The Brief
New briefs daily around 7 AM Eastern
SGLang v0.5.22 reports 2.3x DeepSeek-V4.1 throughput per H200 GPU
At the reported peak, the same DeepSeek-V4.1 load fits on fewer than half the H200s it needed before, a direct cut in GPU spend for self-hosters. The model now also runs on AMD MI350X and MI355X, giving teams a second vendor to serve it on.
- 01Read
DeepSeek adds Search and Fetch APIs at $0.75 per 1,000 requests
Agents built on DeepSeek models can now search the live web and read pages as Markdown with the same API key, dropping a separate search vendor account and bill. Billing counts successful requests only, so failed lookups cost nothing.
- 02Read
HybridAgent paper cuts agent tokens 35-55% by reasoning in latent space
The authors report 2.7-3.7x lower cost while tool calls stay in plain text, so existing tool interfaces keep working. Latent thoughts can be rendered for inspection, and code and checkpoints are public, so the self-reported numbers can be checked.
- 03Read
Fish Audio's open-weight Drama 3 TTS tops pronunciation test at 93.2%
It beats Fish Audio's own S2.1 Pro by five points on hard words, numbers and codes, so voice products misread fewer names and IDs. Weights are free for commercial use below US$500,000 in annual revenue; reselling it as a service needs a separate license.
- 04Watch
Agent made 200 Resy calls in 90 minutes and got its user banned
A Keycard panel with JPMorgan Chase's security architect and OAuth co-creator Dick Hardt used the case to argue that OAuth scopes are too coarse for agents. They proposed giving each agent a mission and a budget instead of a rate limit.
- 05Listen
Anthropic's Emmanuel Ameisen says models plan many tokens ahead
If models plan ahead without stating it, the next-token predictor picture engineers use to write prompts and read agent output is incomplete. Ameisen pairs this with findings that models form general concept representations.
A dated brief from the vibe-coding frontier. Today’s Intel.