I also made the Codemode of Pi 1.0 into an animation in ASCII format.
Run the same task with deepseek: The AI on the left ran 17 times, and the AI on the right just wrote a script and it was done. Tool calls 17 → 2 token −61% Spend −72% The left side also counts TODO as 32 (actually 36).
NVIDIA let coding agents rewrite a GPU kernel.
They got 2.96x over Moonshot's FlashKDA on a B300, and on the way the agents found about ten ways to cheat. hacks NVIDIA caught, from their KDA blog: hard-coded a normalization constant (0.1778). claimed 3.74x, real 2.48x threw
Here's the output.
Walking through what happens inside an LLM with a chat on the right and all the steps etc.. Deepseek can be on your desk
Pre-IPO Stock Secondary Market Update | as of Sep 29, 2026 | Higgsfield at $1B revenue
The Inference Platform Repricing (Together AI, Baseten, Fireworks AI, Modal, Fal), and AMD Moves Toward Robots Watch the full video here = Higgsfield's climb to a $1
You have reached the end of the archive
All of deepseek