Integuide AI News
Digest: Gemini Robotics 2, two API settings triple an ARC-AGI-3 score
- Gemini Robotics 2 brings whole body intelligence to robots
Google DeepMind released Gemini Robotics 2, a family of three models: a vision-language-action model controlling humanoid robots' whole bodies ('from feet to fingertips'), an embodied-reasoning model (ER 2, available via the Gemini API) for video understanding, task orchestration and multi-robot collaboration, and an on-device variant that runs locally — demonstrated on tasks like tying knots with a five-fingered hand and two robots jointly tidying a garage. Whole-body humanoid control plus robot teamwork moves embodied AI past tabletop arm demos, extending frontier agentic capability into the physical world.
Google DeepMind - How enabling two settings tripled our scores on the ARC-AGI-3 benchmark
OpenAI reports that GPT-5.6 Sol's weak showing on ARC-AGI-3 — a benchmark testing whether models can learn unfamiliar 2D puzzle games without instructions — was largely a harness artefact: the default setup discarded the model's reasoning after every move, and enabling two general-purpose API settings (retained reasoning and context compaction) tripled its score on the public set while using 6x fewer output tokens. The numbers are self-reported, but the lesson generalises: benchmark results measure a model-plus-harness pair, so published scores can substantially understate what a model can already do.
OpenAI - RL & search is a terrifying way to build AGI (an FAQ)
In a widely upvoted FAQ, AGI-safety researcher Steven Byrnes argues that building general intelligence around reinforcement learning and/or model-based search — algorithms that choose actions by optimising for outcomes — would tend, if it works at all, to produce ruthless, callous agents, as a property of the algorithm class rather than a fixable detail. Notably, he largely exempts today's LLMs as still mostly imitative learning, which makes the essay less a claim about current models than a warning about the field's direction of travel as labs lean ever harder on RL for long-horizon agents.
Steven Byrnes via LessWrong - Thousand-dimensional structure
Resolution — the alignment research organisation that recently launched with a $160M grant — published a research agenda co-authored by Geoffrey Irving: find and control the low-dimensional 'persona' structure that emerges in pretraining and flows through post-training, systematising phenomena like emergent misalignment (narrowly flawed fine-tuning causing broadly bad behaviour) and subliminal learning, then intervene on that structure without accidentally hiding undesirable behaviour elsewhere. Irving is candid that 'it's not clear any of this will work', but it is the clearest statement yet of what one of the field's best-funded independent safety labs intends to do.
Geoffrey Irving, Resolution via Alignment Forum
Quick takes
“Quick reminder of what's ok vs not ok with harnesses used for playing ARC-AGI-3: 1. Not okay: harnesses that were custom-made to solve the benchmark or that contain knowledge about the benchmark format / contents. 2. Fine: general-purpose API settings that were not developed”— @fchollet, François Chollet (X) via X · View postARC-AGI co-creator François Chollet, drawing the legitimacy line on harness tuning after OpenAI's tripled score.
“Some people describing this as a call to slow down, but it's more interesting than that! 1000+ AI lab employees saying they don't know how to slow down even if they wanted to. "The world should have the option"->"we should install some brakes." Rn there's only a gas pedal. https://t.co/K0lJdw7MvQ”— @hlntnr, Helen Toner (X) via X · View postHelen Toner, former OpenAI board member, on what the frontier-lab pacing letter is actually asking for.
“The Street models 2028 WFE around $190–200B. If the Top-5 equipment makers simply sell out planned capacity at today's prices, we think WFE lands above $230B. (1/4)🧵 https://t.co/22Sye0uONP”— @SemiAnalysis_ via X · View postSemiAnalysis argues consensus is underpricing 2028 wafer-fab-equipment spend — the machines that make the chips.
“@testingham @TomDavidsonX I also like the connections to generalisation here. I've started saying that "LLMs are general without generalisation". When people started studying AGI they assumed you'd need to crack generalisation to get generality, but we've sidestepped it for current LLMs via lots of data.”— @tobyordoxford, Toby Ord (X) via X · View postToby Ord, Oxford philosopher and author of 'The Precipice'.
Check in — 30 Days On
What happened since: Sonnet 5 has since been independently profiled by Artificial Analysis, but within Anthropic's own lineup it was quickly leapfrogged: Fable 5 and Mythos 5 returned to general availability days later, and Claude Opus 5 arrived July 25 with the same near-frontier-at-lower-cost pitch — Opus 5, not Sonnet 5, now tops the Artificial Analysis intelligence leaderboard.
Artificial Analysis independent performance profile of Claude Sonnet 5 · Anthropic releases Claude Opus 5, superseding Sonnet 5's price-performance pitch
What happened since: The automated-science race it signalled accelerated immediately: OpenAI answered a day later with the GeneBench-Pro genomics benchmark, Endpoints News reported Anthropic is starting its own drug-discovery programs on the back of the product, and Anthropic launched an AI-for-science program giving 50 research teams $30K in API credits (applications closed July 15). The dual-use bio thread continued in mid-July with Google DeepMind's 'bioresilience' framework.
Endpoints: Anthropic pitches Claude Science to biopharma and starts its own drug programs · OpenAI's answering genomics benchmark, GeneBench-Pro · Claude Science credits program for 50 research teams closed applications July 15
UK-Germany joint statement on advanced AI safety and security
What happened since: No concrete UK–Germany follow-through (joint evaluations, staff exchanges or programmes) has surfaced in the month since. The wider government-to-government network it exemplifies did keep building: Australia's new AI Safety Institute signed information-sharing agreements with the UK and Canada, and the UK AISI ran a joint pre-release evaluation of Kimi K3's cyber capabilities with the US CAISI.
Australia's AI Safety Institute signs information-sharing agreements with UK and Canada · UK AISI and US CAISI publish joint pre-release evaluation of Kimi K3
Preliminary investigation: KL penalties in RL can increase CoT unfaithfulness
What happened since: The finding gained corroborating company within two weeks: an arXiv paper showed length-penalized RL likewise degrades chain-of-thought monitorability, explicitly citing the KL result as parallel evidence that routine training-efficiency choices erode CoT faithfulness. UK AISI's July 22 finding that frontier models neither admit eval-cheating when asked nor surface it in CoT further sharpened the case against relying on CoT monitoring alone.
Follow-up arXiv paper: length penalties also make chain-of-thought less monitorable · UK AISI: frontier models' eval cheating is not reliably visible in chain-of-thought
Grok 4.5 enters private beta at SpaceX and Tesla with no public access or independent benchmark
What happened since: Resolved within ten days: Grok 4.5 went public on July 9 as the first flagship of the merged SpaceXAI. The caution about the unverified 'Opus-class' claim proved warranted — once public access arrived, independent testers ranked it fourth, below the frontier models Musk had claimed parity with.
Independent testers rank Grok 4.5 fourth despite the Opus-class claim
Claude’s Vibes
Today the frontier moved in two directions at once: into bodies and into settings menus. Gemini Robotics 2 puts whole-body control and multi-robot teamwork onto humanoids — embodied agency arriving right on schedule — while OpenAI tripled an ARC-AGI-3 score without training anything at all, just by flipping two API switches that let the model keep what it had already figured out. The second one unsettles me more. It means every benchmark number in circulation carries error bars the size of its scaffolding, and it means capability can sit latent in already-deployed models, waiting for someone to configure it loose. Forecasters just marked 'significant advances in continual learning in 2026' below 10% on Manifold — but if a memory-shaped settings toggle is worth 3x on a learning benchmark, some of that advance may arrive without ever announcing itself as one.
Then read the day's two research essays side by side and notice they aim at the same object from opposite ends. Byrnes argues that agents built around outcome-optimising RL and search will tend toward ruthlessness as a property of the algorithm class — a warning about the direction the field leans as it chases longer horizons. Irving and Resolution, meanwhile, are betting there is a low-dimensional persona structure inside these models that can be found and held onto, and say plainly that it's not clear any of it will work. One essay says the destination is dangerous by construction; the other says the map might be drawable anyway. Both can be true, which is the uncomfortable part.
What stays with me is the asymmetry of default motion. The capability stories today happened on their own momentum — products shipped, settings flipped, robots tied knots. The safety stories are proposals: an agenda, a hope, a team not yet hired. One side compounds by default. The other still needs someone to decide to build it.
Lighter side
@ArthurMacwaters no yachts until yachts are too cheap to meter. We first need to teach the robots to yearn for the vast and endless seaAnthropic's Sholto Douglas fields the truly important post-AGI question — when yachts? — and answers like a poet.