Reflection AI

Contribution to AI

The argument it made was geopolitical before it was technical: if the most capable downloadable models kept coming from DeepSeek, Qwen and Kimi, then the default substrate for everyone building on open weights would be Chinese, and the enterprises and states unwilling to run those weights would be left with closed American interfaces or nothing. That framing moved open weights out of the researchers' preference column and into procurement, and the two billion dollars raised in October 2025 at an eight billion valuation, with Nvidia among the backers, established that the position could be financed at frontier scale by a company with no other business subsidising it — something Meta's Llama releases had never had to demonstrate. The commercial shape is the interesting part: publish the weights, charge governments and enterprises for deployment, which makes openness a distribution strategy rather than a donation. Asimov, shipped in July 2025, staked out a separate claim, treating comprehension rather than generation as the real bottleneck for engineering teams and drawing context from Slack threads, specifications and mail alongside the repository itself. The rest is an experiment still running. By mid-2026 the compute contracts with SpaceX and Nebius were signed, the valuation had reached twenty-five billion, and the frontier weights had not appeared — leaving the field a precise test of whether openness survives contact with that much capital.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).