Robots can already perform 74% of US physical job tasks, covering 34% of working hours, but businesses have little economic reason to deploy them yet: they are cost-competitive with human labor for only 0.3% of work. The study measures exposure by asking whether current autonomous machines can perform tasks in less controlled settings, then compares that ability with human labor costs. Even if robot prices keep falling at past rates, the cost-competitive share would take 40 years to reach 10%.
Pre-training can make OLMo3-32B abruptly switch between solving a task and copying its tempting pattern, so researchers may need to evaluate intermediate checkpoints instead of trusting the final one. On an “answer+1” arithmetic test, its score fell from 81% at 2.17T tokens to 0% at 2.19T, then recovered to 81.7% at 2.21T. The paper calls this mode-hopping and proposes that competing circuits and changing data windows drive it.
LIFT gives language models a second route between generation steps, so they can carry forward internal information instead of passing everything through the decoded token. During training, each token is paired with a state from a pretrained model, and the new model learns to predict both the next token and next state in parallel. Across 135M to 1B parameters, the paper reports better language modeling, reasoning, and procedural-task results than standard Transformers under the same token budget.
Researchers turn 1,118 Lichess puzzles into a benchmark for testing whether better prompts actually improve frozen language models, without relying on static, easily contaminated test sets. Each answer gets cheap exact-match scoring, while engine checks can evaluate other moves; fresh puzzles and adjustable difficulty keep the test renewable. In a study of six optimization methods across eight models, the full run cost about $800, and the strongest evaluated model solved only about 55% of puzzles.
JusticeAxis gives legal-AI builders a benchmark for the harder task: applying law to established facts without defaulting to rigid statute matching or unsupported discretion. It uses 256 criminal cases from 18 countries, with audio, image, and text evidence plus three lawyer-written judgments per case: the recorded decision and one for each failure mode. The paper reports that open-weight models drift toward unsupported grounds, while frontier models lean toward the statutory default; its JusticeAgent plugin reportedly lifts a frozen open-weight model to commercial-level performance.
Today’s AI lacks the deliberative search that let AlphaGo find Move 37, a warning for anyone treating fluent model output as trustworthy reasoning. AlphaGo paired a policy network that suggested plausible moves with search machinery that built thousands of possible futures and tested their consequences. The essay argues that future systems need this kind of genuine reasoning to produce trustworthy, novel insights in science and medicine.
Praxa gives AI agent builders a way to distinguish a proposed action from authorized execution, its verified external effect, and promotion into serving, making each transition testable instead of implicit. Its harness uses deterministic admission, brokered execution, external read-back, and reconciliation; a repository audit passed 1,027/1,027 unit tests and 89/89 Workerd tests. A 12-task Terminal-Bench pilot found no reliability advantage, so the paper supports the architecture, not production gains or security claims.