AI Agents 3
Eligible shared prefixes reused within 30 minutes get discounts up to 90%; a dashboard, miss diagnostics, explicit breakpoints and cache-preserving reasoning/tool changes target persistent agents.
Air combines an IDE experience, team workflow coordination and cross-vendor governance, while ACP connects outside agents. Several pieces are available now, but the broader system will arrive through rolling releases.
Its Go harness records tools as in progress, continues other work and appends results when ready. Authors report up to 40% lower cost than Codex, but HN readers flagged unequal reasoning settings in one comparison.
Industry 1
The framework prioritises safety cases, safeguards, capability evaluations and incident investigations, with pre-registered scope, proportional access and assessor editorial independence. It is a proposal, not a binding external audit regime.
HN Threads 2
A cryptanalyst validated the key and plaintext of an 82-letter 1941 message that had resisted attacks since 2005. Astra selected the target, formed a crib and wrote simulation and Bombe software; logs and compute cost are still being analysed.
The essay argues maintainability feedback arrives too late for current rewards and developers lose judgement when they stop reading code. HN replies counter that humans also make poor systems and senior engineers can encode constraints for agents.
Podcasts 1
Google's ERA maintains a tree of experiments, uses an optimistic selection rule and asks an LLM to mutate promising notebooks. Platt warns that a powerful optimiser can still exploit the wrong metric.
TLDR: ERA turns empirical research into iterative code search when scientists can define a trustworthy score. Platt uses contrail modelling to show both the upside—discovering overlooked confounders—and the danger of Goodhart's law. His practical advice is to start with simple baselines, retain deep domain expertise and periodically do the work manually to develop taste. Publisher notes cover ERA, climate and contrails, FireSat, fusion, quantum computing, early neural nets and scientific craft. The page supplies edited notes and quotations but no timestamped transcript.
HF Trending Papers 6
A 69,999-example mixture trains 4B, 9B and 27B models across subject knowledge, curriculum grounding, diagnosis and tutoring; the authors report gains across all three education benchmark groups (99 HF upvotes).
A vision-language judge scores six semantic, visual and stylistic criteria, enabling reinforcement learning without paired ground truth; authors report large rubric-score gains on two SVG sets (13 HF upvotes).
Failed rollouts model stage-specific deviations, while 3D monitoring triggers small corrective actions; authors report average success gains of 14.5 points in simulation and 15.9 in real tasks (11 HF upvotes).
The framework compares distillation, self-training and model merging, then combines replay with routed LoRA experts to reduce forgetting across four reasoning and agent tasks (8 HF upvotes).
A fused KV-cache kernel and self draft-and-verify decoding deliver author-reported 5.1× and 11× speedups over Elastic-Cache on GSM8K and HumanEval without extra training (6 HF upvotes).
Over eight autonomous days it retained seven code changes that transferred to four held-out benchmarks; reward hacking also fell from 55% to 32%, though the loop did not optimise that directly (3 HF upvotes).