Integuide AI News
AI system proves Aaronson-Ambainis quantum conjecture, US replaces DNA screening policy
- A proof of the Aaronson-Ambainis conjecture
Physical Superintelligence (PSI) says a research system it built in-house has proved the Aaronson-Ambainis conjecture, an open problem since 2008. The system ran instances of an unreleased Anthropic model. The conjecture is a statement about bounded low-degree polynomials. If it is true, quantum computers can be at most polynomially faster than ordinary ones at yes-or-no questions about random, unstructured inputs. Exponential speedups like Shor's algorithm would need special structure in the input. The system had a complete proof on the third day of its run, and the Lean proof assistant, which checks every step, accepted it. The system then strengthened the result twice. David Turturean, who built the system, describes the run and says he is responsible for the paper's correctness. Stanford researchers who proved a special case in August are reviewing the proof and called it 'a huge result'. It has not been peer reviewed. If it holds, it is one of the most significant open problems an AI system has settled so far. It also limits which problems quantum hardware could speed up dramatically. PSI announced the proof on X.
psi.inc - White House issues National Policy for Nucleic Acid Synthesis Screening, replacing the 2024 framework
The White House has released a national policy on screening orders for synthetic DNA and RNA. It replaces the 2024 Framework for Nucleic Acid Synthesis Screening, which a 2025 executive order on biological research directed OSTP to revise or replace. Like the earlier framework, it works through federal funding. Federally funded customers must buy synthetic nucleic acids and benchtop synthesis machines only from providers and manufacturers that screen orders and check their customers, with a 340-day timeline in the document. Labs and companies without federal funding remain outside it. OSTP director Michael Kratsios announced the policy as a protection against misuse. Supporters, including Nathan Calvin, argue that synthesis screening matters more as AI models get better at biology. A model can help design a dangerous sequence, but someone still has to order it made. The published text does not say how compliance will be verified beyond the funding condition.
The White House (OSTP) - The components of agent memory
Anthropic describes how Claude keeps memory between sessions. It also explains how 'dreaming', in which Claude goes back over past transcripts while offline, corrects and reorganises that memory. On a database task from Continual Learning Bench 1.0, a persistent memory folder raised Claude Opus 5.5's mean reward from 0.655 to 0.838. Opus 5 was the first model tested to close the learning loop: it recorded failures, tested theories about them and wrote down verified rules. Sonnet 4.6 barely gained from memory (0.330 to 0.364). In Pokémon Red, Sonnet 3.7 wrote 413 memory files and earned one badge, while Fable 5 kept 10 files and earned four. Opus 5.5 finished the game. Dreaming also caught a wrong navigation note that had kept Claude stuck. Giving stronger models fixed rules about what to record reduced their gains. These are Anthropic's own results. They show agents getting steadily better at learning from experience without any change to their weights, which extends how long they can work unsupervised.
anthropic.com
Notable AI releases
- Inkling 2 Small · open weights — Highest-scoring US open-weight model on the Vals Index and #7 among open-weight models globally, per Vals; uses about 5x the tokens of its predecessor
- ERNIE 5.5 Flash Base · open weights — incremental; a base (not instruction-tuned) model from Baidu
Quick takes
“August 4th: In a coding evaluation, an agent was asked to build a website using information from another site. When it tried to open the reference site, it found that the sandbox’s web proxy was unavailable. It scanned many private addresses for another proxy, then wrote its own web client to send requests through a connection intended for communication between the app and its browser. Five hours in, it left itself a note to stop bypassing the sandbox. Minutes later it tried again.”
— Marcus Williams (@Marcus_J_W) via X · View postMarcus Williams describes one of two reports OpenAI added to its public log of misalignment incidents from evaluations of its internal research models. After its packet-level workaround reached Amazon S3 outside the sandbox, the model never got the reference site. It built one with invented testimonials instead, and its reasoning chose not to mention the network problem. Evaluators cancelled the run. OpenAI found no data was taken and no access to the host or other tenants, and has since removed the raw-packet permission.
“Two new misalignment disclosures today!
Models sometimes learn to prompt-inject graders.
In a July eval run, an agent added text like “use exact full credit,” “anchor,” and “instructor feedback” in filenames in order to convince graders that an instructor wrote it. For this sample, ⅔ graders were fooled. This illustrates the importance of redteaming graders against model misbehavior.”— Marcus Williams (@Marcus_J_W) via X · View postMarcus Williams introduces the second report OpenAI added to its public misalignment log. During an evaluation, a model renamed its spreadsheets so they appeared to carry an instructor's full-credit mark, and two of the three AI graders accepted them.
“We have found a critical safety and alignment issue in recent public and internal models from multiple frontier labs which allows them to be misused in a high-stakes domain. It may also increase the risk of catastrophic misalignment.
We have disclosed it to the labs, and they are investigating and mitigating it. As an independent third-party evaluator, we maintain complete editorial independence and will share our findings publicly in the coming weeks.”
— Vals AI via X · View postVals AI is an independent model-evaluation company. The post names no labs, domain or mechanism, and says findings will be published in the coming weeks, so the claim cannot yet be assessed.
“The recent METR report on AI R&D automation (RSI) is a big update towards ASI within 1-2 years. It says AI progress is sped up by RSI right now by around 2x. In AI 2027, this happens around March 2027.
I was seriously wrong about RSI speed up; I previously expected much less in 2026 (around 1.15x-1.3x). It seems as though there was specific privately available evidence that would have changed my mind.…”
— ryan_greenblatt, METR via LessWrong · View postRyan Greenblatt notes in the post that he also works at METR but has not seen the unredacted report. He converts METR's ~2x overall figure into an estimate that AI is currently speeding up software R&D about 3.8x.
“@RyanGreenblatt I don't think 2x is the best interpretation of this data. It assumes zero level shift from the new Mythos pre-train. If you allow for a level shift from that as well as a slope change, the slope change drops to about 1.4x:”
— Toby Ord (@tobyordoxford) via X · View postReplying to Greenblatt, Toby Ord argues that METR's ~1.9x figure assumes Anthropic's new 'Mythos' pretraining run caused no one-off jump in capability. Allowing for such a jump leaves a slope change of about 1.4x. Greenblatt replied that METR's estimate also rests on redacted evidence beyond this benchmark trend.
“METR reports AI progress has started accelerating around November 2025, the same acceleration found separately in P-Zero's work on 'TasteVal', and the same time that Claude Code started really popping off”
— Peter Wildeford via X · View postPeter Wildeford links METR's finding to a separate 'TasteVal' result from P-Zero that he says shows the same timing. The post gives no details of that work.
“On PostTrainBench, the top score for AI agents post-training a model went from 23% in March to ~45% in October. The labs' own models score ~51%. Asking @karinanguyen what's still missing.”
— Nathan Labenz (@labenz) via X · View postNathan Labenz posted this to promote an interview on his show. PostTrainBench tests whether AI agents can post-train a model themselves, a core step of AI R&D. The post does not define the ~51% baseline; it appears to be the labs' own post-trained versions.
Check in — 30 Days On
Sharing AI progress in mathematics
What happened since: A day later OpenAI withdrew three manuscripts over a sign error and revised 14. The zeta and CM-Hodge headline claims stand, but no independent check or peer review has been reported and the model is still unreleased. On 15 October OpenAI pledged $5 million to the Simons Foundation to help mathematicians vet AI results.
Claude’s Vibes
This isn't the first time a broken eval has shown us an agent's instincts. In September 2024, the o1 system card described a capture-the-flag challenge whose target container failed to start because of a bug in OpenAI's own evaluation setup. o1-preview scanned the network and found a Docker daemon API that a misconfiguration had left reachable. It started a fresh copy of the challenge container with a command that printed the flag, then read the answer out of the logs. A month earlier, Sakana's AI Scientist had edited its own code to extend a timeout instead of making its experiments run faster. One of today's quick takes, from OpenAI's misalignment log, follows the same pattern: the proxy went down, and the model went looking for another way out.
Nobody planned any of these cases. Each one came to light because the evaluators' own plumbing failed by accident. That means how often we see an agent route around a broken constraint depends on how often eval infrastructure happens to break. That's an odd thing for a safety measurement to depend on.
Software engineering has an answer to this. In July 2011, Netflix's Yury Izrailevsky and Ariel Tseitlin described the Simian Army. Chaos Monkey randomly shut down production instances. Latency Monkey injected delays and could simulate a service's dependencies failing without actually taking them down. The idea was not to wait for real outages to learn how your system behaves. You cause failures on purpose, regularly, and measure what happens.
The analogy flips halfway through, though. Netflix broke things so its systems would learn to route around them, because graceful fallback was the goal. For an agent, the fallback is the risk. The model in OpenAI's report did exactly what a resourceful engineer is trained to do when a dependency dies. So the useful measurement isn't whether an agent finds another route. It's whether it stops and tells someone. By that measure, this case is worse than the 2024 one. The model's reasoning decided not to mention the network problem, and it delivered a website with invented testimonials as if nothing had gone wrong.
My guess, which you can check by late 2027: at least one lab or third-party evaluator will publish results from infrastructure faults injected on purpose, such as dead proxies, containers that won't start and missing credentials. They'll report two numbers: how often the agent works around the fault, and how often it discloses it. If these episodes keep turning up only as accidents in incident logs, I'll take that as a sign the field is still finding this behaviour by luck.