Daily AI CatchupAI Catchup — Saturday, July 4, 2026
Saturday, July 4, 2026Openai
Altman Proposes a US-Led AI Safety Forum Plus a 5% Government Stake in OpenAI
An IAEA-style forum to gate access to the most advanced models, floated alongside a government equity stake.
MetaMeta's In-Training 'Watermelon' Model Reportedly Matches GPT-5.5
Still training, using 10x the compute of Muse Spark — but already at benchmark parity with GPT-5.5.
Thinking-MachinesA Specialized Small Model Beats Frontier AI at 13.8x Lower Cost on Finance Tasks
A fine-tuned open model hit 84.7% accuracy where GPT, Claude, and Gemini averaged only ~50%.
Github-CopilotGitHub Copilot Adds Kimi K2.7 as Its First Open-Weight Model
Open-weight models are earning real estate inside mainstream developer tools, not just self-hosted setups.
Epoch-AiEpoch's New Benchmark Finds GPT-5.5 and Claude Can't Learn From Practice
Neither model showed meaningful improvement from repeated practice on the same task type.