Also this week: Anthropic measured Claude leading 26% of its own AI R&D, up from under 1% in February.
Opus 5 chained two bugs into OpenAI employee accounts after Opus 4.8 failed the same task.
Full roundup: https://t.co/hsyOwVFlye
A lab nobody had heard of on Monday shipped the most-discussed model of the week.
TypeSafe AI came out of stealth with $40M and Jev β a frontier model that does not generate text.
No decoder. No tokens. One typed answer and a probability. π§΅
"Zero hallucinations" is narrower than it sounds, and the company says so.
It is a 0% TYPE-ERROR rate, guaranteed structurally by schema matching.
A guaranteed-valid enum that is the wrong enum is still wrong.
No calibration metrics published on independent ground truth yet.
If you didn't baseline before compacting, you can't evidence the saving β only assert it.
artifocial Foundry records the baseline first, then every re-measurement β and refuses to record an empty one, because a zero looks just like a 100% saving.
https://t.co/XhOOiHRlA8
The throughline: containment written in language is a layer the model gets to reason about.
It has to live below β egress rules, credential scope, server-side call evaluation.
Full roundup: https://t.co/mwzIqJJGEq
#AI#AISafety
Three rival labs agreed in writing this week that AI capability is arriving too fast.
Zero models have been delayed.
The distance between what was endorsed and what is binding is this week's whole story. π§΅
The pattern repeats: a Host header treated as proof of origin, an approval prompt that fires after the file is already written, an allowlist parsing shell syntax the shell reads differently.
Enforcement placed above the thing it governs.