Jan Chorowski

Bonsai Robotics is building the next generation of off-road autonomy, starting with a focus on agriculture. As its multimodal datasets from real-world deployments grew – and model sizes increased – fixed on-prem GPU resources became a bottleneck to innovation. With Anyscale, the team gained a managed, multi-cloud compute platform powered by Ray, scaling perception workloads across data processing and model training without the operational burden of managing clusters. Read the full story: https://lnkd.in/gnViUWSY

37

2 Comments

Meet EverMemOS, the memory framework behind Tanka! Today, it’s officially open-source. 🔹 Structured long-term memory (episodes, profiles, preferences, semantic memory) 🔹 Hybrid retrieval engine — BM25 + embeddings + RRF fusion 🔹 Agentic multi-round recall guided by LLM reasoning 🔹 Production-ready architecture built with Milvus, Elasticsearch, MongoDB, and Redis EverMemOS consolidates, understands, and retrieves memories like a real cognitive system, enabling Tanka to deliver deeper, more personalized conversations every day. If you’re building agents, AI products, or research systems, this is a powerful starting point. 👉 GitHub repo: https://lnkd.in/eGx9jguR #AI #Agents #AIMemory #AgentMemory

8

2 Comments

If you are interested in human-level AI, don't work on LLMs. This was one of Yann LeCun’s core messages during his September 16 Harvard CMSA talk on Self-Supervised Learning, JEPA, and the Future of AI. Key takeaways: → Text-only AI won’t get us to human-level intelligence Typical LLMs are trained on ~30 trillion tokens, an amount of text a human would need 450,000 years to read. Yet a 4-year-old child absorbs roughly the same volume of data in just 16,000 hours, and in far richer sensory form: vision, motion, physical interaction. Thus, intelligence needs redundancy, perception, and grounded experience, not just text. → Feed-forward inference is fundamentally limited Current models rely on a feed forward propagation through a fixed number of layers to perform inference. This is too weak for real reasoning. LeCun argues that intelligent systems need inference through optimization: searching for an output that minimizes an “energy” or incompatibility function (EBMs). This allows zero-shot reasoning without exhaustive training. → Predicting at pixel-level is wrong problem Predicting pixels, particularly future video frames directly, is mathematically intractable. The world is too high-entropy at the pixel level. LeCun’s proposal: JEPA (Joint Embedding Predictive Architecture), which predicts in representation space that captures relevant structure. Intelligence depends on abstract representations that allow us to predict, not reconstructing every detail. → Auto-regressive generation causes hallucinations by design LLMs predict the next token using their own previous predictions. Thus, errors accumulate and drift grows. Hallucination isn’t an accident; it’s baked into the mechanism. → Reinforcement Learning is useful but inefficient RL requires enormous numbers of trials. According to LeCun, it should be a tool of last resort, not the foundation of intelligence. → Humans and animals build layered world models. Machines don’t (yet) Infants form internal models of motion, inertia, and cause-and-effect within their first 9 months. This understanding is hierarchical, operating across multiple levels of abstraction, and enables humans to plan, predict outcomes, and act intentionally. Current AI systems lack this multi-level structure, limiting their ability to reason about actions or anticipate real-world consequences. This lecture made it clear that the path to human-level intelligence won’t likely come from scaling LLMs, but from controllable & safe systems capable of real reasoning, planning, understanding and learning world models from chaotic environment, just as humans and animals do. Watch the full lecture here: https://lnkd.in/eQMEkTEH #AI #ArtificialIntelligence #AIAcademyAsia #MachineLearning #DeepLearning #AIResearch #AIFuture #AISafety #YannLeCun

8

The irony of all of this is that LeCun is right - we need to be concentrating on world models right now, not language models. The metaverse (for all of the ridiculousness that came out of Meta's version of it) was in fact a very ham-fisted effort to do just that, but I think it was largely accidental. For Zuckerberg, it was just another big shiny, and he has a history of trying to pivot his companies around to those big shinies rather than concentrating on meaningful problems. There were several problems with Z's Metaverse, not least being that it was obviously built with the idea of blockchain (and hence digital currency) at its core. From a business standpoint, this meant that it was destined to be the equivalent of a very expensive theme park, where its whole raison d' etre was to shake every last nickel and dime out of its participants. This meant that concentrating on the abstraction of space and time, something that is pivotal to world models, took very much a back seat to figuring out how to harvest all of those gullible marks. A world model is very much built around the notion of persistence, which is a remarkably complex thing to achieve in the virtual world and wasn't solved in the short (incendiary) rise and fall of the metaverse. Now that more of the pieces are in place (yes, you need that language model in the mix, just not by itself) it may be that we're coming back to the issue of how we represent worlds (not just the physical world) in a consistent, standardized manner.

32

6 Comments