The artificial AI researcher is not reshuffling and repackaging known ideas, like in the humanities, but identifies a new research problem, plans out an experiment, conducts the experiment, and writes up the results. It does not borrow more than human AI researchers do.
MACHINE LEARNING
-
Top Canadian and UK Universities for AI Education
By
–
UBC, Alberta, Montreal, Toronto, Waterloo, McGill, Simon Fraser … so many amazing universities in Canada for anyone wanting to study AI. There’s outstanding places in the UK too: Oxford, Cambridge, Imperial, UCL, … and tons in Europe. Frankly from any of these places you can
-
State Space Models vs Transformers: Mechanistic Evaluation of Recall
By
–
new paper! 🫡 why are state space models (SSMs) worse than Transformers at recall over their context? this is a question about the mechanisms underlying model behaviour: therefore, we propose using mechanistic evaluations to answer it!
-
Evals and Benchmarks Don’t Capture What You Think
By
–
There’s also so many layers to most to really understand what they are capturing, it’s pretty interesting. Most evals and benchmarks do not capture what you would initially assume.
-
AI Evals and Benchmarks Capture Model Progress
By
–
I love AI evals and benchmarks, so cool to see interesting ways of capturing model progress
-
AlphaEvolve: DeepMind Applies Natural Selection to Self-Improving AI Code
By
–
Une IA qui s’auto-améliore, sélectionne ses meilleures versions, et recommence en boucle.
— VISION IA (@vision_ia) 28 mai 2025
DeepMind ne copie plus les humains. Elle copie Darwin.
Alpha Evolve, c’est la sélection naturelle appliquée au code.
➡️ Extrait de ma dernière vidéo sur Alpha Evolve : https://t.co/KGEjKJY0dF pic.twitter.com/6IrY4998kzUne IA qui s’auto-améliore, sélectionne ses meilleures versions, et recommence en boucle. DeepMind ne copie plus les humains. Elle copie Darwin.
Alpha Evolve, c’est la sélection naturelle appliquée au code. Extrait de ma dernière vidéo sur Alpha Evolve : -
AI Models’ Reasoning Capabilities and High-Stakes Limitations
By
–
AI isn’t ready for high-stakes contexts where precision is critical. In this year’s AI Index, we chart the progress of top models’ complex reasoning capabilities. See more insights from the technical chapter of the #AIIndex2025: https://
stanford.io/43w3rEo -
Mistral AI Releases Codestral Embed Model
By
–
Mistral AI released the new Codestral Embed model. "Codestral Embed significantly outperforms leading code embedders in the market today: Voyage Code 3, Cohere Embed v4.0 and OpenAI’s large embedding model."
-
Codestral Embed: New Code Embedder Model Released
By
–
Very happy to release Codestral Embed, our first code embedder. It can use up to 3072 dimensions, ordered by relevance. Using embeddings of 256 dimensions with int8 precision is already sufficient to outperform all existing models.
-
Reinforcing General Reasoning Without Verifiers
By
–
Reinforcing General Reasoning without Verifiers Zhou et al.: https://
arxiv.org/abs/2505.21493 #ArtificialIntelligence #DeepLearning #ReinforcementLearning