn this. After hundreds and thousands of hours of interaction with you AI, you AI will definitely learn this. It's a simple formula. To other, the context grounding score ask whether the shown insight is supported by the right evidence. This is interesting, it means is this really a faithful thing a machine that you as a human really trust, no? And how can it optimize that you trust this machine more? Now, this is a highly individual thing, no? And then of course the learning lift which I think is great, measures whether the feedback improves your latest decisions. So, if you tell the AI, "Hey, I would prefer you behave in this way. " Or the AI tells you, "Hey human, I would prefer you do something in another way. " No, you will have to find, you know, this is like a marriage that you have here in a relationship. Hey, the learning lift and it can measure because the AI is always on. Everything you type, everything you say is analyzed by the AI, and the AI can absolutely analyze and notify you when your learning lift here is above average. Isn't this what we always wanted? Okay, so here we go. Now, together the metrics ask you whether the agent choose you the right action, use the right evidence to present to you, and improved itself after the feedback after you interacted and said, "Yeah, I like this. I do not like this. " You notice, no? So, those are now our additional elements that we have to train an AI on.

And we do it only on code because code is simple. Code is verifiable. Code is executable. There's no interpretation. There's no whatever. So, the framework above makes here the proactivity testable. Before let's say Google deploys here a full level three coding agent. So, you see the future of AI is absolutely amazing, no? And it means that coding agents that we have cloud code today will mature you from the tools that execute prompted task. Imagine a human ask a machine to do something into level three partners that may notice the drift, the uncertainty, or emerging opportunities before you as a human ask the machine to do it. Before you have to type any prompt. And proactivity in agent coding should be judged not by how often an agent acts, but whether it surfaces at the right the right inside time, no? I think this is interesting. I think Google has some insight here that sometimes when an AI interrupts you proactively, that humans do not really prefer to be interrupted 20 times by an AI like an open claw when this open claw thing wants to tell you, yeah, that you as a human Okay. So, with enough evidence and stay silent when the intervention is unwarranted. And at the end you will find in this preprint by Google, Google Labs, sorry, that the proactivity in the agentic coding scheme should be judged not by how often an agent acts now, no? But whether it surfaces here as an coding agent, as cloud code, whatever, the right insight at the right time for you as a human being having enough insight into your human thought process, into your professional level, into your professional argumentation, into the way you think as a human. [clears throat] Because the AI has to learn this. It has to simulate detect a pattern in your behavior so that the AI is able to become a better companion with the time. So therefore it needs a lot of evidence and there's also the option but I like to stay silent when the AI intervention is simply not really appreciated by the human being. I hope you had a little bit of insight into what is coming to you in the next weeks or months. I find it absolutely fascinating to have here the, let's call it a Chinese view here looking back what we have up until now, where we are here with our skills, tool, our agentic skills, our autonomous agents. And then look here at Google where they go now for proactive agents where you as a human you don't have to ask anything anymore because this AI is monitoring you completely. Hope to see you in my next video.