The Brief
New briefs daily around 7 AM Eastern
Google launches Gemini Decisions API at $0.10 per million input tokens
Unlike Jev, the decision model it is compared against, gemini-decision-1 accepts images, audio, video and PDFs and returns graded probabilities. Routing and policy checks on media become one cheap scoring call instead of a general model writing text per question.
- 01Read
Runway open-sources Praxis-1, a 27B robot world model, under Apache 2.0
Robotics teams get weights, inference code and LoRA fine-tuning code under a commercial license, and Runway reports 46.2% on RoboLab-120, ahead of open VLA baselines such as pi0.5. Video pretraining becomes a usable data source for robot policies.
- 02Try
Official MCP filesystem, git and fetch servers patch six vulnerabilities
The fixes are in the official reference implementations. Symlink races in the filesystem server could expose files outside the directories an agent was allowed to touch, so older versions leave that sandbox boundary open.
- 03Read
United Airlines blocks unauthorized AI agents from booking and check-in
The block covers travel agencies and aggregators as well as United's own platforms, so consumer travel agents that act through shared MileagePlus credentials lose access to a major US carrier and need first-party or partner channels instead.
- 04Watch
Duolingo agents merged 700+ PRs and 225,000 test lines since March
Test coverage on Duolingo's iOS app rose from 11% to 24% with engineers acting only as final reviewers. The loop of picking untested code, fixing CI and merging under review carries over to other large, repetitive migrations.
- 05Read
Anthropic says Claude ported nearly all AMD GPU kernels in three days
Kernel porting is the low-level work that gates serving a model on new hardware. Anthropic now serves Claude across Trainium, TPUs, NVIDIA and AMD, and says a model did that work for AMD with little human help.
A dated brief from the vibe-coding frontier. Today’s Intel.