September 27, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Safety A report said underground sellers were offering stolen AI-model access for as little as three percent of legitimate subscription prices. IT之家 ↗
- AIGC Engram is a sampler that converts AI hallucinations and broken outputs into experimental musical material. The Verge ↗
- Policy Anthropic CEO Dario Amodei was scheduled to have dinner with President Donald Trump amid growing AI-policy tensions. TechCrunch ↗
- Agents Databricks said 91% of enterprises already use at least two AI coding tools, making governance more difficult than adoption. X / Databricks ↗
- Agents GitHub said its Copilot app can run agents in parallel, with each session receiving an isolated Git worktree and context. X / GitHub ↗
- AIGC NVIDIA said it collaborated with Meta to improve the efficiency of Meta’s Muse Realtime Avatar model. X / NVIDIA AI ↗
- Safety OpenAI said it fixed a bug that degraded image understanding in GPT-6 Sol and GPT-6 Luna across its API and Codex. X / OpenAI Developers ↗
- Industry Anthropic’s CEO received a humorous Saturday Night Live treatment as the company’s public profile expanded beyond technology circles. TechCrunch ↗
- Models Cohere promoted Parse 5 as an enterprise document parser that extracts tables, images, bounding boxes, and flowcharts. X / Cohere ↗
- Agents Researchers reported that AI agents now perform more model-development work, while humans still make the major research decisions. The Decoder ↗
- Funding A Tsinghua-linked quantum-AI startup reportedly reached a valuation of about 1 billion yuan while targeting large-model infrastructure. 量子位 ↗
- Research A report described a compact quantum-computing box completing an end-to-end workflow while keeping all data on-device. 量子位 ↗
- Research An anonymous Yutu model reportedly ranked first on two benchmarks after testing showed strong coding performance. 量子位 ↗
- AIGC Recraft demonstrated its V4.1 Flash model generating a detailed neo-retro Japanese poster series from a single prompt. X / Recraft ↗
- Open Source NVIDIA released an open 100-million-parameter model that identifies up to eight simultaneous speakers in real time. The Decoder ↗