🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

🔬 Excited to announce our ICLR 2025 Workshop on *Weight Space Learning*! Applying deep learning to model weights could revolutionize how we develop, analyze & manipulate NNs. Submit by Feb 10 In-person @ Singapore + remote options ICLR 2026 #ICLR2025 weight-space-learning.github.io

⏳Final hours to submit to the 1st workshop on Weight Space Learning at #ICLR2025 @ICLR_conf❗ 🚀Explore learning directly from model weights! 🌐 weight-space-learning.github.io ⏰ Submit by Feb 10 2025 AoE

We just dropped a fun panel discussion with François Chollet Zenna Tavares and Kevin Ellis on program synthesis. This was a powerhouse conversation with three legends in the field, enjoy.

A crucial point that everyone should be internalizing: in the age of test-time search, it's pretty much always possible to reach any level of capability by simply expending more compute. So it’s not just about, "can you do it?". The key is how efficiently you can do it.

Today we are announcing ARC-AGI-2, an unsaturated frontier AGI benchmark that challenges AI reasoning systems (same relative ease for humans). Grand Prize: 85%, ~$0.42/task efficiency Current Performance: * Base LLMs: 0% * Reasoning Systems: <4%

The $1,000,000 ARC Prize 2025 competition is back! And introducing ARC-AGI-2 the only unbeaten benchmark (we're aware of) that remains easy for humans but now even harder for AI. New ideas are still needed to reach AGI. We've got lots of great updates for 2025 --

Excited to have Machine Learning Street Talk (Machine Learning Street Talk) as a launch partner for ARC-AGI-2, featuring a deep dive interview with co-founders Mike Knoop and François Chollet

The chart below is IMO the most important thing we published today (h/t Bryan Landers). It shows that "scaling up" existing ideas, even the latest AI reasoning systems with log-linear accuracy/compute characteristics, is insufficient for AGI. We still need some architectural or

ARC-AGI-2 is here — the new leading AI program synthesis benchmark. It challenges models to reason, abstract, and generalize — not just memorize. Design led by François Chollet with key contributions from the Ndea team.

Are You Smarter Than A.I.? An interactive article by The New York Times covers ARC Prize and François Chollet "Some experts predict that A.I. will surpass human intelligence within the next few years. Play this puzzle to see how far the machines have to go."

Interactive NYT feature on François Chollet & ARC Prize.

Beyond Brute Force: my conversation with star AI researcher François Chollet and Zapier’s Mike Knoop * ARC AGI 2 and why it breaks base LLMs * test-time adaptation * the search for true machine intelligence All links in reply to this post below

Ndea co-founder François Chollet featured in The Atlantic. "The Rise of Fluid Intelligence — François Chollet is on a quest to make AI a bit more human." A story about the increasing relevance of ARC-AGI by Matteo Wong.

Join our ML Theory group next week on April 17th as they welcome Konstantin Schürholt for a session on "Weight Space Learning: Treating Neural Network Weights as Data" Thanks to Anier Velasco Sotomayor, Thang Chu and Andrej Jovanović for organizing this event 🥳

Join us tomorrow, April 17th as we host Konstantin Schürholt for a session on "Weight Space Learning: Treating Neural Network Weights as Data" Learn more: cohere.com/events/cohere-…

Our inaugural workshop on weight space learning at ICLR in Singapore has started! Join us in Topaz 220-225 to talk all things NN weights and check weight-space-learning.github.io!

Text-to-LoRA (T2L) — a new hypernetwork by @SakanaAI, that: - Can compresses many existing LoRAs into itself. - Quickly generates new LoRA adapters for a big LM, using just a text description of a task. Once trained, it can create new LoRAs on the fly. Here's how T2L works:

.François Chollet spoke at Y Combinator AI Startup School about going from AI to AGI. Why did pretraining scaling hit a wall? Does Test-Time Adaptation scale to AGI? What's next for AI? Plus Ndea's goal to build a lifelong-learning, self-improving, DL-guided program synthesis engine.

Customizing Your LLMs in seconds using prompts🥳! Excited to share our latest work with HPC-AI Lab, VITA Group, Konstantin Schürholt, Yang You, Michael Bronstein, Damian Borth : Drag-and-Drop LLMs(DnD). 2 features: tuning-free, comparable or even better than full-shot tuning.(🧵1/8)

LLMs can synthesize many programs, but how should we search among them? New from Sakana AI - AB-MCTS frames code generation as an adaptive tree search, guided by external feedback. Beats baselines on synthesis benchmarks including ARC-AGI.