The Brief
New briefs daily around 7 AM Eastern
SemiAnalysis launches GDPx; the best frontier model scores just 12%
GDPval, still the general benchmark on most model cards, was designed before agents and no longer separates frontier models. GDPx uses specialist-built tasks, and its scores show safety classifiers pulling down Opus 5.5 and GPT-6 Astra.
- 01Read
SemiAnalysis: 62.2 GW of 2028 US datacenter builds face election risk
About three-quarters of 2027 capacity is already being built, so near-term compute supply holds. 2028 is exposed: Texas alone adds close to 30% of that year's US gigawatts and has frozen environmental permits since Sept. 21.
- 02Read
Brazilian court fines two law firms for hidden prompt injection
The court's own AI tool, Galileu, surfaced white-on-white text telling AI to respond superficially, and the firms were fined about $16,000. Hidden instructions in a document now carry a legal penalty, not only a technical risk, for whoever plants them.
- 03Read
Microsoft releases Mage-VL-Cowork, a 5B open computer-use agent model
The weights are Apache-2.0 and serve on Transformers, vLLM and SGLang, so a team can run screen-driving agents on its own hardware instead of paying per call to a hosted computer-use model. The model card excerpt carried no benchmark results.
- 04Read
Sakana AI will pretrain its own LLMs; Sawara Small due spring 2027
Sakana's earlier Namazu models were post-trained on other labs' open weights. Pretraining from scratch gives a Tokyo lab control of its data and training recipe, and Sawara is designed from the start for tool use and coding.
- 05Watch
Mintlify says AI agents now make up about 80% of its docs traffic
That share was about 10% at the start of the year. When most readers are agents, page size and accuracy count for more than visual design, and Coinbase reports about 100 million agent payments over x402, so agents are also buyers.
A dated brief from the vibe-coding frontier. Today’s Intel.