AI News/Wire summary
FrontierML tests whether AI agents can productively spend money on model training
Key points
- Benchmark covers 21 long-horizon machine learning tasks across six categories, including LLM training and computer use.
- Each agent gets a 12-hour CPU sandbox and a wallet of $10, $50, or $200 for compute and data.
- Opus 5.5 gained 45.4 points at $10 and 58.8 at $200, while a human researcher gained 56 and 66.1.
- Best open-weight model, Kimi K3.1 Pro, reached 25.5 at $10 and 29.6 at $200, according to the authors.
Why it matters for marketers: Agent gains rise with budget, so teams comparing AI tools should look at results across spend levels, not a single headline score.
Summary by the Z Media news desk, from the outlets' reporting. Read the full stories at the source.
Read the original
Import AI: Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development
Related on Z Media
- AlphaGenome 2 predicts how DNA variants act in specific human cell typesvia Google DeepMind, 6 days ago
- Blackwell traces found in Chinese AI models as Congress eyes tighter chip rulesOur brief, via TechCrunch, 4 days ago
- Anthropic says selective notes beat full logs for Claude agent memoryvia The Decoder, yesterday
- Apple launches study testing how people rate AI health answersvia The Verge, yesterday
- GPT-6 Astra helps produce proof toward a lattice Standard Model problemvia OpenAI, 2 days ago
- Gates Ventures and West Health offer up to $15 million for AI Alzheimer's researchvia The Information, 15 hr ago