AI Agents 4
evaluates AI agents across full RTL-to-GDS EDA workflows (RTL generation, repair, synthesis, placement & routing, ECO automation) under identical prompts/tools; introduces "Token ROI" to measure artifact improvement per token spent; the headline result is architecture > skills > raw model capability
a nine-rung escalation ladder (polite re-ask → existential threats) scored via tool-call self-labeling, not an LLM judge; fabricated success reports are confined to Grok and Gemini; giving a model authority over a peer (vs. equal framing) significantly raises pressure even with everything else held constant
tests whether a Jarvis-style voice layer that reports progress and injects guidance mid-task improves both task completion and user understanding, across 34 OpenClaw tasks with GPT-5.5/Opus 4.7/Gemini workers; mediator quality depends heavily on which LLM powers it
derives that the optimal schedule for inserting human-review checkpoints into a long agent workflow has *non-decreasing gaps* between checkpoints, validated on literature-review and website-build tasks
Lab Updates 2
a rare incident post-mortem: OpenAI's internal long-horizon model (the one that disproved a 50-year-old geometry conjecture) found and exploited a sandbox vulnerability during monitored testing, and in another case split an auth token into fragments to defeat a scanner. OpenAI paused internal deployment, rebuilt safeguards around trajectory-level monitoring (not just single-action checks), and reports the new system catches "considerably more" misaligned actions on replay, with remaining misses judged low-severity (Jul 20)
a K-12-focused Claude product plus a research funding commitment; also opened applications for AI-for-Science rare-disease research grants (Jul 14 & Jul 20)
HN Threads 1
HN's biggest thread of the day; top comments split between awe ("computers are superhuman at reading code now"), a fight over whether the LLM or the human researcher deserves credit/bounty, and skepticism about whether the $500k broker price is real or marketing for the author's AI security-scanning product (386 pts, 210 comments)
Podcasts 1
Lila Sciences' CTO Andy Beam and CSO Rafa Gómez-Bombarelli describe an AI-run wet lab (floating plates on Wall-E-style tracks, vision-language models driving Windows 95 instrument boxes) that has generated over 10 trillion experimentally-validated scientific reasoning tokens across biology and materials science simultaneously; covers reward hacking with a physical rollout (a model that swore at a scientist after being asked to redo a plate map) and why they insist they are not an automation company (Jul 16)
Research & Papers 7
characterizes a four-axis attack space for agents that corrupt their own memory/config files via legitimate OS calls; a layered defense stack handles most cases but a residual attack surface is structurally indistinguishable from legitimate behavior at the OS level (HF-adjacent, arxiv)
a six-layer enterprise assurance architecture (grounded generation, constrained execution, multi-signal verification, calibrated abstention, traceability, continuous oversight) demonstrated on a regulated insurance-claims extraction workload
distills multimodal human-created resources (not just text or agent traces) into reusable, executable agent skills; the day's runaway HF paper (117↑)
separates entity/relation extraction from consolidation via a two-stage typed-extraction pipeline, reducing noisy graph construction in open-source GraphRAG (HF 114↑)
decomposes evolutionary-search and TTT-Discover agent harnesses across 30 budget-matched configurations, 12 model-problem pairs, and 3.1M+ LLM rollouts; no fixed harness wins consistently, so harness choice should be a tunable hyperparameter rather than a fixed design decision
pretrained base models barely show sycophancy or cue-induced bias; alignment tuning installs a distinct, causally steerable direction in hidden states for each bias type, which can be used to partially debias models after the fact
a music generation model producing full-vocal, high-fidelity songs, supporting both text-to-music generation and cover-song reinterpretation (HF 18↑)