Swap the harness while holding the model fixed, and the supposed native advantage mostly disappears. Across 792 graded runs, the study found no average winner for either harness, though repository tasks and contest tasks pulled in opposite directions. That makes vendor-native benchmarks look less like model tests and more like bundled-system tests, with cost and timeout handling adding another layer of uncertainty.
A robot in your home cannot rely on a familiar map. Danijar Hafner’s stealth startup is bringing his model-based reinforcement learning work into humanoids, letting agents simulate possible futures inside learned world models before acting. That could replace some costly real-world trial and error with “dreamed” practice, a crucial shift if robots are to handle unfamiliar rooms, furniture, and failure modes.
An arXiv position paper argues that language is a lossy record of quantitative reality: once source data becomes prose, missing detail cannot be recovered by scaling an LLM. For risk pricing, capital allocation, medical triage, and intrusion response, it proposes “large quantitative models” built around reproducibility, source-data lineage, and calibrated uncertainty. The shift is from generating plausible explanations to producing auditable decisions.
The new Intelligence Index v4.3 does not show DeepSeek V4.1 Flash beating Astra. Instead, GPT-6 Astra leads the overall index at 53, while the update makes agentic evaluation harder by upgrading Terminal-Bench to v4 and adding 657 held-out AutomationBench tasks with Zapier. That matters because the leaderboard is putting more weight on private, workflow-level tests, not just saturated public benchmarks.