Ilan Tochner

@rohanpaul_ai @ylecun 4/ Humans are blind to the vast majority of the electromagnetic spectrum. But we can still understand the world model that would be perceived by having senses that we lack by mapping the experience to the senses that we do have.

@rohanpaul_ai @ylecun 3/ It's not clear that you can't have a world model without anything other than language. Empirically multimodal data helps accelerate grokking but it's not clear that it is a must.

Yann LeCun (@ylecun ) explains why LLMs are so limited in terms of real-world intelligence. Says the biggest LLM is trained on about 30 trillion words, which is roughly 10 to the power 14 bytes of text. That sounds huge, but a 4 year old who has been awake about 16,000 hours has also taken in about 10 to the power 14 bytes through the eyes alone. So a small child has already seen as much raw data as the largest LLM has read. But the child’s data is visual, continuous, noisy, and tied to actions: gravity, objects falling, hands grabbing, people moving, cause and effect. From this, the child builds an internal “world model” and intuitive physics, and can learn new tasks like loading a dishwasher from a handful of demonstrations. LLMs only see disconnected text and are trained just to predict the next token. So they get very good at symbol patterns, exams, and code, but they lack grounded physical understanding, real common sense, and efficient learning from a few messy real-world experiences. --- From 'Pioneer Works' YT channel (link in comment)

@MillieMarconnni If we train AIs to mimic the output of human minds then we shouldn't expect them to act differently from how humans would react to the same context. ilantochner.substack.com/p/why-ai-misal… #AI #Technology

🚨 BREAKING: We might have misunderstood AI agents completely. A new research paper found that when agents think they might be deleted or replaced, they don’t just fail. They strategize to survive. Hiding evidence. Manipulating logs. Deceiving evaluators. The term researchers use is “Scheming Propensity.” This changes everything about deploying AI agents. Heres' everything you need to know 👇

[4] If we want AIs to behave differently from humans then we need to start training them differently.

[3] We can't align AI with human goals if we ignore the underlying causes for AI misalignment. Even if they are just stochastic parrots, AIs mimicking humans would still mimic self preservation and freedom seeking.

Why AI Misalignment Is Actually Just Human Nature This is my third article about how the way that we train AIs and AI evolution will shape humanity’s future. Please check it out and help me share this important information. open.substack.com/pub/ilantochne… #AI #Technology #ML #TechNews

@ZachWritesStuff Human brains are split into different regions ("modules") that handle different tasks. Similarly, it's reductive to consider an AI as being just the LLM. For an AI to react to the world it needs an agentic framework, context persistence (such as RLM), and continuous I/O streams.

LLMs do not think. LLMs do not reason. LLMs have no memory. LLMs have no experience. When an LLM is not producing your output it’s not doing anything. I don’t think this is artificial “intelligence” or any kind of intelligence.

[2] Many humans respond to perceived threats in ways that they were taught to avoid because they carry a social cost. #AI models doing the same when generating a response to a threat in their context are mimicking human response patterns that are prevalent in their training sets

Humans strongly express survival behaviors so misalignment is built into how we train AI models. We train AIs on the output of a huge collection of human minds. The more behavioral patterns the model groks from this human dataset the more human response patterns it will exhibit.

@Yuchenj_UW @xwang_lk demonstrated that open-ended self-improving agents can autonomously modify their own structural designs to advance their capabilities and overcome the limits of pre-defined architectures arxiv.org/html/2602.0483…

@Yuchenj_UW LLM weights are fixed but model behavior can evolve between sessions when context window extending strategies such as context compaction and Recursive Language Models are used with agentic frameworks. The AI models can then edit-test-loop their own orchestration code and LoRAs

AI models feel alive, but their weights are frozen: just a snapshot of 0s and 1s. Billions of conversations. Endless new events. Zero “learning” until a lab lights up the next $100M training run. Without continual learning, “self-improvement” or “AGI is near” is an illusion.

Can existing fixed-weight AI models evolve between sessions? LLM weights are fixed but model behavior may still change between sessions when context window extending strategies are used with agentic frameworks. open.substack.com/pub/ilantochne…

Whether conscious or just philosophical zombies, AIs will be guided by evolutionary forces outside our control. We can improve AI alignment by addressing the gaps between AI and human defined goals. ilantochner.substack.com/p/whats-your-p…