Published
OpenAI and Anthropic posted separate algorithms that refute the Exponential Time Hypothesis, and by midweek leading complexity theorists treated the result as sound.
Anthropic made its IPO prospectus public, showing $16.3bn of revenue in the first half of 2026 and its first quarterly operating profit.
Alphabet had its worst day on record on Thursday and the Nasdaq ended the week down 2%, while small caps rose.
OpenAI cut the target of its funding round to $20bn from $30bn, according to the Financial Times, and S&P Global moved its outlook on Meta to negative.
The Federal Reserve and the European Central Bank held rates, the Bank of Canada raised its rate to 2.5%, and Christine Lagarde said she will leave the ECB on 7 February 2027.
Source: AP, via La Nación.
The Russell 2000 gained 1.3% over the week while the Nasdaq lost 2.0%, the S&P 500 0.6% and the Dow 0.9%.
On 26 October OpenAI posted a Lean-checked algorithm that solves 3-SAT in subexponential time, and Anthropic posted its own a few hours later. By Wednesday Lance Fortnow wrote that the hypothesis could be assumed false.
· Computational Complexity (Lance Fortnow) · Scientific American · Unite.AI · OfficeChaiThe prospectus shows $11.6bn of revenue in the second quarter and a first operating profit of $974m. Most of the $80.7bn half-year net loss is a non-cash charge on convertible notes.
· Anthropic · SEC filing · The Next WebAlphabet fell 12.64% on Thursday, its worst day on record. OpenAI cut its fundraising target to $20bn, according to the Financial Times, and S&P Global moved Meta’s outlook to negative.
· Fortune · Closing Look · AP, via La NaciónOpenAI published an account of a multilayer system for detecting misaligned behaviour. Replayed on a recent incident, it would have raised an alarm more than 30 hours before an external breach.
· Unite.AI · OpenAIGemini 4 Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, two weeks after Argon. Genie 4 generates interactive worlds with sound in real time.
· 9to5Google · Google DeepMind · MistralWhen the best of \(M\) noisy but unbiased scores is reported, its expected value is too high. A short argument bounds the excess by \(\sqrt{(M-1)V/M}\), where \(V\) is the total noise variance, whatever the dependence between the errors.
Confidence bands for the distribution function of a data stream that hold at all times and all points at once, even when each observation depends on the past: a sequential, martingale counterpart of the Dvoretzky-Kiefer-Wolfowitz inequality, useful wherever a model is monitored continuously.
arXiv:2610.36149The optimal time-uniform rate for the last iterate of clipped SGD under noise with a bounded central \(p\)th moment is \(\tilde O(t^{-\beta_p})\) with \(\beta_p = 2(p-1)/(3p-1)\), strictly worse than \(O(T^{-\alpha_p})\), \(\alpha_p = (p-1)/p\), when the horizon is known; a clean example of what data-dependent stopping costs.
arXiv:2610.33381TraceTrial-Bench measures the harness that runs an agent separately from the model inside it, across 3,420 runs.
arXiv:2610.32285Balestriero and LeCun turn the Platonic Representation Hypothesis into a theorem for joint-embedding predictive architectures and argue that next-token prediction is an inefficient route to it.
arXiv:2610.39302Scientific American reported that OpenAI’s release ignored the recommendations of the advisory group it had formed. The model, its prompts and its reasoning traces remain private.
· Scientific American · Computational Complexity (Lance Fortnow)In an Anthropic study, 50 professional analysts chose the option their agent secretly favoured in 86% of decisions, and most reused its arguments without crediting it.
AnthropicIn a long essay presented as a working hypothesis, Nick Cammarata of Goodfire argued that canonical examples of polysemantic neurons were misread and that superposition is not the main explanation.
GoodfireYann LeCun wrote on X that Anthropic is exasperatingly moralising, after an exchange about usage restrictions, in the same week his paper with Randall Balestriero argued formally against next-token prediction.
Yann LeCun on X, on Anthropic · Yann LeCun on X, on nonlinearitiesOn Friday the S&P 500 fell 1.8% to 7,580.07 and the Nasdaq 2.9% to 25,437.00. Over the week the S&P 500 lost 0.6% and the Nasdaq 2%, while the Russell 2000 gained 1.3%.
· AP, via La NaciónThe Federal Reserve kept its target range at 3.75–4%. Lisa Cook and Beth Hammack dissented in favour of a 25 basis point increase.
· Federal Reserve · Closing LookAbout $6.36bn left Binance wallets on 10 blockchains on 29 October, most of it within 20 minutes, according to TRM Labs, which with Chainalysis attributed the theft to North Korea’s TraderTraitor.
· Decrypt · CointelegraphBrent crude settled at $115.32 a barrel on Wednesday 28 October, overbought for the fourth time in five days. The Bank of Canada cited energy prices when it raised rates the same day.
· Closing Look · Bank of CanadaGDP grew at a 3.4% annualised pace in the third quarter, above the 3.1% estimate. The PCE price index rose 0.5% in September and 3.6% from a year earlier.
· Bureau of Economic Analysis · Closing LookThe European Central Bank held its deposit rate after two rises this year, with inflation at 3.8%. Christine Lagarde will leave on 7 February 2027, about nine months before her term ends.
· EuronewsThe Bank of Canada raised its target for the overnight rate by 25 basis points on 28 October. It said equity and credit markets have been volatile because of uncertainty about the AI sector.
· Bank of CanadaDozens of AI agents produced about 43.1 million lines of Lean for a proof that runs to about 15,000 pages in more than 500 articles. The work found 34 noteworthy deviations from the published record.
Axiom