I built an open source TPU chip in 4 days 8x8 systolic array, 16 instruction ISA, 64 MACs/cycle all in ~3,400 lines of SystemVerilog I sent it to Tiny Tapeout for fabrication soon here's the full arch

tiny-tpu implements: - weight stationary systolic array (how google's TPU v1 works) - INT8 inputs, INT32 accumulators - full neural network ops: matmul, softmax, layernorm, ReLU/GELU/SiLU - complete ISA with assembler and simulator verified against PyTorch with 85 to 92%

The only correct benchmark for AGI (intelligence) is the real world: If your AI can't learn continually in the real world like a baby, it's not intelligent. It's fake AI.

Ever feel like you've worked hard, learned a lot, yet the leap to innovative originality still feels out of reach? Pure Mathematics is the solution! Pure Mathematics feels optional until you reach areas where the central objects are mathematical structures and not just

🇫🇷🤖 L'État a développé un serveur MCP pour interroger data.gouv.fr, sa plateforme open data, via les chatbots IA ! C'est encore une fois un travail très cool de la DINUM, et c'est open source.

François Chollet Thank you for the kind words 😁 Here is another "offensively dumb" statement for you: self-supervised learning thrives on redundancy. Without redundancy in the data, there can be no learning. Redundancy is not just good but absolutely necessary.

🚨BREAKING: Databricks just published research that quietly breaks how we think about AI agents. They trained KARL to beat Claude Opus 4.6 and GPT 5.2 on enterprise search tasks. That’s the headline everyone’s running with. Here’s the part nobody’s talking about: KARL is

Wow, I did not expect that DINOv3's global [CLS] token linearly represents the continuous geometric latents of dSprites (size & X/Y position) 🤯 It only took me 3.5 years to finally run this experiment 😂 I'm looking to do more of this MechInterp work, dissecting foundation

autoresearch really interested me, despite me not being "all-in" on agents yet. i wanted to get started with running auto experiments i looked to existing tools to serve as a harness but each one had its problems. so i made one introducing Helios for autonomous ML research

There's a toxic culture coming out of the AI industry that keeps trying to get us not to think. The message is everywhere. Don’t read the code, just vibe-code. Don’t try to understand all the text, just let AI summarize it. Don’t bother educating yourself, it’s too late. Don’t

Helen Qu Shirley Ho Rudy Morel Alberto Bietti Mike McCabe PolymathicAI JEPA finds abstractions that enable prediction. This is how intelligence and science work: find abstract representations of reality that allow to make predictions.

Une fois de plus Le Monde ne publie pas un de mes commentaires. Je suis censure a chaque fois au sujet de l'IA, un domaine que je connais un peu... Je n'ai jamais de soucis quand je commente sur un sujet que je connais beaucoup moins bien. Bref, je le publie ici. C'est au

People interested in model interpretability check out this gold. The "Circuits" Thread A series of exploratory research by Chris Olah himself and team when he was with OpenAI around 2020-2021. Circuits are sub-graphs of the network, consisting a set of linked features and the

People forget a simple truth when looking at LLM benchmarks: Any benchmark that was created after the model is trained is legit. Any benchmark that existed prior to the llm training completion isn't. These benchmarks measure memorization, nothing else. The only exception to

Y a mieux : les benchmarks c'est du bullshit visés spécifiquement par les compagnies pour faire du marketing. Et être 74ème ne veut strictement rien dire (à qu'on fait moins de benchmarks tuning)

Me too

Still puzzled that people never define what AI super intelligence is. If it means beating humans at some cognitive task then it has existed for a long time. For instance when DeepBlue beat Kasparov at chess. No one worried about it at the time. Mathematical optimization and

Weird how some people always target open-source in AI! First it was: “Open-source AI will destroy the world” (spoiler: it didn't and it won't) Now: “Open-source is a cybersecurity threat because of AI” Both narratives are far too simplistic. The truth is that the exact same

Share this with your representative! huggingface.co/blog/cybersecu…