Il y a 3 ans, les capacités de l'IA en mathématiques et en sciences se limitait à la multiplication à 3 chiffres pour GPT-3. Aujourd’hui, ces mêmes capacités sont évaluées à travers des questions de physique de la matière condensée. Quiconque pense que l’IA ralentit est
RESEARCH
-

Powerful AI: Definition, Timeline, and Public Skepticism
By
–
"What powerful AI (I dislike the term AGI) will look like, and when (or if) it will arrive, is a huge topic in itself. It’s one I’ve discussed publicly and could write a completely separate essay on (I probably will at some point). Obviously, many people are skeptical that
-

ByteDance Releases USO Open Style-Subject Optimized Model
By
–
ByteDance released USO, an open ‘style-subject optimized customization model’
— The Rundown AI (@TheRundownAI) 1 septembre 2025
It can preserve subjects and apply new artistic styles to create customized images
Available on Hugging Facepic.twitter.com/KNT359g98xByteDance released USO, an open ‘style-subject optimized customization model’ It can preserve subjects and apply new artistic styles to create customized images Available on Hugging Face
-
OpenAI Model Code Transparency and Verification Concerns
By
–
The lack of full visibility into the Python code the OpenAI family of models runs is baffling to me – I see this from Deep Research sometimes Being able to see the code is such an important tool for verifying the result, and I can't think of a cromulent rationale for hiding it
-
Flow State at Midnight: Working on Worthy Problems
By
–
hard to beat flow state at midnight while working on a worthy problem
-
Mistral AI Model Continues Development and Improvement
By
–
Mistral est évidemment une brique sur laquelle nous nous appuyons. Le modèle est encore perfectible, et nous y travaillons.
-
GPT-5 Pro needs transparency in code execution and analysis
By
–
That said, one thing OpenAI needs to do is let me see the analysis being done by GPT-5 Pro. I can't confirm its work without being able to see the python code it ran and what the results were. The fact that the error it spotted was real helps, but the thinking trace is obscured.
-

AI Autonomously Critiques Methods and Performs Advanced Statistical Analysis
By
–
The fact that my only prompt was "critique the methods of this paper, figure out better methods and apply them" and that it spontaneously checked the relationships in my tables, ran monte carlo simulations on my point estimates and did sensitivity analyses was kind of nuts,
-
Deep-Thinking AI Models Undersold: Market Potential Unrealized
By
–
GPT-5 Pro and Gemini 2.5 Pro Deep Think are both very impressive models for hard problems. I think they were both undersold during their respective launches, in part because I am not sure the labs themselves really understand the market for a slow, "deep-thinking" model, yet.
-

TiKMiX: Dynamic Data Influence Mixture for Language Model Pre-training
By
–
TiKMiX Take Data Influence into Dynamic Mixture for Language Model Pre-training
