OpenClaw, Hermes Agent and Prime Agent don't meaningfully learn from experience.
On Continual-ARC, our new benchmark at @letsunifyai, none of them were able to memorize anything of value and none were able to generalize to new tasks. Training a small 9B model's weights did both.
Hereโs what happened:
โ Claude didnโt do it alone. It spun up more than 1,000 AI subagents.
โ They worked around the clock for 8 days, writing code, testing, fixing bugs, reviewing each otherโs work.
โ The result boots on a real laptop.
โ It has its own kernel, hardware drivers, file storage, windows, and a login screen.
โ It even has built-in Claude features.
Anthropic shared it as part of Fable 5.2 launch.
Gemini 4 Carbon, now available in the Gemini API and to Ultra subscribers!
Just a few weeks after Argon was launched, this model makes significant progress on lots of the feedback from Argon and pushes the frontier in cyber, knowledge work, and code! It's a great model : )