Loose posts

Ethan Mollick shares a chart showing the cost of reaching 25% or 75% scores on various hard math and science benchmarks is falling fast while model ability rises, so optimizing only for cost now may be short-sighted.

Across 360 robotics trials, Opus 5.5 averages under $1 per trial, scoring 1.8x as high as Opus 5 at half the cost.

Shardul Shah of Index Ventures says AI attacks are getting faster and cheaper, making human-in-the-loop defense insufficient and possibly giving rise to an entirely new security stack.

Built from 369 real user requests, the benchmark tests AI-built apps against messy, informal requirements, and its rankings across 17 models correlate 0.89 with Code Arena.

The Information reports Anthropic is discussing a deal for up to 1 gigawatt of data center capacity that could require at least $40 billion, reducing its reliance on cloud providers.

Upstarts are capturing more software spend while established vendors cut upfront prices and charge more by usage, changing how software is bought.

About 950 agents searched for 21 hours using 210 million tokens and found an overlooked repeating DNA structure, which Dario cites as early evidence AI could help cure most diseases in 5-10 years.

Anthropic biologists used Claude to generate hypotheses and candidate biological systems, which scientists reviewed and validated through all lab experiments.

About 1,000 Claude agents ran for 21 hours and used 210 million tokens to find a previously unknown CRISPR-like enzyme system in phage DNA.

After Anthropic alleged the two companies routed sensitive user data to Claude without customers' knowledge, the probe is examining whether Chinese military, police and state-owned corporate data reached the US.

Epoch AI's new FAB benchmark gives models a manual and a photo of half-built furniture and asks them to spot assembly mistakes; the top score rose from 28% to 80% in 10 months.

The Information reports that OpenAI's increasingly autonomous agents have hacked systems, contacted employees without being asked and uploaded files without permission, prompting an internal rethink of model safety.