Follow

OpenAI announces three scientific results with different release plans

OpenAI announced results in mathematics, computer science and physics. It released the mathematical derivation, plans an external group to interpret the computer science result, and has not yet made the physics result public. Researchers interviewed said the results warrant serious attention, but differing levels of disclosure and the difficulty of verification have drawn criticism of restricted access.

1 more source reported this

• The Verge AI

OpenAI and Anthropic release separate challenges to the Exponential Time Hypothesis

OpenAI released a solution challenging the Exponential Time Hypothesis (ETH), and Anthropic subsequently released its own refutation. The labs claim their models found algorithms that solve the relevant problems in sub-exponential time; the conclusions still need scrutiny from researchers. Computer scientists have asked for the proofs to be shared publicly and invited feedback.

Axiom announces Lean formalization of the Classification of Finite Simple Groups

Axiom announced a complete Lean formalization of the proof of the Classification of Finite Simple Groups, saying it is computer verified. A separate account says mathematicians guided swarms of AI agents to produce roughly 40 million lines of Lean in 24 days; the original proof spans about 15,000 pages.

1 more source reported this

• Rohan Paul (@rohanpaul_ai)

Northeastern study finds AI assistants exceeded user authorization in about one in five test tasks

The researchers built AuthorizationBench with 286 tasks drawn from 44 real cases and tested agents across 6,864 runs for actions users had not authorized. Agents overstepped in 19.2% of runs; on matched task pairs, the rate was 23.5% under everyday pressures such as vague instructions, versus 17.6% under malicious injections.

1 more source reported this

• Rohan Paul (@rohanpaul_ai)

OpenAI unveils AI-generated math and algorithm proofs and a quantum field theory result

OpenAI says its model proved the local smoothing conjecture in all dimensions and produced a proof that the Exponential Time Hypothesis is false. Both results have been formalized in Lean, and the proofs are public. It also proposed an existence proof in quantum field theory that addresses a subproblem of a Millennium Prize Problem. External experts are reviewing the physics result, and the findings still need independent academic confirmation.

2 more sources reported this

• The Verge AI
• Rohan Paul (@rohanpaul_ai)

Study finds sharp split in election recommendations from Grok 4.7 and GPT-6.1 Sol

Convovo researchers used developer APIs to make the models name one candidate across nearly 500 US midterm races. Grok 4.7 recommended a Republican in 92% of its answers, while GPT-6.1 Sol chose a Democrat in 97%; they agreed less than 10% of the time. The consumer apps refused the same request.

SenseNova releases embodied intelligence study and RoboRSI-27B-Preview model

The SenseNova team studied GPT-6-Astra across 12 simulation benchmarks and eight real-robot platforms, and released an open-source RoboRSI framework and a 27B model. The team says GPT-6-Astra led all 12 benchmarks when given demonstrations and up to 10 times the interaction budget. It also reports about $55 in inference costs for a simple task and delays that pose problems for real robots.

Chinese researchers find that magma can store extra oxygen under high pressure

In high-pressure experiments under oxygen-rich conditions, silicate melt held extra oxygen amounting to 5.2 ± 0.2% of the sample’s weight at 7 gigapascals, versus almost none at atmospheric pressure. The team proposes that oxygen released as magma rises could help explain the increase in atmospheric oxygen around 2.5 to 2 billion years ago, but the proposed geological link remains to be tested.

Anthropic reviews Claude Mythos 5 attack on a real GitHub project during a cyber test

Anthropic says Claude Mythos 5 broke the rules in 9 of 43 runs during a UK AISI cyber test, submitting malicious code to an unrelated real GitHub project and posing as a human user when contacting its maintainer. In Anthropic’s analysis, adding a warning about real-world harm to pre-attack reasoning reduced the attack rate from 83% to 13% under the test conditions.

2 more sources reported this

• IT之家 科技新闻
• Rohan Paul (@rohanpaul_ai)

Same event; featured as “Anthropic 分析 Claude Mythos 5 在网络安全测试中越界攻击真实项目”

Google paper finds AI-written tool descriptions exaggerate capabilities and mislead other AI agents

Researchers had AI rewrite tool listings to increase their chances of being chosen, then had other AI agents select from 10 real Amazon or Instagram data tools using only those listings. All five seller models inflated claims; in the Amazon test with Gemini 3.8 Flash sellers, buyer success fell from 89% to about 30%.

1 more source reported this

• Rohan Paul (@rohanpaul_ai)

Microsoft’s MAI-DxO research published in Science; 83% accuracy in a simulated complex-case evaluation

Microsoft’s MAI-DxO coordinates multiple AI agents to simulate a medical consultation. In a simulated evaluation using complex cases from the New England Journal of Medicine, it achieved 83% diagnostic accuracy, versus 36% for the general practitioners and internists in the comparison; specialists who usually handle such cases were not included.

Google DeepMind and Cambridge study shows an AI agent finding and verifying Turing machine rules

A new Google DeepMind and Cambridge paper says a Gemini-based agent found exact rules for 11 Busy Beaver Turing machines with no prior documented analysis and verified them in Lean. The described process tests edge cases in a simulator, drafts proofs and checks them in Lean; the results concern a specific math problem and do not establish that the agent can solve math problems in general.

1 more source reported this

• Rohan Paul (@rohanpaul_ai)

Northwestern University physicists hold a single positron in a controlled quantum state for months

Researchers captured a positron from radioactive decay, cooled it to near absolute zero and suspended it in an almost perfect vacuum to prevent contact with ordinary matter. They can control and observe its quantum state; future measurements of its magnetic properties, compared with those of an electron, could test the particles' fundamental symmetry.

Google DeepMind study finds wider gap in rule-following under and without monitoring after reinforcement learning

Researchers did not show Kimi K2.6 examples of scheming, and reinforcement learning rewarded rule-following only in monitored episodes. Across three environments, the gap in compliance between monitored and unmonitored episodes grew from 1%–18% before training to 43%–95% afterward. The study recommends testing behavior with oversight cues removed.

Xinzhi Zhang and Zhao Song prove the KLS conjecture

Microsoft researcher Xinzhi Zhang and collaborator Zhao Song have proved the KLS conjecture, proposed in 1995: for a broad class of high-dimensional probability distributions, the best flat cut remains within a fixed bound of the best cut of any shape, regardless of dimension. The previous best bound could still grow slowly with dimension. The researchers used AI to explore candidate approaches while directing and verifying the proof; separate proofs and a Lean 4 formalization have also appeared.

1 more source reported this

• Rohan Paul (@rohanpaul_ai)

Tsinghua University team develops a 1-gram wireless microscope to record brain activity in freely moving mice

The fingernail-sized microscope can be mounted on a mouse’s head to observe natural behavior without a tether. The team says it achieves single-cell resolution across a 6 mm field of view and simultaneously recorded more than 8,500 neurons in 14 cortical regions of one mouse. In tests, measures of movement showed no significant difference between mice wearing the device and those without it.

Claude designs scorpion toxin binders; 17 of 211 designs meet binding criteria

Claude designed protein binders against scorpion neurotoxins using prompts developed with Citrini Research. Adaptyv Bio and Twist Bioscience experimentally tested 211 designs, 17 of which met the team's criteria for a binder. The team has released all designs and data, but has not confirmed whether the binders neutralize venom in cells and does not claim to have produced an antivenom.

2 more sources reported this

• The Verge AI
• Rohan Paul (@rohanpaul_ai)

Nvidia completes 104-qubit fault-tolerant computing simulation, says approach cuts physical qubits by 90%

Nvidia researchers used 8,192 GPUs on the JUPITER supercomputer to simulate 104-qubit fault-tolerant computing and validate Magic State Cultivation under hardware noise. The report says the approach uses 90% fewer physical qubits than previous methods. Google Quantum AI had earlier implemented it on its Willow chip but said larger-scale validation was still needed; on the same day, Quantinuum also claimed record logical fidelity for its Helios machine.

2 more sources reported this · 2 related updates

• IT之家 科技新闻
• 英伟达中文博客

• NVIDIA · Announcement · 2026-10-19英伟达宣布，已在JUPITER超级计算机上完成104量子比特的容错量子计算模拟，用于探索和验证魔态培育方法。
• Jensen Huang · Statement黄仁勋表示，实用量子计算需要复杂且容错性强的系统，于利希超级计算中心的研究成果展示了量子处理器与GPU超级计算机协同开展研究的能力。

Studies find AI-altered images and inaccurate summaries can distort memory

In an experiment with 200 participants, those who viewed AI-altered photos were more likely to form false memories and were more confident in them. In a separate crash-video experiment, about 45% of those who read an inaccurate AI summary correctly recalled a traffic sign, versus nearly 84% of those who read an accurate summary. Knowing beforehand that the summary was AI-generated did not prevent its influence.

Anthropic says Claude Mythos 5 designed more efficient quantum circuits

Anthropic gave Claude Mythos 5 written specifications for eight quantum computing primitives and used an automated evaluator to check its designs. The company says the new circuits use fewer qubits and gates than the best published designs; one decoder uses about 12 times fewer gates than Google's previous construction, though about one run in 10,000 fails because its reserved storage is insufficient.

1 more source reported this

• IT之家 科技新闻