Morning Edition · Sunday, June 28, 2026
MirrorCode Benchmark Finds Frontier Models Can Rebuild Real Software From Scratch, Up to a Point
Epoch AI and METR ask models to rebuild working programs without any source code, and the best result is a 56 percent solve rate on tasks estimated to take human engineers weeks.
MirrorCode, a long-horizon coding benchmark from Epoch AI and METR, gives a model only a compiled binary it can run, natural-language documentation, and example input-output pairs, then asks it to rebuild the program without ever seeing the…
Continue the AI Intelligence Brief
Track frontier labs, chips, export controls, model releases, regulation, and AI infrastructure.
- 5 AI intelligence signals a day
- Frontier labs, compute, and chips
- Model releases and AI infrastructure
- Source-grounded analysis with confidence labels
The Global Intelligence Brief stays free.
Part of a tracked trend
Oversight and Evaluation Lag Accelerating AI Capabilities
Over the next 3-6 months, evidence mounts that governance, evaluation, and agent-safety methods are failing to keep pace with capability growth, driving investment in interpretability, agent-manipulation benchmarks, and institutional-reform proposals.
More from this edition
- AI's First Revenue Census Puts Generative-AI Sales at $110 Billion, Above Infrastructure Depreciation
- Washington Orders Anthropic to Cut Foreign Access to Two Top Models, Citing a Narrow Jailbreak
- DeepSeek Open-Sources DSpark, Claiming Up to 4x Throughput on V4 via Speculative Decoding
- EU Picks an Italian Startup to Build a 400-Billion-Parameter Open Model in 24 Languages
- Open-Vocabulary Segmentation Reaches Production as Apps Build on Meta's SAM 3
- Anthropic Asks Washington for Power to Block Dangerous Models, Plus $350 Million for Labor Fallout
- Alibaba Sues the Pentagon to Get Off Its Chinese-Military-Company List
- Anthropic Puts Claude Inside Slack as a Taggable Teammate With Shared Memory
- A Venture Investor Accuses AI Leaders of Using Doom Warnings as a Fundraising Tool
- Brain-to-Image Decoding Advances, and DeepMind's Hassabis Frames It as AI for Science
- Meta's Muse Spark Shows the Efficiency Turn in Frontier Models, and a Break From Open Weights