AI
4 items · 4 articles in this topic
A thin day for AI releases. The items are an open implementation, an inference benchmark, an agent-tooling project and a governance interview.
- 1
NVIDIA published a blog post on serving GLM-5.3 at about 608 tokens per second across two DGX stations using vLLM. The headline figure is vendor-reported.
Hacker News (front page) ↗15 min read
- 2
A Show HN project implements the DeepSeek V3 architecture in roughly 200 lines of pure JAX, which makes the design easier to study.
Hacker News (front page) ↗5 min read
- 3
Turbo-Browser-MCP offers batch form filling, DOM deltas and submit guards for browser automation over MCP. The submit guards address the risk of agents submitting forms unintentionally.
Hacker News (front page) ↗12 min read
- 4
Helen Toner, the former OpenAI board member who voted to fire Sam Altman in 2023, tells The Verge that AI CEOs still cannot be trusted. The interview covers the firing and AI safety policy.
The Verge ↗57 min read