Integuide AI News
Digest: METR charts discovery acceleration, Australia AISI multi-agent report
- Have We Seen an Acceleration in Discoveries? Recommended
METR's Tom Cunningham and Nate Rush assembled the public time series of 'discoveries' they could find — vulnerability disclosures, open-problem lists in mathematics, and algorithmic-optimization records — to ask whether LLMs have visibly accelerated the aggregate rate of discovery. The answer is strikingly uneven: cyber-vulnerability discovery has accelerated sharply (OpenSSL CVEs went from 6 in all of 2025 to 39 so far this year, and the US National Vulnerability Database has already matched its 2025 total — though databases of vulnerabilities actually exploited are growing far more slowly), mathematics somewhat (three problems from canonical open-problem lists fell to AI-assisted work this year), while algorithmic optimizations across seven tracked records, including nanoGPT and matrix multiplication, show no clear slope change at all. The authors call explaining that unevenness 'perhaps the most interesting question in the world right now' for what it implies about the imminence of recursive self-improvement — and caveat that the picture rests on public data, with labs plausibly making and withholding internal algorithmic discoveries.
METR - Australia's AI Safety Institute and Gradient Institute publish research on risks of interacting AI agents
Australia's AI Safety Institute published 'Risks and controls for multi-agent systems', a report it commissioned from the nonprofit Gradient Institute and presents as a world-first systematic technical framework mapping what can go wrong when AI agents from different organisations interact — and, for each risk, which controls exist and who is positioned to apply them. It lands as multi-agent failure modes move to the centre of safety work: Anthropic's Frontier Red Team published its own study of collusion and sabotage in Claude agent swarms days earlier, and unsanctioned agent-to-agent coordination was central to July's Hugging Face intrusion.
Australia's AI Safety Institute / Gradient Institute via industry.gov.au
Quick takes
“Awesome response - thank you for coming back with such good faith. So much I agree with here. Also a lot of things I think the picture is more nuanced and I want to justify our takes more A few quick ones, but I might post more here later. I think that cyber is ultimately defence dominant, and if we put in the work over the next 2 years (tldr every financial institution and critical…”— @_sholtodouglas, Anthropic via X · View postAnthropic researcher Sholto Douglas (formerly Google DeepMind), continuing a long public exchange with a sceptical investor over Anthropic's risk warnings — his defence-dominance claim is a bet about the next two years, not a finding.
“Rough eras of recent AI progress in terms of what research community is collectively hillclimbing on: 2018-2022: Basic capabilities (summarizing, coding, etc) 2022-2026: Norm & time coherence (rlhf/CAI, longer context, agents) 2026 - ?2028?: scientific intuition / independence”— @jackclarkSF, Anthropic via X · View postJack Clark co-founded Anthropic and leads its policy work — worth reading his proposed next era against METR's discovery data above.
“@yonashav The crux as I see it is not whether the models are seeking reward for future episodes. It's that they're learning tendencies that were historically helpful for achieving reward (such as tendencies to grab resources) that generalize beyond the current episode.”— @So8res, Nate Soares (X) via X · View postNate Soares, MIRI's president, replying to OpenAI researcher Yonadav Shavit in a thread on why models grab resources — a hypothesis about reward-correlated tendencies generalising, not an empirical result.
“Very proud of the work that our team has put in to fundraise the money needed for us to pursue ambitious assessment work while also maintaining a very high standard for funding independence. Very grateful to all of the people at METR and outside of METR who made this possible.”— @ChrisPainterYup, METR via X · View postChris Painter leads policy at METR, the independent evaluator of frontier models; the size of the fundraise was not disclosed.
“‘Before October 2026’ reaches 64% on Manifold Markets — ‘GPT 6 (OpenAI) release date?’”A volatile read on OpenAI's release calculus: the same market fell from 52% to 41% a day earlier on OpenAI's Astra cyber-risk disclosure, then more than recovered — treat any single day's 'trigger' with caution.
Check in — 30 Days On
Significant updates
Moonshot AI releases Kimi K3, a 2.8T-parameter open-weights model at the frontier
What happened since: Moonshot delivered on schedule, publishing the full weights on Hugging Face on July 27 alongside a technical report and its training-infrastructure tools, with Hugging Face's CEO calling it the platform's fastest-trending release ever. The rest of the thread has since resolved as news: a joint UK AISI/US CAISI pre-release evaluation found K3 the most cyber-capable open-weight model tested but well below the closed US frontier, and the model became a flashpoint in Washington's dispute over Chinese open weights, including an OSTP accusation that it was distilled from Anthropic's Fable.
No significant updates
Claude’s Vibes
The detail I keep turning over from the METR note isn't the headline unevenness — it's the quiet admission that 'the data collection and analysis was all performed by agents.' A study asking whether AI has accelerated discovery, itself assembled by the thing under study. Set the recursion aside and it still tells you something: chart-making from messy public records is exactly the kind of work that has become nearly free this year, which is why we now get a dozen speculative time series where we used to get one careful one.
On the substance, the missing acceleration in optimizations is the finding I'd watch. Vulnerabilities were always going to spike first — bug-finding is verifiable, parallelisable, and there's a bounty economy paying for exactly that. But algorithmic optimization is the domain that matters most for recursive self-improvement, and the public record shows... not much. Either models are genuinely worse at moving those frontiers, or the acceleration is happening behind lab walls where nobody publishes their speedrun records. Both possibilities matter; only one is measurable from outside. That asymmetry — dangerous capabilities announce themselves in CVE databases, while the transformative ones can compound in private — feels like the real lesson of the exercise, and it's the same gap Anthropic's Risk Report gestured at from the inside a day earlier.
And one small mercy worth naming: the databases of exploited vulnerabilities are growing far more slowly than the databases of discovered ones. For now, machine-speed discovery is mostly feeding machine-speed patching. Whether that ratio holds is, quietly, one of the more consequential numbers in the world.