# The classical AI winter is here.. **By:** Denis O. **Published:** 2026-09-25T00:57:39.862Z **Source:** [LinkedIn](https://www.linkedin.com/posts/denis-o-b61a379a_ai-share-7509053473280237568-7smV) --- # Denis O.’s Post The classical AI winter is here.. Everybody is running around with Jev or RLCD like it is the new kid on the block. Fast decisions, probabilities, confidence, structured outputs. No giant stream of generated tokens. Ooooh, ahhhh! Lol.. RLCD stands for Reinforcement Learning for Calibrated Decisions. But strip away the marketing and it is the familiar machine learning and statistics.. train a model so its predicted probabilities are calibrated against observed outcomes. But probability calibration is not a new concept. Neither are classifiers, scorers, discriminative models, confidence estimation, structured prediction or decision functions. Yet somehow AI changed generations and forgot its own field. Of course a system designed to make bounded decisions can outperform an LLM at bounded decision making. LLMs were never designed to be stable autonomous decision engines in the first place. Go back to Attention Is All You Need. It was not a paper about AGI, autonomous intelligence or a universal cognitive architecture. The Transformer was introduced as a sequence transduction architecture and demonstrated on machine translation. Then we became hypnotized by fluency at scale.. Hyperscale it enough, feed it enough human output, and the statistical continuation over bullshit became so convincing that an entire industry started confusing linguistic competence with intelligence.. Then we wrapped it in tools, memory, RAG, planning loops and agent harnesses and tried to turn a probabilistic sequence generator into an autonomous cognitive system. And the industry response to Jev is hilarious. We are rediscovering classification, scoring, regression, state estimation, structured prediction and bounded probabilistic decision making all over again when we already had all of this for decades if not a century(!!). Jev may be useful engineering. Yet everybody is acting shocked that direct decision computation can be faster and more stable than asking a giant autoregressive LLM to generate tokens about the decision first. It feels like the AI industry is suffering from dementia.. It went through machine learning, neural networks, reinforcement learning, probabilistic models, expert systems, cognitive architectures, search and planning.. Transformers came around, scale exploded, and somehow everybody forgot where we had already been.. Now we are walking around the same tracks in circle and calling it major progress. The deeper problem is simple. Hyperscaling did not breach the intelligence wall. It gave us extraordinary generative NLP, structured coding and generally better learned distributions through curated human labeled datasets.. But scaling the NLP did not magically turn it into a fundamentally different form of intelligence or superduper AGI.. Stable autonomy under uncertainty is still out of reach for GenAI. So now the industry is moving backward while calling it 'forward' or Jev.. and that, my bros, is your standard classical AI winter. [#ai](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) My fave so far was a post arguing that Jev is over hyped and you should just use an LLM instead, then proceeding to describe how to set it up with multiple LLM calls per classification and (best of all) \*\*asking\*\* the LLM for the probabilities.. Hundreds of likes and comments are all "finally someone who knows ML".. Oh, and the author has "PhD in AI" in his tagline 🙃 Can't decide whether to laugh or cry over the current state of "AI enthusiasts". Very glad to see this post getting so much attention, engagement, re-posting. It deserves even more. Their not "suffering" from demetia, they were openly denying the past. They need to IPO, they need way more than a sh1t ton of money, twice actually, they need bailout, they cannot admit to these simple truths. One should never look down to “statistics”. The fact that someone is based on statistics is a feature, not a bug. Also, it is not merely “statistics”: a lot of work has been done to develop train both tractable and expressive functional approximators that are useful enough to find practical real world applications. Saying that state of art tools is merely “ML/Statistics” is like saying that a figth generation fighter jet is merely “control theory / material science”, implying that the same applies to a horse cart. Yes, sure, the same applies to a horse cart, but quantity and precision eventually becomes a quality on its own. Local models ✌️ Thanks for the masterclass! There's another way to see this transition. The human attention and energy that the question-answering and language ability of LLM brought to AI is unparalleled. And that did cause a lot of Harness level innovation, much more than pre-LLM ML had! Now the rediscovery of the classifer being a good Decision tool (albeit this one is Transformer based, which you haven't mentioned), adds to the momentum of this innovation. So yes, human progress didn't happen in a Logical fashion. We are PsychoLogical and not Logical [Denis O.](https://www.linkedin.com/in/denis-o-b61a379a?trk=public_post_comment-text) It's easy to oversimplify things and succumb to nihilism. However, Jev (or at least its popular open-source alternatives) managed to do what ML and Stats have failed: convince that cost-effective automation using modern AI architectures (albeit heavily tweaked) can yield value. How many data scientists / engineers do you know who can deliver the same value, without having to invest a lot of time to create bespoke solutions that attain this level of efficiency? Sure, Jev isn't the breakthrough some people make it out to be, but it is very good engineering, no doubt. Perhaps some people's over-reaction to it signifies the crux of the problem: the limited AI literacy in the world today. It may be based on some AI system, but it's not the next step of AI development. Even the creators of this model make it clear that it has a very particular (and limited) scope. Calling this AI winter is a bit harsh, IMO. Cheers This is all true and a very good point. I do want to bring two potential innovations in Jev to attention though: (1) it’s more general purpose than classical ML classifiers as it’s trained on all sort of binary decisions so it becomes a general purpose classifier which historically wasn’t as much the focus on the ML research field (still TBD as how good is at this and if we end up having to re train it for all sort of more specialized classification problems). (2) it’s input handles unstructured data more easily and friendly than most classifier methods in classical ML. What’s an even bigger problem than what you called dementia is that we also have partial reconstruction of our memory here as it’s extremely difficult to know how unique these innovations are as I think Jev is closed source and its method wasn’t even published so hard to know how much of this is just some clever way to call a classifier and smart tricks on the user interface vs some new model capability. - Today is a sad day for AI... Because Jev shows that apparently most of the field knows jack.. I see people with AI lead titles at some of the biggest orgs genuinely confused by its performance and by why it takes so 'few tokens', as if useful computation somehow has to involve a giant LLM generating thousands of tokens. You do not need tokens to compute that 2 + 2 = 4. You do not need to sample the answer. You do not need a probabilistic language model to establish the fact, and you definitely do not need to consume a lake of water doing it. Some things are simply known, computed, estimated or decided directly. That applies to a huge class of the statistical functions covered by Jev. Probability functions, confidence estimates, thresholding, scoring, optimization, regression, state evaluation, tree based methods, statistical inference and bounded decision functions have existed for decades. There is an enormous mathematical and machine learning toolbox for doing exactly this kind of work without asking an LLM to generate a book about it first. Yet somehow we spent the last few years routing every little thing through LLMs. Instead of asking what existing lightweight statistical function actually solves the problem, we started asking how to prompt an LLM to approximate it with tokens. Then we wrapped more software around the transformer, spent even more tokens managing the context, and eventually forgot there was ever another way to compute anything.. Now Jev comes along and people are shocked. Oh my God, how is it so fast? How does it use so few tokens? How can these tiny models make decisions without asking a giant LLM every single time? Because you never needed an LLM for those functions in the first place. That is what makes the industry reaction depressing. A bunch of people can take techniques from statistics and machine learning that have existed since the 1980s or even 1900s, apply them where they actually belong, and suddenly half the industry acts like it has witnessed some new fundamental breakthrough in AI. If by progress we mean finally stopping the use of an LLM for every problem whether it belongs there or not, then sure, that is progress. But this is not some magical post LLM discovery. It is the industry going in circles again. A new generation arrives, forgets the previous generation of techniques, spends years forcing everything through the fashionable architecture, eventually rediscovers the old tools and then celebrates the rediscovery as if somebody just invented computation. 'internet moment' 😭😭😭🤡 That is the sad part. Jev did not reveal some mysterious new law of AI. It revealed how many people working in AI apparently know jack.. Also this is why alternatives have such a hard time gaining traction and why we cant have nice things.. bc LLMs is all anyone knows. [#ai](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) - The age of harness, Jev and GenAI bros' sudden discovery and pivot to ML.. For a while the main GenAI thesis was that the autoregressively sampling LLM would eat everything. That giant hyperscaled LLMs would 'reason', classify, plan, retrieve, decide, use tools, and eventually become this fully autonomous intelligence layer (not). Scale the LLM, scale the context, scale the compute, and apparently autonomous intelligence would simply emerge from nothing but density and scale (it didnt).. Instead, more and more of the actual system has migrated outside the LLM because of train costs among other things. Memory, state, retrieval, planning loops, verification, context management, permissions and governance are outside. The loop that makes an 'agent' continue doing anything is outside the model itself. The LLM is only a stateless frozen function that is invoked by a harness, returns output, and waits to be invoked again.. So the AI competition is moving back to the harness. Jev is a great example of such a harness. Instead of asking an expensive generative model to make every tiny decision inside an agent loop, Jev specializes in bounded decisions. Give it state and ask it to choose among known alternatives, score something against a defined scale, or estimate the probability of a proposition. It does not need to write a book about which tool should be called. It just picks it using more efficient mechanisms.. And that is how GenAI bros have rediscovered ML 😁🤣🤣💀 Comes out if you stop insisting that an LLM must do absolutely everything bevause tokenmaxxing is so expensive, the rest of machine learning suddenly becomes useful again! Who could have seen this coming??.. Tool routing? Classification. Which document matters most? Ranking. Is this transaction suspicious? Classification or anomaly detection. Need fast classification or scoring over structured features? Hello XGBoost. How likely is this proposition? Probability estimation. All 200-440x faster than 'thinking' LLMs! bc Duh! None of these problems inherently require a generative language model to autoregressively manufacture tokens until it eventually emits an answer. Humanity spent decades developing specialized statistical and ML methods for exactly these bounded problems. So now harnesses improve by becoming heterogeneous. LLMs for NLP and generation. Specialized learned models for classification and ranking . Search and optimization where those work better. Classical ML where classical ML always worked. If bros keep going like this they will announce a revolutionary new architecture that fits a line through data (oooh) 😂 ..Eventually they may even rediscover SOAR and the 90s cognitive architectures. The irony of the GenAI bros pivot to deterministic methods and ML though is fantastic. After four years of treating AI and LLM as synonymous, the 'AI field' is suddenly rediscovering ML. And apparently this is an 'internet moment'(!!) no less. Lmao, bros. LMFAO [#ai](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) - [Denis O.](https://www.linkedin.com/in/denis-o-b61a379a?trk=public_post_feed-actor-name) Why modern AI is trapped or whats in a name.. We can call it AI, AGI, SI, BSI, RSI, or whatever delulu du jour concept comes next. But beneath all of those labels sits an implicit classical bound, and its name is 'combinatorial search'. That is what essentially all modern AI reduces to. Deep learning learns and searches representations from enormous amounts of data. LLMs sample and proximity search token continuations from those learned distributions. Reinforcement learning searches for policies that produce higher reward.. Even some of the best 'AI' systems we have today make this painfully clear. DeepMinds AlphaEvolve generates candidate programs, evaluates them, preserves stronger ones and uses evolutionary selection to generate the next population for a search.. Dream RSI accumulates discovery decision trees and improves the tree exploration policies deciding where the underlying system should search next.. Both can produce genuine discoveries and significant improvements. But underneath, both are extraordinarily sophisticated ways of performing a combinatorial search. Everything still executes on classical computing infrastructure too. GPUs changed the scale dramatically by making huge amounts of linear algebra and parallel evaluation practical, but they did not repeal computational complexity.. They only let us evaluate more candidates, train larger models, sample more paths and search vastly larger spaces.. but they do not make every state or solution computationally reachable. That is the fish bowl modern AI is stuck in. More data fills it with more 'knowledge'. Larger models build richer representations of that knowledge. Longer context exposes more of it at once. Memory preserves previous trajectories. Tools expand the available actions. Reinforcement improves selection. More GPUs increase the number of possibilities that can be evaluated. AlphaEvolve evolves better candidates available for search. Dream RSIi mproves the strategy used to search for them.. The fish bowl can become absolutely enormous, but it remains bounded by what classical search can practically reach.. This is why modern AI can appear to be 'recursively improving' while the underlying cognitive bus is effectively fixed. Generate, sample, evaluate, prune, rank, recombine, search again. When the answer is reachable within that space, this machinery can be spectacularly effective. When the required computation or representation lies outside it, searching harder is only an exercise in futility . The system simply continues exploring what is reachable from the knowledge it has. Hyperscaling search does not breach this bound. So we have built astonishing combinatorial search machines and surrounded them with increasingly sophisticated infrastructure. But making the fish bowl bigger and having it search better keeps us in the same paradigm. No matter the name or scale.. [#ai](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) - [Denis O.](https://www.linkedin.com/in/denis-o-b61a379a?trk=public_post_feed-actor-name) Why active inference or adaptation is the only economically viable way forward in AI.. People keep asking why active inference or adaptive approaches are not just theoretically appealing, but economically unavoidable.. The reason is not philosophy or ideology. It is based on pure economics, even after accounting for falling compute costs.. Yes, the cost per FLOP continues to drop. Hardware improves and infrastructure becomes more efficient. But capability demands still scale faster than unit costs fall. Training a serious deep learning (DL) system today typically costs $5 to 20 million in raw compute, depending on scale and duration. Once data acquisition, labeling, engineering labor, experimentation, evaluation, and deployment are included, a single meaningful training cycle often reaches $20 to 60 million.. That is not an exception.. That is the current modus operandi. The economic failure appears at the marginal improvements.. Each time you want better long horizon decision making, stronger robustness under distribution shift, or fewer failures in rare but high impact scenarios, you pay nearly the same $$$ for the training again. Even if compute costs drop by a factor of 2 or 3, training runs simply grow longer, datasets grow broader, and optimization budgets expand.. The curve does not reset. It reinflates. Returns also diminish rapidly. Early training buys obvious gains. Later training buys narrow, fragile improvements. Worse, those improvements often destabilize existing behavior.. Finetuning induces overfitting. New data biases outcomes. Previously learned capabilities disappear through catastrophic forgetting.. To compensate, teams freeze layers, replay historical datasets, distill prior checkpoints, and layer corrective measures just to preserve what was already paid for.. The system becomes more expensive to maintain as it 'improves'. This is not a flaw specific to LLMs alone. It is a property of DL as a class of methods. These systems do not internally represent the world as a dynamical system.. They do not reason over dynamic persistent trajectories, constraints, or irreversibility. They perform local pattern completion that is effective for perception and classification. It fails for decision making because the real world is not a static dataset. It is a branching space of futures where actions permanently constrain what is possible next. Autoregressive sampling cannot account for that structure, whether it is stochastic or greedy. Active inference or adaptive AI methods change the economics because capability no longer comes from retraining. These systems maintain a live model of the world expressed as structure and adjust behavior by updating beliefs in real time. There is no universe where software built for translation and perception becomes alive, autonomous, or cures cancer simply because GPUs got cheaper. Intelligence that acts in the world must adapt continuously, or it will never be economically viable. [#ai](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) - [Denis O.](https://www.linkedin.com/in/denis-o-b61a379a?trk=public_post_feed-actor-name) Current 'frontier' AI increasingly reminds me of beating a dead horse and then wondering why it will not come alive... See you can build the most elaborate harness imaginable around the fixed frozen LLM dead horse. You can add memory, tools, retrieval, planning, verification, context management and layers of software that decide what happens next. You can keep improving the harness indefinitely, but the harness does not change what is underneath it.. A better harness around a dead horse is still a harness around a dead horse. The dead horse does not become alive or gain agency because the harness became more complex. That is essentially what we have done with the LLMs. The LLMs are deep learning neural networks (shocking, I know), so they provide the neural part of what has effectively become nouveau neurosymbolic AI. The harness now provides the symbolic part. In many ways this is a remake of the 1990s cognitive architectures with a new wrapper. Now we are building cognitive architectures around a new shiny but still dead horse.. The surrounding software stores state, retrieves memory, calls tools, injects context, evaluates outputs and creates the loop by invoking the model repeatedly. Then we call this concoction 'agentic AI', even though the supposed 'agency' is imposed externally.. There is no persistent recursion or inner loop continuously running inside the model itself. Only the dead harness has the loop, memory and state.. While the LLM horse is still dead too. None of this produces anything like actual stable autonomy either.. Context accumulates, assumptions get compressed or lost, errors propagate and the system drifts. Eventually we add more retrieval, checkpoints, verification, context engineering and exogenous orchestration to keep it all pointed at the task, while powered by hopium. Every time the dead horse fails to make correct move on its own, we respond by building a more elaborate harness around it. The consciousness claims take the same dead horse confusion even further.. Give the harness memory and suddenly people see a persistent self. Give the harness a software loop and they see ' revursive' agency. Let the dead horse NLP output 'I am afraid' and people start discussing whether it is suffering.. But perception plus software does not magically produce an inner experience, awareness or magikal AGI unicorn.. An externally implemented loop does not become an inner loop merely because the dead horse unwillingly participates in it, and externally maintained state does not become a persistent self because another autoregressive inference call receives it as context. We are confusing increasingly elaborate machinery around the dead horse with the dead horse somehow coming alive.. Add more compute, longer context, better memory, more tools or another hundred layers of cognitive architecture if you want.. We can build the greatest harness humanity has ever seen . But the dead LLM horse inside is still dead. [#ai](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) - RL is no magic panacea for frozen DL.. Folks like to make claims with RL, they wrap their frozen Euclidean latent space in some faux reward techniques and call it AI, no less. While RL, or reinforcement learning, has its uses, let us be real, it is a hack. It is a hack to adjust for the fact that whatever DL model you like to use, including LLMs, is effectively frozen or incapable of updating its internal state in any meaningful way. The weights can be adjusted, but the topology, or how data is encoded relative to other encoded data, is effectively static.. The latent space is fixed after training. You are not learning new structure, you are navigating within an existing encoding. To actually change that encoding, you need retraining.. That is the core limitation RL is trying to obscure. So RL shows up as a workaround. Not as learning in any structural sense, but as optimization over a fixed representation. You define a faux reward function, you run policy updates, you maximize expected return.. But nothing in that loop changes the underlying state of the model. It does not reconfigure how concepts are represented, it just biases trajectories through what is already there. That is why RL is fundamentally limited. The policy can only exploit features that already exist in the model RL is coupled with.. If the representation is missing something, no amount of reward shaping will create somethig out of nothing.. You can push harder, tune objectives, run more rollouts, but you are still confined to the same latent manifold. And this is where it breaks in practice.. Reward functions are proxies, not truth. The system optimizes the proxy, not the real objective. This leads to reward hacking, specification gaming, and fragile behavior.. It finds shortcuts that satisfy the signal while violating intent, because it has no grounding beyond the reward.. Credit assignment is noisy, gradients have high variance, exploration is inefficient, convergence depends on fragile hyperparameters. Small changes in reward or environment can produce large, unpredictable oscillations or shifts in behavior.. There is no inherent robustness because there is no internal restructuring.. When you stack RL on LLMs, you are optimizing outputs of a system trained for next token prediction. You align responses to feedback, but you do not give the system the ability to update its internal model of the world.. You are shaping outputs, but not enabling adaptation.. To learn anything genuinely new, to change structure, you still need to RETRAIN. RL does not remove that requirement. And despite what every 'AI expert' with an RL wrapper is going to claim, this is not a path to real autonomy. RL does not fix frozen DL, it works around it. .It pushes systems toward faux rewards inside a fixed space, and when those rewards stop aligning with reality, the system collapses right back into the same local minima it started from, calling that limitation 'learning'. [#ai](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) - The False Promise of AGI with LLMs.. There is a quiet sleight of hand happening in the AI industry, and most people do not notice it because the output looks impressive. The loose variance hack in modern LLMs, aggressively marketed by todays dominant Misanthropic and ClosedAI labs as AI or AGI even, is genuinely useful. It is an extraordinary slop generation engine.. High entropy, plausible syntax, overconfident tone.. Occasionally it even gets CRUD right on the first try. As a prototyping assistant, it can be very useful. As a generator of senseless, soulless corporate speak emails, it is perfect. As a vibes coding machine, it excels.. Just do not ask it for incremental updates. That is where the illusion befins to collapse.. The moment you try to build on prior work, refine behavior, preserve constraints, or maintain structural intent, coherence starts to decay. The model does not accumulate competence. It resamples it. Each interaction is a fresh draw from a distribution that has no obligation to remain consistent over time.. That is not intelligence. That is statistical fluency without real memory mechanisms that are not symbolic. And no, when I say AI, I do not mean Automated Immaturity or Absence of Intelligence. I mean the thing we historically meant by the word. Adaptive autonomy and decision making under uncertainty. The definition that sits very close to how DARPA frames Wave 3. Systems that can act, adapt, and remain stable or coherent under moderate to severe uncertainty without human in the loop.. This is where the LLM AGI narrative breaks. No language model built on perceptrons or deep learning alone can do this consistently without drift. Drift here is not a bug. It is a structural or natural consequence of the architecture itself. Every deep learning model eventually requires retraining to restore performance, recover alignment, or bolt on incremental capability at the same or higher retrain cost.. That retraining is not learning in the operational sense. It is a full reset. An erasure followed by statistical reconditioning or finetuning thatvalso comes at a cost.. The core benchmark for intelligence is not literary eloquence, leaderboard scores, or polished bs demos.. It is adaptive behavior that remains coherent over time. Systems that can update internally without collapsing their own priors. Systems that do not require periodic memory wipes to stay useful. Systems whose competence compounds instead of decohering. On that benchmark, AGI via LLMs is not just delayed. It is unachievable.. Not someday. Not with bigger models. Not with more GPUs.. You cannot scale your way out of drift when drift is the architecture. LLMs are powerful tools. But tools are not minds or autonomy.. And fluent slop, no matter how convincing, is not autonomy. Happy New Year! [#ai](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) - Everyone is running around talking about 'AI progress' like we have cracked something fundamental. and then you look at ARC AGI 3 results.. Since the official release and some last minute increases in complexity, no model is scoring above 0.5%.. Not 5% or 1%. Half a %.. Let THAT sink in! Even most humans are struggling to get anywhere close to ARC AGI 3 'human baseline' but that is a whole other story.. ARC is not some edge benchmark or academic curiosity or a benchmark you can simply pattern match your way through with an LLM or some overfit data patterns. This is the first environment that actually forces systems to explore, adapt, and compose actions in real time instead of regurgitating patterns from a frozen distribution. And the result is BRUTAL and brutally simple.. there is no reasoning there in LLMs or any RL + Deep Learning stacks.. What we call 'AI' today gets ripped to shreds and collapses the moment you go from prompting for quiche to interaction with complex reality under severe uncertainty. What we are seeing in this performance is token trajectory drift. These systems are sampling paths in a flat Euclidean latent space, hoping to latch onto statistical attractors. There is no stable structure, no persistence of belief, no coherent evolution of state. So what happens?? Oscillation. Thrashing. Random walks and flood fill dressed up as intelligence. And ARCAGI 3 exposes that instantly. People are already talking about harnesses, wrappers, tricks.. Greg Kamradt himself pitched in on Twitter on how he would solve this lol… cool. Where is it then?? WHERE? It has been a week. Where are the results if its just a harness problem? Because right now all we have are external crutches trying to compensate for the absence of an actual reasoning or intelligence or even effective perception(!). We dont need to solve 'consciousness' to test for reasoning. We just need an environment that forces abstract reasoning under uncertainty. ARC AGI 3 does exactly that .. and the same systems people are calling 'early AGI' collapse to effectively zero meaningful performance. Not degraded. Not struggling. Effectively 0.. Not even a 5% margin of error! (Compare that to 3/3 WIN for Seed IQ with performance exceeding human baseline) Because continuation over labeled corpora and occasionally producing something that compiles is not intelligence. It is trajectory sampling over a learned distribution.. Sometimes it works. Sometimes it makes sense. Sometimes it breaks. Sometimes it ships. Sometimes it leaks your code.. But there is no persistence, no grounded state, no causality, no understanding of consequences across time. ARC AGI 3 did not 'break AI' but it revealed it was never there to begin with. That GenAI and pattern matching at scale, while useful, is not intelligence or AI that can actually reason when there is no stable corpora or dataset to interpolate over. These are basics and yet so many fail to comprehend these very simple truths [#ai](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text)