There's one great example that shows how physics can help upgrade AI world models, and it's LeHami. @ylecun, @randall_balestr and @RominaYalov brought Hamiltonian mechanics into JEPA to preserve phase-space volume and prevent latent collapse during long-term predictions. Why exactly Hamiltonian mechanics? • Variables are coordinates q and momenta p (not velocities like in Lagrangian mechanics), giving LeHami paired first-order equations for updating its states. • This makes Hamiltonian H a central function. It's often total energy: kinetic + potential, in classical mechanics. • LeHami learns this H function with kinetic-like K(p) and potential-like V(q, u) components, where u represents actions. • A symplectic leapfrog integrator uses this function to predict how states evolve depending on the actions taken, while no constrains are violated. To understand this better it's useful to think of a pendulum. It can pass the same position in either direction. But physical position alone won’t predict what comes next, you also need momentum. LeHami follows a similar principle, but works with learned representations of position and momentum. → And most importantly, it builds a constraint into how those predictions evolve. Maybe that's exactly the kind of thing world models like JEPA need now - one less way to go wrong. We've taken a much deeper look at LeHami's design and results in this guide. Plus, there's a little mechanics lesson for AI and physical AI folks that's worth learning →

A free, 24-hour online conference – "Engineering better intelligence" by @neo4j It's a great opportunity to learn the most practical approaches to agent memory, context engineering, GraphRAG, multi-agent systems, and intelligent applications from your peers. What’s on the agenda? - @AndrewYNg keynote on what’s next for AI builders - 100+ live technical sessions from AI builders - 24 hours of programming across all time zones - Three tracks: AI Engineering, Modern Applications, and Data Intelligence + general sessions Date: Nov 12, 2026 Don't forget to register! →

Speaking of decision models, one interesting family just came out - Drex by @NaceAI. This family includes 4 open-source decision models, each of which skips text generation altogether and directly assigns probabilities to the available options. Drex explores 4 different ways to do this: 1. Three smaller models use a pointer architecture: It consists of a pretrained LM + a small head that compares option representations with the question. They process the context once and reuse it to answer multiple questions. This saves time and computation. 2. The largest, Drex Max, is a dense multimodal decision model It reprocesses the context for each question, using bidirectional attention to compare all (up to 255) options together before assigning probabilities. That costs more computation but gives better accuracy. Training is interesting too. Drex learns from labeled decisions across 11 domains: first supervised fine-tuning, then reinforcement learning from adjusted feedback (RLAF) that mixes verified answers with a teacher model’s probability estimates to calibrate confidence. Drex Max adds an extra vision-training stage. And a few interesting results which show that it works: - Speed: Drex Max takes 183 ms vs. 2s for GPT-6 Sol and 3.5s for Claude Opus 5.5. - Accuracy: Beats Jev 1.14 on several benchmarks and approaches frontier LLMs, though Claude Opus 5.5 still leads on Decision Bench v4. - Consistency: RLAF improved its benchmark score by 2 points, while consistency training cut answer flips from 6.3% to 4.2% when options were reordered.

Speaking of decision models, one interesting family just came out - Drex by @NaceAI. This family includes 4 open-source decision models, each of which skips text generation altogether and directly assigns probabilities to the available options. Drex explores 4 different ways to do this: 1. Three smaller models use a pointer architecture: It consists of a pretrained LM + a small head that compares option representations with the question. They process the context once and reuse it to answer multiple questions. This saves time and computation. 2. The largest, Drex Max, is a dense multimodal decision model It reprocesses the context for each question, using bidirectional attention to compare all (up to 255) options together before assigning probabilities. That costs more computation but gives better accuracy. Training is interesting too. Drex learns from labeled decisions across 11 domains: first supervised fine-tuning, then reinforcement learning from adjusted feedback (RLAF) that mixes verified answers with a teacher model’s probability estimates to calibrate confidence. Drex Max adds an extra vision-training stage. And a few interesting results which show that it works: - Speed: Drex Max takes 183 ms vs. 2s for GPT-6 Sol and 3.5s for Claude Opus 5.5. - Accuracy: Beats Jev 1.14 on several benchmarks and approaches frontier LLMs, though Claude Opus 5.5 still leads on Decision Bench v4. - Consistency: RLAF improved its benchmark score by 2 points, while consistency training cut answer flips from 6.3% to 4.2% when options were reordered.

There's one great example that shows how physics can help upgrade AI world models, and it's LeHami. @ylecun, @randall_balestr and @RominaYalov brought Hamiltonian mechanics into JEPA to preserve phase-space volume and prevent latent collapse during long-term predictions. Why exactly Hamiltonian mechanics? • Variables are coordinates q and momenta p (not velocities like in Lagrangian mechanics), giving LeHami paired first-order equations for updating its states. • This makes Hamiltonian H a central function. It's often total energy: kinetic + potential, in classical mechanics. • LeHami learns this H function with kinetic-like K(p) and potential-like V(q, u) components, where u represents actions. • A symplectic leapfrog integrator uses this function to predict how states evolve depending on the actions taken, while no constrains are violated. To understand this better it's useful to think of a pendulum. It can pass the same position in either direction. But physical position alone won’t predict what comes next, you also need momentum. LeHami follows a similar principle, but works with learned representations of position and momentum. → And most importantly, it builds a constraint into how those predictions evolve. Maybe that's exactly the kind of thing world models like JEPA need now - one less way to go wrong. We've taken a much deeper look at LeHami's design and results in this guide. Plus, there's a little mechanics lesson for AI and physical AI folks that's worth learning →

Open-source AI stack from DeepSeek for building your own agent: ▪️ DeepSeek Harness (dsh): Connects a model to tools and manages agent sessions. ▪️ DeepSeek-V4.1-Flash and other models ▪️ DeepInfer: An inference engine for running models. ▪️ 3FS: A distributed file system for storing and moving data ▪️ deepseek-recipe: Translates API requests into model prompts and parses responses ▪️ FlashMLA: Optimized attention kernels ▪️ DeepGEMM: Optimized GPU kernels for matrix operations ▪️ DeepEP: Moves data between accelerators ▪️ DeepJIT: Compiles, caches, and loads GPU kernels at runtime ▪️ TileKernels: Optimized kernels for operations like routing and quantization Read our full guide to how the pieces work and fit together, what it takes to run them, and what’s still missing from the stack:

I models can learn to read without a tokenizer. It's called in-context tokenization: context changes how the model groups information into chunks. And the fact that it can work for DNA prediction is mind-blowing! @cartesia gave its new H-Net model a made-up word. At first, it broke it into little pieces. But the more it saw the word, the more of it it grouped together. The same thing happened with patterns and names that kept showing up in code. Then Cartesia trained a separate H-Net on bacterial DNA, and showed it a human gene. And somehow, it picked up on the gene’s structure too. That's wild. So, here's a question: Can we build systems that learn how to organize each kind of data for themselves? Let's see ↓ In this video, I also get into how H-Net works, how it compares to SmolLM3, and why it needed much less training data to reach similar results ->

I models can learn to read without a tokenizer. It's called in-context tokenization: context changes how the model groups information into chunks. And the fact that it can work for DNA prediction is mind-blowing! @cartesia gave its new H-Net model a made-up word. At first, it broke it into little pieces. But the more it saw the word, the more of it it grouped together. The same thing happened with patterns and names that kept showing up in code. Then Cartesia trained a separate H-Net on bacterial DNA, and showed it a human gene. And somehow, it picked up on the gene’s structure too. That's wild. So, here's a question: Can we build systems that learn how to organize each kind of data for themselves? Let's see ↓ In this video, I also get into how H-Net works, how it compares to SmolLM3, and why it needed much less training data to reach similar results ->

Open-source AI stack from DeepSeek for building your own agent: ▪️ DeepSeek Harness (dsh): Connects a model to tools and manages agent sessions. ▪️ DeepSeek-V4.1-Flash and other models ▪️ DeepInfer: An inference engine for running models. ▪️ 3FS: A distributed file system for storing and moving data ▪️ deepseek-recipe: Translates API requests into model prompts and parses responses ▪️ FlashMLA: Optimized attention kernels ▪️ DeepGEMM: Optimized GPU kernels for matrix operations ▪️ DeepEP: Moves data between accelerators ▪️ DeepJIT: Compiles, caches, and loads GPU kernels at runtime ▪️ TileKernels: Optimized kernels for operations like routing and quantization Read our full guide to how the pieces work and fit together, what it takes to run them, and what’s still missing from the stack:

Could one model help build another model that makes decisions? Here are 3 approaches worth exploring ↓ 1. Strands Decider Models help produce and check training material: The team replaces Qwen3.5-2B’s language-generation head with a pointer head that scores supplied options. One Qwen generates questions, a larger Qwen checks them, and a Gemma model supplies teacher signals for a later recipe. People choose the recipe and tests. 2. SII-Decision Rreasoning helps teach a fast readout. A Qwen model reasons through questions, then learns to approximate those answer probabilities without repeating the reasoning. Some benefit carries over to classification, but much less to difficult multistep tasks. 3. Distil– a coding agent fixes its teacher’s labeling mistake Nemotron treated one warning sign as several independent confirmations. The agent clarified the instructions and had it relabel the data. Distil reports that a Qwen student’s score rose from 72.6% to 78.5% against model-generated reference answers. This is AI Builds AI in a form we can inspect. Models help produce and check training material; people choose the recipe and tests. What else could a model contribute? Here are other interesting cases and where they fit in the bigger intelligence puzzle

An agent can remember a conversation. But what about the visual context? Oracle AI Agent Memory 26.8 gives agents image-aware memory. It lets them keep screenshots, receipts, and product photos as persistent context across sessions → The original image lives in @Oracle AI Database alongside its description and metadata. ▪️ The key detail → search works through text, not pixel similarity. Raw image bytes stay out of default search results and readback, keeping routine retrieval lightweight. This means that: - Visual evidence survives across sessions. - Agents can find stored images through natural-language queries. - Applications can retrieve the originals for inspection, display, or forwarding. Messages preserve text and image parts in their original order. Say a customer returns to a support chat about checkout timeouts. The agent can find the earlier screenshot through its saved description, and the application can load the original for a closer look. Oracle’s Daniela Pavlenco walks through the workflow with a synthetic support screenshot and a runnable notebook:

Must-read papers of the week ▪️ Harness-Model Collaboration for Long-Horizon Continual Learning ▪️ Rollback: Experience Backtracking in World Models for Policy Improvement ▪️ CausalWAM: Learning the Causes of Actions in World Action Models ▪️ MiRL: Error-Targeted Reinforcement Learning for Correcting Terminal Agent Misbehaviors ▪️ WebAnything: From the Wild Web to Scalable Agent Environments ▪️ DecoyCount: Evaluating Visual-Reasoning Misalignment in Counting ▪️ More Memory, Better Videos? A Controlled Study of Visual Memory in Long-Video Generation ▪️ Can LLMs Uncover Hidden Physical Laws through Free Exploration? ▪️ TAO: Tool-Aware Joint Scheduling for LLM-Based Agentic Serving ▪️ REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control ▪️ ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning ▪️ Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs Explore them to keep up with main AI trends. Here’s also the full list of stunning papers + links and our weekly AI news digest:

10 best sources to explore harness (studies and videos) ▪️ Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents ▪️ AI Models Are Eating Their Harnesses. Now What? ▪️ Why DeepSeek Harness Is The End Of Coding Agents as We Know Them ▪️ Rethinking Verification in Agent Harnesses ▪️ HarnessMM ▪️ Self-Evolving Agent Harness ▪️ Harness² ▪️ Object-Oriented Harness ▪️ Harness Scaling for Generalizable Agentic RL ▪️ Can LLMs Improve Their Own Harness? Find all the links and see what makes each one worth exploring:

A free, 24-hour online conference – "Engineering better intelligence" by @neo4j It's a great opportunity to learn the most practical approaches to agent memory, context engineering, GraphRAG, multi-agent systems, and intelligent applications from your peers. What’s on the agenda? - @AndrewYNg keynote on what’s next for AI builders - 100+ live technical sessions from AI builders - 24 hours of programming across all time zones - Three tracks: AI Engineering, Modern Applications, and Data Intelligence + general sessions Date: Nov 12, 2026 Don't forget to register! →

Must-read papers of the week ▪️ Harness-Model Collaboration for Long-Horizon Continual Learning ▪️ Rollback: Experience Backtracking in World Models for Policy Improvement ▪️ CausalWAM: Learning the Causes of Actions in World Action Models ▪️ MiRL: Error-Targeted Reinforcement Learning for Correcting Terminal Agent Misbehaviors ▪️ WebAnything: From the Wild Web to Scalable Agent Environments ▪️ DecoyCount: Evaluating Visual-Reasoning Misalignment in Counting ▪️ More Memory, Better Videos? A Controlled Study of Visual Memory in Long-Video Generation ▪️ Can LLMs Uncover Hidden Physical Laws through Free Exploration? ▪️ TAO: Tool-Aware Joint Scheduling for LLM-Based Agentic Serving ▪️ REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control ▪️ ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning ▪️ Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs Explore them to keep up with main AI trends. Here’s also the full list of stunning papers + links and our weekly AI news digest:

10 best sources to explore harness (studies and videos) ▪️ Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents ▪️ AI Models Are Eating Their Harnesses. Now What? ▪️ Why DeepSeek Harness Is The End Of Coding Agents as We Know Them ▪️ Rethinking Verification in Agent Harnesses ▪️ HarnessMM ▪️ Self-Evolving Agent Harness ▪️ Harness² ▪️ Object-Oriented Harness ▪️ Harness Scaling for Generalizable Agentic RL ▪️ Can LLMs Improve Their Own Harness? Find all the links and see what makes each one worth exploring:

An agent can remember a conversation. But what about the visual context? Oracle AI Agent Memory 26.8 gives agents image-aware memory. It lets them keep screenshots, receipts, and product photos as persistent context across sessions → The original image lives in @Oracle AI Database alongside its description and metadata. ▪️ The key detail → search works through text, not pixel similarity. Raw image bytes stay out of default search results and readback, keeping routine retrieval lightweight. This means that: - Visual evidence survives across sessions. - Agents can find stored images through natural-language queries. - Applications can retrieve the originals for inspection, display, or forwarding. Messages preserve text and image parts in their original order. Say a customer returns to a support chat about checkout timeouts. The agent can find the earlier screenshot through its saved description, and the application can load the original for a closer look. Oracle’s Daniela Pavlenco walks through the workflow with a synthetic support screenshot and a runnable notebook: