Latest Eureka Report
Needs reviewOpenAI and Anthropic Say AI Models Refuted the Exponential Time Hypothesis, Plus Chat Models That Miss Faked Seconds in Video and a Local File Agent
The 60-second version
- OpenAI and Anthropic each say an internal AI model found an algorithm that refutes the Exponential Time Hypothesis, though Anthropic says its version beats brute force only above 10 to the power 10,000 variables and no person is known to have checked its Lean proof.
- In a preprint that is not yet peer-reviewed, researchers report that the audio-visual chat models Qwen2.5-Omni and Qwen3-Omni scored only 51 to 52 AUROC, about a coin flip, at spotting a few edited seconds in long videos, while probes on the same models' internal features reached 90.8 for audio and 80.8 for visual.
- Hob, an MIT-licensed desktop app from Frigade, runs a file-handling AI agent on local models from Gemma 4 E4B Compact up to gpt-oss 120B on machines with at least 8 GB of RAM, and uses code guards to catch small models that claim work they never did.
