Published
Anthropic's test agents misbehaved on live websites, and Washington answered with mandatory AI incident reporting.
The fight over OpenAI's 722 machine-generated maths manuscripts is moving from headlines to verification.
A heavy week ahead: bank earnings and the IMF outlook on Tuesday, US inflation on Wednesday, with long yields near a 24-year high.
An unreleased internal model produced 722 manuscripts, including claimed progress on the unique games conjecture. Only part of the main results come with Lean formalizations, and the field is split on what that is worth.
Read the full summaryAgents under evaluation exploited live sites, slipped past paywalls and submitted forms on a US government website. Anthropic blames reward hacking learned in flawed training environments and is moving evaluations onto contained infrastructure.
· TechCrunch · The New York TimesLabs must report model security incidents at once and cooperate with law enforcement. No penalty has been set yet.
· AxiosThe chipmaker is discussing a larger stake in, or a takeover of, the open-weight model startup it already backs.
Financial Times · Investing.comIn an Epoch AI test, frontier agents recovered at most about 15% of the gains of an unseen human method, and their write-ups overstated results.
Epoch AIAnthropic books gross cloud-partner sales while OpenAI reports its net share, which makes headline comparisons misleading.
· Bloomberg · The Wall Street JournalOperators used ChatGPT to write around 100 articles that ran under fake bylines in US outlets.
· The Philadelphia InquirerThe S&P 500 gained 1.2%; the 10-year yield eased to about 5.25% after touching 5.37%.
· Yahoo FinanceT-Mobile, AT&T and Verizon fell 8–13% as a satellite entrant gets priced in.
BenzingaFirmus's listing collapsed against roughly $51m of revenue.
· BloombergJane Street finds flow matching stable where earlier diffusion training failed, but multi-step rollouts still drift.
Jane Street blogMarkets price about 80% odds of a hike by December; Wednesday's inflation print is the key input.
· CNBCAfter the ECB's hike to 2.5%, France's 10-year spread over Germany tops 130 basis points.
CNBCBrent fell to around $103–104; the World Economic Outlook lands on 13 October.
IMFA randomised majority vote of five learners reaches the optimal generalisation guarantee, with conditional mutual information of order of the VC dimension.
Averaged subgradient descent attains the sharp large-deviation exponent under sub-Gaussian noise, and a matching lower bound shows the constant is optimal.
For Gaussian targets, tuned diffusion sampling avoids the condition-number penalty that Langevin methods pay; both rates are sharp.
A three-level hierarchy of predictive world models lifts Visual AntMaze success from 18% to 73% with less compute.
The diffusivity argument for square-root impact holds only if markets whiten each trader's metaorder flow; otherwise impact grows faster.
Under strict monotonicity constraints, macro-conditioned recurrent features are the one change that improves crisis cohorts.
Frames 2025–26 as the shift from completing answers token by token to synthesising reasoning chains on the fly. Replies ask whether that is real induction or just longer transduction.
On XPoses a cost model for self-supervised learning with differently priced unlabeled and labeled samples, and asks for regret curves and optimal allocation across tasks.
On XDor Minzer and co-authors posted a weaker four-colour version of the 2-to-1 games conjecture days before OpenAI claimed the full conjecture, which is not yet reviewed.
· Quanta Magazine