@aymericroucher - AI Dynamics

Top AI models stay close in a fast-moving packed peloton

By

–

01 March 2026 13h53

Funny how the top models from different labs stay within a tiny distance from each other. The peloton is moving extremely fast, but stays quite packed, no breakaway so far. And that is despite large differences, e.g. in available compute. What's the forcing function that makes

→ View original post on X — @aymericroucher

1 March 2026

Encouragement to build with smolagent’s CodeAgent and propose improvements

By

@aymericroucher

–

20 February 2026 16h03

Nice work!
Don't hesitate to build a bit with smolagent's CodeAgent if you're interested in code-actions agents, I'm less involved now but would still love to look at what improvements you could propose!

→ View original post on X — @aymericroucher

20 February 2026

Custom Taalas hardware runs Llama-3.1-8B at 17k tokens per second

By

@aymericroucher

–

20 February 2026 15h53

Custom hardware from Taalas runs Llama-3.1-8B, at 17k tokens per second
17k Absolutely insane (For the record Cerebras is crazy good and they're at 2k on the same model)
And latency is very low too! Their chatbot is here: https://
chatjimmy.ai It's genuinely a eerie

→ View original post on X — @aymericroucher

20 February 2026

Hugging Face Icon Now Available in Font Awesome

By

@aymericroucher

–

11 February 2026 12h00

Tell your vibe-coding tool: you can now use the @huggingface icon in Font Awesome ! I posted the original request for addition in early 2025: thanks to the Font Awesome team for following up!

→ View original post on X — @aymericroucher

11 February 2026

From March 2025 to Feb 2026: AI writes all code

By

@aymericroucher

–

04 February 2026 19h39

Devs, march 2025: Amodei is insane, no way AI writes all our code next year
Devs, feb 2026:

→ View original post on X — @aymericroucher

4 February 2026

Kill conditions for agent orchestrators to stop error loops

By

@aymericroucher

–

27 January 2026 18h34

Interesting!
One thing that I have to always do: whenever a Claude Code starts to go wrong, I kill it, because when they fill their context with error logs and wrong ideas, they tend to keep digging their hole.
A good agent orchestrator would have "kill conditions" to just

→ View original post on X — @aymericroucher

27 January 2026

LeCun: child perceives more visual data than LLM

By

@aymericroucher

–

21 January 2026 2h24

This is false: LeCun does the calculation in a video, I need to find it: a child in their first 4 years of life perceives more visual data than the largest LLM trainings use.

→ View original post on X — @aymericroucher

21 January 2026