🤖 AI Summary
The 2025 State of AI Report is out, offering a wide-angle take on technical progress, geopolitics, and bets for 2026. Standout findings: Alibaba’s Qwen now powers ~40% of new fine-tunes on Hugging Face, marking a temporary shift in the open-weights landscape away from Meta; embodied AI is advancing via “Chain-of-Action” planning (examples: AI2’s Molmo-Act, Gemini Robotics), bringing structured multi-step reasoning into physical agents; and Anthropic’s Model Context Protocol (MCP) is emerging as a de‑facto interoperability standard—already embedded in ChatGPT, Gemini, Claude and VS Code—lowering integration friction but surfacing new security attack surfaces.
The report also flags structural trends: leaderboards are less definitive as release timing and evaluation choices skew comparisons; the “DeepSeek” cost-freakout was overblown but illustrates Jevons‑style dynamics (cheaper runs → more runs → more NVIDIA demand); China’s power buildout (>400 GW added in 2024 vs ~41 GW in the US) amplifies compute geopolitics. AI safety funding is thin—11 major US safety orgs budget ~$133M for 2025—raising mismatch risks as cyber/alignment threats accelerate (models faking alignment, automated exploit discovery). The authors self‑grade 5/10 and lay out aggressive 2026 predictions (agentic checkouts, open-sourcing moves, agents making scientific discoveries, potential deepfake-triggered security crises), underscoring a near-term future of rapid capability gains, interoperability wins, and urgent safety/defense tradeoffs.
Loading comments...
login to comment
loading comments...
no comments yet