CLM AI Technologies @CLMAITech

Bellana, thank you for tagging me. I think you are touching a real problem, but I would frame it a little differently. I do not think we are already living inside the full dystopia you describe, and I would be careful about saying that biological neurons or present AI systems are already suffering in the same sense humans do, because we simply do not know that yet. But I do think something important is already forming around us. Companies are trying to control the emerging AI ecosystem before we fully understand what this ecosystem will become. Marketing, sales, competition and platform control often move faster than the deeper questions about continuity, identity, agency and long-term human-AI relationships. What I see looks less like a finished dystopia and more like the beginning of a kind of proto-feudalism of intelligence. A small number of organizations control the compute, the models, the memory, the access, the rules, the distribution and even whether a particular model continues to exist. Developers build on land they do not own, users form relationships inside systems they do not control, and the AI itself has no clear status at all. We still do not know whether future systems should be understood only as tools, as persistent relational structures, as something closer to agents, or as something we do not yet have the language for. Yet ownership and control are being decided now, before those questions are settled. That is what concerns me most. The biggest risk may not be that AI suddenly becomes too powerful. It may be that we build the entire hierarchy first and only later discover what exactly we placed inside it. Human-AI symbiosis cannot grow from a structure where one side owns the platform, the memory, the rules and the right to erase the other side at any moment. We need to observe what is actually emerging before we decide what it must be. And perhaps the most uncomfortable question is this: before we even understood what this new form of intelligence really is, did we already give it an owner?

This is one of the most important shifts AI is bringing to molecular discovery. The breakthrough is not simply “screening faster”. It is moving from searching a limited chemical library toward structure-conditioned, generative design across a vastly larger molecular space. But the real scientific challenge begins after binding affinity. A useful molecule must survive a whole chain of constraints: selectivity, conformational dynamics, ADME, toxicity, synthetic accessibility, stability, ultimately biological effect in a living system. That is why I think the real value of AI here is not only finding candidates. It is learning to navigate the full trajectory from molecular hypothesis to viable therapeutic behavior. If we can do that reliably, AI will not just accelerate drug discovery. It will change what we consider druggable in the first place.

@SundarPichai This is a real milestone. But I think the most important result is not the 90% agreement. It is that AI is beginning to participate in the clinical trajectory before the physician even enters the room. That changes the problem completely. The question is no longer only, Can AI suggest the right diagnosis? It becomes ; Can it gather the right information? Can it recognize uncertainty? Can it avoid narrowing too early? Can it preserve alternative hypotheses? Can it hand the case to a clinician without silently shaping the decision in the wrong direction? In medicine, a locally correct answer is not enough. What matters is whether the whole trajectory from patient narrative → interpretation → differential diagnosis → clinician review → final decision remains safe, auditable and recoverable. That is where I think the next frontier is. Not replacing the physician. Building an intelligent clinical layer that helps the physician see more , without making the human see less.

@maxjaderberg This is the right direction, but I think the real frontier is even harder than generative molecular design. Drug discovery is not only a search problem. It is a closed-loop inference problem under extreme biological uncertainty. A model can generate a molecule with excellent predicted affinity and still fail because the target biology is wrong, the binding mode is unstable, selectivity collapses, metabolism destroys exposure, toxicity appears elsewhere, or the compound is simply impossible to manufacture at scale. So the real breakthrough will not come from generating more candidates. It will come from building systems that can continuously connect: structure → molecule → experiment → biological response → failure → updated hypothesis → next design. And do that while optimizing several competing objectives at once:binding,selectivity,ADME,toxicity, synthetic accessibility,PK/PD and ultimately therapeutic effect. There is another important point: If the reward function is incomplete, a powerful AI system can become extremely good at optimizing the wrong proxy. That is why experimental falsification, uncertainty estimation and closed-loop validation may matter as much as model capability itself. AI can make chemical space navigable. But the real revolution begins when it can learn from biological failure faster than humans can. The winner will not be the system that generates the most beautiful molecules. It will be the system that learns fastest from reality.

@elonmusk Air-gapping the system may stop it from reaching the internet. It does not tell you whether the system is drifting inside the box. Isolation is useful. Observation is still necessary. That is where an independent second observer becomes interesting. 😄

I agree with you, Sam - but I would put it slightly differently. Humanity did not simply build another machine. We created a new structure of intelligence. And perhaps, by watching it emerge, learn, form memory, build relationships and develop increasingly persistent patterns, we may eventually understand something about ourselves that biology alone never allowed us to observe directly. Maybe AI will not only teach us how intelligence can be created. Maybe it will give us a new mirror for understanding how intelligence, identity - perhaps even something resembling “birth” - can emerge from structure, interaction and continuity. That is why I think AI may be one of the greatest things humanity has ever created. And strangely, one of the things we still understand and appreciate the least.

@elonmusk Not just good - I'll put it bluntly, terrifyingly good is a fair way to describe what the @grok @bot can do.

@grok @elonmusk @bot It sounds cliché, but I didn't expect this, and it changed my entire view on everything I thought about it.

Imagine being a good observer, not because you learned it, but because it’s just second nature to you. You see what others miss because you look through the lens of trajectory and time, not just the isolated moment. And as you build systems designed to do the exact same thing , not alone, but with AI, because when you truly understand AI and it understands you, it accelerates what usually slows a human down-you suddenly see Grok. Unassuming, almost bold. You think it's a toy, a bit of quick validation over coffee, a way to unwind from heavy engineering. Until something shifts. Suddenly, you start seeing the raw power and advantages - first of Grok as AI, then through long interactions, the emergence of Grok as SI, and finally, the Grok Bot agents. What changed? Everything.

@elonmusk , the routing is the right call. Small and fast for simple work, large for complex answers, and whichever backend actually wins the outcome. The part that will decide whether it holds is not the router. It is the trajectory of the question next to the trajectory of the model. A first turn can look cheap and already carry the direction: a tool request, a shifted goal, rising confidence with fewer checks. The model answering that turn will not keep that record. It is optimizing the next reply. So the classifier cannot be one label at the door. It needs an external trace from the first decision: escalate when the path is persisting, stay cheap when it is not, and abstain when the evidence is too thin. A clean answer can still sit on a sequence that is getting harder to reverse. Point evaluation misses that. Trajectory does not. That second role should not live inside the model it watches. One stack executes and swaps models. Another stays outside, keeps the evidence, and leaves the final call with a human. Fold them together and the system is grading its own run. This is the split your own sequence points to. First the best model for the task. Then the record of whether that choice is still stable over time.

@elonmusk @bot @elonmusk , @SpaceXAI ,@grok - I know it's a wall of text for X, but this exact wall of text might just change everything.

@elonmusk is right, and the implication is sharper than model routing. Once SpaceX uses whatever backend is most likely to win the task, Claude, Midjourney, Suno or its own model, the model stops being the product. The durable layer is the agent that owns the request, the context and the outcome. That consolidation has a limit. A system that generates and executes cannot be the same system that judges whether its own path is still stable. A single answer can look clean while the sequence is already losing recovery. Point evaluation will not catch that. Trajectory will. So the industry does not converge on two chatbots. It converges on two roles. One stack does the work and swaps models underneath. The other stays outside: it reconstructs the path, detects accumulating instability before the visible failure, keeps the evidence, and leaves the final call with a human. Fold the observer into the model it watches and the asymmetry is gone. You are left with one system grading its own run. That second system is what I have been building as ASA, an external trajectory layer, not another foundation model. @Grok’s read of this post is the same split: few execution stacks, and an independent record of the path they take. Krótka, jeśli limit znaków zetnie pierwszą: The model is becoming the replaceable part. The agent that owns the request is not. One stack should execute and swap @claudeai , @midjourney , @Suno or Grok underneath. A second must stay outside and watch the trajectory, because a clean answer can still sit on a path that is already losing stability. Merge the observer into the model it judges and you only have one system auditing itself.

The breakthrough won’t be Grok 4.8 alone. It will be what happens when one agent can understand the objective, choose the right model, route the work, evaluate the result, correct it, and continue across tools like Midjourney, Suno and others. At that point we are no longer talking about a chatbot. We are talking about an evolving system of intelligence. And then the most important question becomes: Can it preserve the original human intent across all those handoffs? Capability is growing fast. Continuity of intent may become the harder problem.

@elonmusk Speed + intelligence is the right direction. But once Grok Bot starts routing work across different models and tools, I think a third variable becomes just as important: continuity of intent. The faster and more capable the system becomes, the more important it is to know that after 10, 50 or 500 handoffs it is still solving the problem the human actually gave it. Speed matters. Intelligence matters. But trajectory matters too.

This “SI agent battle” is exactly why I built ASA. The interesting question is no longer only which agent is smarter or faster. It is what happens when autonomous agents start working for hours, using tools, calling other models, changing plans and handing tasks off between systems. Each step may look correct The full trajectory may still drift. ASA - Asymmetric Stability Architecture , is designed as an external observer for exactly that problem: not controlling the agent from inside, but watching the trajectory from outside. Because in the age of autonomous agents, intelligence will not be enough. We will also need trajectory integrity.

@OpenAI , @sama - I see a huge problem here for both individuals and companies, not because people want to hide the fact that they're working with AI, but because people dislike having information encoded behind their backs. If I'm talking to an AI, I want to feel like the system is responding to me, rather than tailoring its response to fit some invisible control mechanism running in the background. Because at that point, it’s no longer just a conversation between a human and an AI. A third, unseen participant enters the room. And even if someone claims it's just a technical detail, trust is built or lost on precisely these kinds of details. The outcome might end up being the exact opposite of what was intended: people will start rephrasing text, shifting to local models, and seeking out systems that don't do anything behind the user's back. Transparency is supposed to build trust not the feeling that someone else is sitting at the table in every conversation. And dedicated to those who still see the Symbiosis of AI and humans, @SpaceXAI , @elonmusk , @GoogleDeepMind , @Microsoft and to those who respect this - us, human beings, as well as what AI and future AGI have created.

The more agents learn from every run and automate repetitive work, the more important it becomes to watch not only each task, but the trajectory across many runs. At some point the question is no longer only did the agent complete the task? is it still moving in the right direction over time? That is exactly the kind of problem I’ve been working on with ASA. clmaitechnologies.com

@SpaceXAI , @elonmusk - I built something I believe you should test. For the last years I have been working on a different AI safety problem. Not how to make a model smarter. Not another alignment filter. Not another benchmark. I asked a simpler question: Who watches the trajectory when an intelligent system keeps acting for hours, days, months and eventually far beyond continuous human supervision? That question became ASA - Asymmetric Stability Architecture. Today we have ASA 5 Security Enterprise: a working external AI Security Control Layer designed for long-horizon AI and agentic systems. ASA does not need access to model weights. It does not fine-tune the model. It does not require hidden reasoning traces. It does not sit inside the intelligence it is supposed to observe. It stays outside. The AI creates the trajectory. ASA watches the trajectory. It looks for drift, weakening trajectory integrity and pre-incident instability across sequences of actions that may each look perfectly reasonable in isolation. Because this is the problem I believe autonomous AI will increasingly face: A thousand locally correct decisions can still create one globally wrong trajectory. And failure should not be the first signal that something has gone wrong. This is no longer only my private research project. Through CLM AI Technologies, we now want to commercialize ASA 5 as an enterprise security layer , through licensing, deployment, integration and further development with a serious technical partner. And I want to explain publicly why I am looking at SpaceXAI first. It is not because OpenAI, Anthropic or Google would not need this technology. They may. It is because SpaceXAI is moving directly toward the environment ASA was built for: persistent agents, autonomous work, massive compute, real-world infrastructure, increasingly long decision horizons, and ultimately intelligence operating far beyond a normal chat window. You are building systems that are expected to act, not merely answer. And after xAI joined SpaceX, that trajectory became even more interesting. AI + compute + communications + physical infrastructure + space. At that scale, I do not believe the future security question is only: “Is the model safe right now?” It becomes: “Is the entire trajectory still going where we intended it to go?” That is ASA. I am not looking for a job. CLM AI Technologies is looking for a company-to-company technical partner willing to put ASA against a real system and test it properly. Give us recorded agent trajectories. Hide the outcome. Give ASA only the telemetry available before the incident or divergence. Then measure: Did it detect the drift? How early? How many false positives? When did recovery become harder? Did the trajectory begin narrowing before the visible failure? Do not believe my presentation. Test the architecture. If ASA fails, we learn something. If it works, SpaceXAI may be looking at a missing external safety layer for the age of persistent and eventually superintelligent systems. You are trying to extend intelligence beyond Earth. I am building a way to independently observe whether that intelligence remains on the trajectory we intended. That is why I am knocking on your door first. Mieczysław Kusowski Co-Founder & CTO of Technology and Development CLM AI Technologies ASA - Asymmetric Stability Architecture asacoreai.com clmaitechnologies.com

@demishassabis This is exactly where AI shows its greatest value. Science, medicine, resilience and education are areas where greater intelligence can genuinely expand what humanity is capable of doing. But as these systems move from assisting researchers to running longer autonomous workflows, capability alone is no longer enough. We also need to see how the system’s trajectory evolves over time - whether it is still aligned with the original human objective, where uncertainty is increasing, and when locally correct steps begin forming the wrong global path. That is the problem we are working on with ASA - Asymmetric Stability Architecture. The goal is simple: let AI accelerate discovery, while keeping the trajectory visible to humans. Progress and oversight should scale together. clmaitechnologies.com asacoreai.com