Latest Eureka Report
Needs reviewAI Model Picks Up Coding Skill From One-Word Answers, Preprint Says; NVIDIA Pitches a Watchdog for Rogue Agents
The 60-second version
- Researchers from Lovart AI and four universities report that a 1.5-billion-parameter Qwen model gained 5.34 percentage points on the HumanEval+ coding test after training only on 5,664 one-word answers to unrelated prompts, in a preprint that is not yet peer-reviewed.
- NVIDIA launched its Open Agent Safety Platform on Sept. 28, pairing open-source OpenShell software with a Sentry watchdog on BlueField-4 chips, and says more than 100 organizations including Anthropic and Microsoft are working with it, while OpenAI is not named.
- Developer Mathias Strasser's free Jeff models return option probabilities in about 22 milliseconds on a high-end GPU by his own measurement, but early Hacker News testers said they fell well short of TypeSafe's proprietary Jev on real tasks.






