Younes Bensouda Mourri

Google released a new 270M Gemma 3 variant last week, a lightweight, open-weight LLM that's perfect for local fine-tuning experiments and general tinkering. I couldn't resist and re-implemented it from scratch in pure PyTorch: https://lnkd.in/gSGzZvZv A few interesting tidbits about the Gemma 3 architecture: 1) One of the few recent architectures (besides gpt-oss) that uses Sliding Window Attention with a ratio of 5 local to 1 global layers 2) Dual RMSNorm: unlike standard Pre-Norm or Post-Norm transformers, Gemma 3 uses RMSNorm both before and after its attention + MLP blocks (I added the Qwen 3 figure showing the standard placement for comparison) 3) Vocabulary size: similar to other Gemma 3 models, it uses a huge vocabulary (~256k tokens), which should help with better multilingual coverage The code in my notebook requires about 1.49 GB of RAM (for inference) and runs comfortably on a laptop / Mac Mini. (Note that while it has good instruction-following capabilities given its small size, it's of course not a go-to model for local inference; the focus is more on being fast and lightweight for research and experimentation purposes.) Happy tinkering!

4,528

132 Comments

What is Function calling & MCP for LLMs? (explained with visuals and code) Before MCPs became popular, AI workflows relied on traditional Function Calling for tool access. Now, MCP is standardizing it for Agents/LLMs. The visual below explains how Function Calling and MCP work under the hood. Today, let's learn: - Function calling by building custom tools for Agents. - How MCPs help by building a local MCP client with mcp-use and using tools from Browserbase MCP server. In Function Calling: - The LLM receives a prompt. - The LLM decides the tool. - The programmer implements a procedure to accept a tool call request from the LLM and prepare a function call. The tool call request is found in the LLM's response when you prompt it. - A backend service executes the tool. This Function Calling takes place within our stack: - We host the tool. - We implement a logic to determine the tool to invoke and its parameters. - We execute it. So Function Calling requires us to wire everything manually. MCP simplifies this! Instead of hard-wiring tools, MCP: - Standardizes defining, hosting, and exposing tools. - Makes it easy to discover tools, understand schemas, and use them. - Demands approval before invoking them. - Detaches implementation from consumption. For instance, whenever you integrate an MCP server, you never write a line of Python code to integrate the tools. Instead, you just integrate the MCP server and everything beyond this follows a standard protocol handled by the MCP client and the LLM: - They identify the MCP tool. - They prepare the input argument. - They invoke the tool. - They use the tool’s output to generate a response. Everything happens through a standard (but abstracted) protocol. So here’s the key point: MCP and Function Calling are not in conflict. They’re two sides of the same workflow. - Function Calling helps an LLM decide what it wants to do. - MCP ensures that tools are reliably available, discoverable, and executable, without you needing to custom-integrate everything. For example, an agent might say, “I need to search the web,” using function calling. That request can be routed through MCP to select from available web search tools, invoke the correct one, and return the result. Check the workflow in the diagram below. In this setup, to build a local MCP client, I used mcp-use because it lets us connect any LLM to MCP servers & build private MCP clients, unlike Claude/Cursor. - Compatible with Ollama & LangChain - Stream Agent output async - Built-in debugging mode, etc Find the mcp-use GitHub repo in the comments! ____ Find me → Avi Chawla Every day, I share tutorials and insights on DS, ML, LLMs, and RAGs.

269

11 Comments

An AI story with Yann LeCun. A journey through neural networks, deep learning, and open-source foundation models, as well as one of sailing, building flying machines, and Parisian brasseries. A documentary by Barbara Rosario for her AI Stories series. https://lnkd.in/eXt7DNqT

1,365

42 Comments

📝 Announcing our paper that proposes a unified cognitive and computational framework for Artificial General Intelligence (AGI) -- going beyond token-level predictions -- one that emphasizes modular reasoning, memory, agentic behavior, and ethical alignment 🔹 𝐓𝐡𝐢𝐧𝐤𝐢𝐧𝐠 𝐁𝐞𝐲𝐨𝐧𝐝 𝐓𝐨𝐤𝐞𝐧𝐬: 𝐅𝐫𝐨𝐦 𝐁𝐫𝐚𝐢𝐧‑𝐈𝐧𝐬𝐩𝐢𝐫𝐞𝐝 𝐈𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞 𝐭𝐨 𝐂𝐨𝐠𝐧𝐢𝐭𝐢𝐯𝐞 𝐅𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧𝐬 𝐟𝐨𝐫 𝐀𝐫𝐭𝐢𝐟𝐢𝐜𝐢𝐚𝐥 𝐆𝐞𝐧𝐞𝐫𝐚𝐥 𝐈𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞 𝐚𝐧𝐝 𝐢𝐭𝐬 𝐒𝐨𝐜𝐢𝐞𝐭𝐚𝐥 𝐈𝐦𝐩𝐚𝐜𝐭 🔹 In collaboration with University of Central Florida, Cornell University, UT MD Anderson Cancer Center UTHealth Houston Graduate School of Biomedical Sciences, Toronto Metropolitan University, University of Oxford, Torrens University Australia, Obuda University, Amazon others. 🔹 Paper: https://lnkd.in/gqKUV4Mr ✍🏼 Authors: Rizwan Qureshi, Ranjan Sapkota, Abbas Shah, Amgad Muneer, Anas Zafar, Ashmal Vayani, Maged Shoman, PhD, Abdelrahman Eldaly, Kai Zhang, Ferhat Sadak, Shaina Raza, PhD, Xinqi Fan, Ravid Shwartz Ziv, Hong Yang, Vinija Jain, Aman Chadha, Manoj Karkee, @Jia Wu, Philip Torr, FREng, FRS, Seyedali Mirjalili ➡️ 𝐊𝐞𝐲 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬 𝐨𝐟 𝐓𝐡𝐢𝐧𝐤𝐢𝐧𝐠 𝐁𝐞𝐲𝐨𝐧𝐝 𝐓𝐨𝐤𝐞𝐧𝐬' 𝐂𝐨𝐠𝐧𝐢𝐭𝐢𝐯𝐞‑𝐂𝐨𝐦𝐩𝐮𝐭𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐀𝐆𝐈 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤: 🧠 𝐅𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤: Integrates cognitive neuroscience, psychology, and AI to define AGI via modular reasoning, persistent memory, agentic behavior, vision-language grounding, and embodied interaction. 🔗 𝐁𝐞𝐲𝐨𝐧𝐝 𝐓𝐨𝐤𝐞𝐧‑𝐏𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧: Critiques token-level models like GPT-4.5 and Claude 3.5, advocating for test-time adaptation, dynamic planning, and training-free grounding through retrieval-augmented agentic systems. 🚀 𝐑𝐨𝐚𝐝𝐦𝐚𝐩 𝐚𝐧𝐝 𝐂𝐨𝐧𝐭𝐫𝐢𝐛𝐮𝐭𝐢𝐨𝐧𝐬: Proposes a roadmap for AGI through neuro-symbolic learning, value alignment, multimodal cognition, and cognitive scaffolding for transparent, socially integrated systems.

221

9 Comments

the micro1 research team ran rubric-driven math evals across geometry, Olympiad reasoning, and prompt dependence. TLDR: xAI performed the worse and OpenAI performed the best. --- Here's an overview of the results: grok-4 expert finished last in every domain, scoring 3.9% on prompt dependence, 37.0% on geometry, and 15.2% on Olympiad tasks. GPT-5 set the benchmark with 84.8% in geometry and 40.3% in prompt dependence, while Gemini 2.5 Pro edged ahead on Olympiad. Overall, GPT-5 was the strongest performer, outpacing Gemini Claude 4 Opus, and Grok 4 Expert across the majority of tasks. full report coming soon.

238

15 Comments

🚨The AI inference race just shifted. The latest MLPerf results tell a very clear story: NVIDIA’s GB300 NVL72 set records on the new DeepSeek-R1 reasoning inference benchmarks with the Blackwell Ultra architecture. NVIDIA also set records on the new benchmarks this round: Llama 3.1 405B interactive, Llama 3.1 8B, and Whisper. What stood out to me technically: ✅ NVIDIA Blackwell Ultra architecture - designed for AI reasoning inference at scale ✅ NVFP4 acceleration - purpose-built 4-bit floating point format with improved accuracy ✅ Disaggregated serving - helped deliver nearly 50% throughput gains by separating context and generation processing 🤯 ✅ Full-stack optimizations - key to delivering performance records But the bigger picture was Rubin CPX at AI Infra Summit. It’s a new class of GPU purpose-built for long-context inference: 30 PFLOPS NVFP4 compute 128 GB GDDR7 memory 3× faster attention 6.5× performance gains over GB300 Why is this so big? Companies like Cursor (AI code editing), Magic (100M-token windows), and Runway (gen-video) are already planning integrations. The ecosystem is lining up: Azure, CoreWeave, Dell Technologies, Google, Oracle, Hewlett Packard Enterprise, Supermicro. The direction seems clear: inference performance gaps are widening, and the next wave of applications will demand hardware that keeps up. NVIDIA is leading the way, with Rubin CPX expected to be available at the end of 2026. 💡What’s your take... are we entering the era of specialized inference processors? #NVIDIABlackwell #NVFP4 NVIDIA AI NVIDIA Data Center #nvidiapartner

525

41 Comments

My colleague and friend Hartmut Neven once told me the best way to think about quantum computing was to imagine chopping an onion. You could do it the ‘traditional’ way: peeling the onion, later by layer, and chopping each layer in turn. That’s a classical computer.  But what if there was a different way? What if you could hold the whole onion and run your knife through it at once, getting the result but so much faster. That’s a quantum computer. Right now, ‘doing the whole onion' means quantum computer chips that can complete computations in under 5 minutes that would take today’s fastest supercomputers 10 septillion years. (To save you a Google: a septillion has 24 zeros ) When you combine computers of that speed and scale with AI, the possibilities are endless - from discovering new medicines quicker, to accelerating new types of energy like fusion. That’s the potential the Nobel judges recognised yesterday, when they awarded Google Quantum AI's Michel Devoret, his former Google colleague John Martinis, and Berkeley’s John Clarke the 2025 Nobel Prize for Physics for their pioneering discoveries in quantum. We already have their breakthroughs to thank for technologies we use every day, from mobile phones to cameras. Now they’re steering us towards the future with quantum - and it’s incredibly exciting to think that these are still just the early days. Huge congratulations to Michel, John and John! More in Harmut's blog: https://lnkd.in/dx5havXE (And glad to have the chance to revisit this photo with Hartmut, Demis and Eileen 🙂)

170

9 Comments

Stanford just published their 2025 Computer Vision lectures. For the first time since 2017, it’s finally here. For FREE. This is the course that was co-created by Andrej Karpathy and Dr. Fei Fei Li. 🔸 What you’ll learn: ↳Image classification & backprop ↳CNNs, detection, & segmentation ↳Transformers & self-supervised learning ↳Generative models: GANs, VAEs, diffusion 🔸 Why it matters: ↳The gold standard in computer vision ↳Where CNNs became accessible to everyone ↳Taught by Stanford’s leading researchers If you missed it then, don’t miss it now. (Playlist linked in the comments) ♻️ Repost to help someone learn AI the right way ➕ Follow me, Sairam, for practical AI learning that works ----- Want to go deeper? Gradient Ascent has you covered. Join 18k+ readers from Google, Meta, Netflix, and over 151+ countries worldwide

1,792

134 Comments

Some signs that Waymo has jumped on the end-to-end machine learning bandwagon. Junko Yoshida has a scoop from attending The Autonomous: signs that Waymo is doing tentative deployments of end-to-end machine learning in their fleet. Complete with a quote from Waymo that seems to support this (or at least does not deny it). Such a change has deep implications for scalability as well as safety. More about this in the article. I fully expect that Waymo will be smarter than a simplistic pixels-in/driving-out approach, as indicated in the story. But what that will look like is speculation at this point. https://lnkd.in/e4vTmRTs

88

9 Comments

🦾 Great milestone for open-source robotics: pi0 & pi0.5 by Physical Intelligence are now on Hugging Face, fully ported to PyTorch in LeRobot and validated side-by-side with OpenPI for everyone to experiment with, fine-tune & deploy in their robots! π₀.₅ is a Vision-Language-Action model which represents a significant evolution from π₀ to address a big challenge in robotics: open-world generalization. While robots can perform impressive tasks in controlled environments, π₀.₅ is designed to generalize to entirely new environments and situations that were never seen during training. Generalization must occur at multiple levels: - Physical Level: Understanding how to pick up a spoon (by the handle) or plate (by the edge), even with unseen objects in cluttered environments - Semantic Level: Understanding task semantics, where to put clothes and shoes (laundry hamper, not on the bed), and what tools are appropriate for cleaning spills - Environmental Level: Adapting to "messy" real-world environments like homes, grocery stores, offices, and hospitals The breakthrough innovation in π₀.₅ is co-training on heterogeneous data sources. The model learns from: - Multimodal Web Data: Image captioning, visual question answering, object detection - Verbal Instructions: Humans coaching robots through complex tasks step-by-step - Subtask Commands: High-level semantic behavior labels (e.g., "pick up the pillow" for an unmade bed) - Cross-Embodiment Robot Data: Data from various robot platforms with different capabilities - Multi-Environment Data: Static robots deployed across many different homes - Mobile Manipulation Data: ~400 hours of mobile robot demonstrations This diverse training mixture creates a "curriculum" that enables generalization across physical, visual, and semantic levels simultaneously. Huge thanks to the Physical Intelligence team & contributors Model: https://lnkd.in/eAEr7Yk6 LeRobot: https://lnkd.in/ehzQ3Mqy

377

10 Comments

You can fine-tune OpenAI’s gpt-oss models in just 10 mins with Unsloth AI’s Colab notebook. Unsloth speeds up training by 1.5×, cuts VRAM usage by 70%, and enables context lengths 10× longer – all without sacrificing accuracy. - You only need about 14 GB VRAM to fine-tune 20B with Unsloth, instead of 65 GB with other tools. This is possible because Unsloth uses custom kernels, memory offloading, and weight conversion tricks. - Prepare your dataset -- If you want to keep the model’s reasoning skills, include at least 75 percent reasoning examples and 25 percent direct-answer examples. - Run the notebook -- Unsloth handles loading, formatting, and saving the fine-tuned model automatically, ready for inference. Link to notebook: https://lnkd.in/gtzkwq5q #AI #GenAI #LLMs

55

1 Comment

🔥 Cool paper dropped today: “GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning” by Lakshya A Agrawal, Omar K. and others. Instead of using RL’s scalar rewards, GEPA leverages natural language feedback from system traces to reflectively evolve prompts. They got ~10% better performance than GRPO with 35x fewer rollouts. This supports a broader point that we’ve been focused on at Veris AI: when you have an environment that provides rich feedback, RL is just one optimization tool in your toolbox. The magic happens when your environment generates high quality structured, interpretable feedback that can be consumed in multiple ways. You don’t have to just collapse them into scalar rewards. Smart system design = flexible feedback mechanisms that unlock diverse optimization strategies. Side note: Berkeley folks have been killing it recently with everything going on at Skylab and BAIR. Go bears.

68

4 Comments