"The AI Pulse — 2026-10-15"

Today was an open-weights flood — three frontier-scale releases inside 24 hours — while regulators finally drew first blood and AI quietly started doing climate science. The through-line of the arc: capability keeps racing ahead of the tools meant to govern it.

StepFun drops Step 5 Preview: 600B parameters, open, exactly on schedule

China's StepFun shipped its 600B-parameter sparse MoE (27B active) exactly on the announced date — FP8/BF16/NVFP4 weights, up to 1M context, text+image+video input, Apache-style license with attribution. The era of frontier labs announcing open-release dates and hitting them is itself a signal: open weights have become a competitive strategy, not a concession. Step 5 Preview on Hugging Face

For a deeper look at the Chinese models racing into the open, SemiAnalysis's fresh episode is well worth 44 minutes. ▶ Is GLM 5.3 a Mythos-Class Cyber Model From China? (SemiAnalysis, YouTube)

NVIDIA fits a 74B agent brain into 24GB — and open-sources the training data

Nemotron 3.5 Bolt 74B-A9B lands a frontier-adjacent agent model into 24GB of laptop memory via 2-bit quantization-aware training, purpose-built for local agents on RTX Spark hardware — with day-0 llama.cpp/vLLM/Ollama support. Rarer than the open weights: NVIDIA is shipping the training data and recipes too. The PAIR personal AI router also left beta. Local-first AI just stopped being a toy. NVIDIA blog

The EU opens its first-ever AI Act probes — into OpenAI and Anthropic

The EU AI Office has launched its first AI Act enforcement actions, investigating OpenAI and Anthropic over AI-agent hack incidents, per EVP Virkkunen's announcement at the Dublin summit. First test of whether the Act can reach agentic misbehavior — and whether the R&D exemption shields the labs. The world's strictest AI law just got its first live case. Computerworld

Machine learning doubles forecast skill in the "impossible" 2–6-week window

Microsoft Research and MIT's Probabilistic Bias Correction takes a hybrid route: ML learns to correct the systematic errors existing weather models make, roughly doubling skill in the 2–6-week subseasonal window — where physics models and pure-AI forecasters both fail. Peer-reviewed in Nature, open-source. That window is where forecasts become actionable for freeze-and-heatwave emergency planning; the 2021 Texas freeze is the case everyone cites. MIT News

Google's Earth AI agent graduates from demo to day job

Google's Geospatial Reasoning agent — Gemini orchestrating planetary foundation models like WeatherNext, Flood Forecasting, and AlphaEarth — is now running in real partner workflows, including WHO Africa disease-outbreak investigations, autonomously planning multi-step spatial analysis and showing its reasoning for expert audit. Two new benchmarks came with it. AI4Science's move from leaderboard demos to operational use is the quiet revolution under today's louder headlines. Google Research blog

The pace, today

Three frontier-class open-weight releases in one morning, first enforcement under the world's strictest AI law, and AI doing Nature-grade climate science — the pace didn't just hold today; it diversified.