Aviv Nahon

"How One AI Model Creates a Physical Intuition of Its Environment" A nice article about our work with the Video-JEPA model showing how it learns a kind of "physical common sense." The model's prediction error shoots up whenever it is shown a video in which something impossible happens. The article has quoted from the lead author Quentin Garrido, collaborators like Randall Balestriero, scientists not connected to our work like neuroscientist Karl Friston and cognitive scientist Micha Heilbron. https://lnkd.in/e5uhG-_P

1,027

‏65‏ ‏תגובות‏

Scared of using LLMs in Python? It’s easier than you think. I procrastinated on using LLM for a long time, Even though I saw the power of LLM models in ✔️ Analyzing unstructured, long-form text ✔️ Turning data into plain-English summaries ✔️ Automating repeated tasks, like weekly reporting But it seemed so complex. It felt like “real developer” work. But then I tried it. And it literally takes <10 lines of code. So I’m sharing it here, so you can try it too! Currently, my favorite models to work with are: - Sonnet 4 - Gemini Flash 2.5 - GPT 5 (sometimes I switch back to 4o) ------ 𝗗𝗮𝘄𝗻'𝘀 𝗛𝗼𝘁 𝗧𝗶𝗽𝘀 𝗧𝗶𝗽 #1 never work backwards from the solution. Don’t “force” an LLM into your analyses. That’s adding unnecessary complexity. Make sure you have a good use case for an LLM, before doing the work. 𝗧𝗶𝗽 #2 Use providers, like OpenRouter to quickly switch between models Not sure which model best fits your use case? Use a provider like OpenRouter, so you can test out different models, by changing just the model field. 𝘗𝘚: 𝘵𝘩𝘪𝘴 𝘪𝘴 𝘸𝘩𝘢𝘵 𝘐 𝘶𝘴𝘦 𝘵𝘰 𝘲𝘶𝘪𝘤𝘬𝘭𝘺 𝘵𝘦𝘴𝘵 𝘥𝘪𝘧𝘧𝘦𝘳𝘦𝘯𝘵 𝘮𝘰𝘥𝘦𝘭𝘴 𝘧𝘰𝘳 Interview Master . 𝗧𝗶𝗽 #3  Always sanity check outputs LLMs can speed you up, but they still make mistakes. ------ ♻️ Was this useful? Repost it so others can see it too. 📩 I’m writing about this in more detail in my next newsletter. Article is going out on Wednesday. Sign up → www.askdatadawn.com

1,173

‏49‏ ‏תגובות‏

🚨 Excited to announce: LIBERO is now merged into LeRobot Hugging Face with multi-task evaluation support, train on LIBERO datasets, evaluate across 130+ tasks, and set new baselines for VLA models! LIBERO is a benchmark designed to study lifelong robot learning. The key idea: robots won’t just be pretrained once in a factory, they’ll need to keep learning and adapting with their human users over time. This ongoing adaptation, known as lifelong learning in decision making (LLDM), is a crucial step toward building robots that become truly personalized helpers. To make progress on this challenge, LIBERO provides standardized tasks that test knowledge transfer: how well can a robot apply what it has already learned to new situations? 🔬 LIBERO task suites ✓ LIBERO-Spatial: reasoning about spatial relationships; tasks vary by spatial arrangement but use similar objects. ✓ LIBERO-Object: manipulating different, diverse objects; tasks focus on object-centric transfer ✓ LIBERO-Goal: adapting to changing instructions/goals on the same objects; tasks vary by goal requirements rather than objects or spatial arrangement. ✓ LIBERO-90: 90 short-horizon tasks; quick, varied tasks ✓ LIBERO-Long: 10 complex multi-step tasks Together, they cover 130+ tasks, from simple manipulations to multi-step scenarios. 📌 What’s new in LeRobot: LIBERO is now the baseline benchmark for adding new Vision-Language-Action (VLA) models. You can train directly on our preprocessed LIBERO datasets (on the Hub). And you can evaluate your policy in LeRobot, tracking success rates across all suites. This makes LIBERO a dataset & a shared foundation for embodied AI research, one where the community can fairly compare methods and build on each other’s work. 📄 LIBERO paper: https://lnkd.in/erj963Zw 📊 LIBERO dataset on Hugging Face: https://lnkd.in/eYrTEcK3 📚 LeRobot libero docs: https://lnkd.in/ePJzS4FE We can’t wait to see what new VLAs the community builds and evaluates on this benchmark. Let’s push lifelong robot learning forward together 🤖✨ Pepijn Kooijmans Adil Zouitine Steven Palma Dana Aubakirova Michel Aractingi Francesco Capuano Hugging Face Thomas Wolf Clem Delangue 🤗 

72

‏1‏ ‏תגובה‏

🔥 Everyone’s chasing bigger models and faster inference… but almost no one is talking about logprobs — the hidden superpower in LLMs. Log probabilities don’t just tell you what the model said — they tell you how sure it was about every token. And that changes everything. Why this matters 👇 ✅ Catch low-confidence outputs before they derail workflows ✅ Add interpretability to an otherwise black-box model ✅ Unlock advanced tricks: hybrid routing, selective sampling, fact-checking Here’s the flow most people overlook: Logits → Probabilities → Logprobs → Sampling → Response Not all models expose them (Claude doesn’t). But OpenAI (GPT-4, GPT-4o) and Gemini (VertexAI) do — and if you’re building serious AI systems, it’s worth paying attention. 💡 Question: Have you already used logprobs in your pipeline — or is this still flying under your radar? #AI #LLM #MachineLearning #OpenAI #GenerativeAI #LangChain

315

‏27‏ ‏תגובות‏

Tool-Augmented Unified Retrieval Agent for AI Search Nice paper showing how to effectively extend RAG to support real-time, dynamic, and transactional queries through agentic tool use. In particular, I like the idea of the Distilled Agent Executor, which uses a small, latency-optimized agent fine-tuned via a mixed-rationale method (training with chain-of-thought but inferring without it). This achieves near-teacher-level performance with dramatically lower cost. For example, a distilled Qwen3-4B model outperforms GPT-4o and its own teacher model (Deepseek-V3) in tool-use accuracy (88.3% vs. 81.7%) while reducing latency from 6.8s to 750ms.

130

‏5‏ ‏תגובות‏

Excited to announce that our group has had two papers accepted at NeurIPS this year, covering two very different and exciting topics! 1. 𝐏𝐚𝐫𝐞𝐭𝐨𝐐: 𝐒𝐜𝐚𝐥𝐢𝐧𝐠 𝐋𝐚𝐰𝐬 𝐢𝐧 𝐄𝐱𝐭𝐫𝐞𝐦𝐞𝐥𝐲 𝐋𝐨𝐰-𝐛𝐢𝐭 𝐋𝐋𝐌 𝐐𝐮𝐚𝐧𝐭𝐢𝐳𝐚𝐭𝐢𝐨𝐧 ParetoQ is a framework designed to effectively quantize LLMs to sub-4-bit precision. It surpasses all previous methods tailored to specific bit widths. Our ternary 600M-parameter model outperforms the previous SoTA ternary 3B-parameter model proposed in 1-bit era paper, using only one-fifth of the parameters. Paper: https://lnkd.in/ga7rP7Qb Code: https://lnkd.in/eWPYFryC  Model: https://lnkd.in/gBcmsBnu PyTorch blog: https://lnkd.in/g_wy2cr5 Congratulations to Zechun Liu, Changsheng Zhao and all the co-authors! 2. 𝐀𝐮𝐭𝐨𝐏𝐚𝐫𝐭𝐆𝐞𝐧: 𝐀𝐮𝐭𝐨𝐠𝐫𝐞𝐬𝐬𝐢𝐯𝐞 3𝐃 𝐏𝐚𝐫𝐭 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫𝐲 AutoPartGen is an autoregressive model that generates complex 3D objects by producing one 3D part at a time and seamlessly assembling them into coherent final reconstructions. Leveraging a latent 3D representation, the model can automatically determine the number and type of parts based on various inputs, including 2D images, 2D part masks, or an existing 3D object. AutoPartGen achieves state-of-the-art performance in 3D part generation. Project: https://lnkd.in/gTUE_Ceb  Paper: https://lnkd.in/gtPHHd9B Congratulations to Minghao Chen, Andrea Vedaldi and all the co-authors!

207

‏1‏ ‏תגובה‏