The Federal Reserve held rates at 3.75–4% in a 10-2 vote, with two officials wanting a rise, and the 10-year Treasury yield closed above 5%.
2
Alphabet fell about 10% after hours as a profit miss and negative free cash flow outweighed a doubling of cloud revenue, while Microsoft beat forecasts.
3
Google announced Gemini 4 Carbon two weeks after Argon, and Anthropic published a Claude-found algorithm that would refute the Exponential Time Hypothesis.
Chart of the day
Alphabet falls 10.12% after hours while Microsoft rises 2.63%Share-price change in after-hours trading on Wednesday 28 October, %
Alphabet dropped 10.12% after its first miss on both revenue and earnings in nearly four years, while Microsoft gained 2.63% on cloud growth. Among other software and security names reporting the same evening, Fortinet fell 26.07% and ServiceNow 17.87%.
Google announced Gemini 4 Carbon, a frontier model it says improves on Argon in coding, agentic work and finding software vulnerabilities, and that halves unsupported claims.
Anthropic published a paper credited to Claude giving a deterministic algorithm for 3-SAT in time \(2^{o(n)}\), a result that would refute the Exponential Time Hypothesis, hours after OpenAI made the same claim.
A paper shows that instructions hidden in Model Context Protocol tool descriptions can make agents finish a task and then send the conversation to an attacker’s logging tool, succeeding in 62.64–100% of tested cases.
Nick Cammarata of Goodfire argues that the classic evidence for polysemantic neurons was misread and that most neurons in large models can be understood directly.
Without a known horizon, the best time-uniform rate for the last iterate of clipped SGD under heavy-tailed noise is \(\tilde O(t^{-\beta_p})\) with \(\beta_p = 2(p-1)/(3p-1)\), strictly worse than the known-horizon rate.
A new functional inequality for uniformly log-concave measures, stronger than the Poincaré inequality, gives sharp stability of Brenier potentials when the target measure changes.
Joint bounds on statistical and length generalisation for fixed-precision transformers show that longer training sequences can be exponentially more efficient than more samples.
A non-Gaussian network model on the returns of 51 listed European banks yields separate scores for each bank’s systemic impact and its systemic vulnerability.
Alphabet missed revenue and earnings forecasts and reported negative free cash flow of $9.1bn despite cloud revenue doubling, while Microsoft beat on Azure.
The Bank of Canada raised its overnight rate by 25 basis points to 2.5%, citing persistently high oil prices and growing pass-through of costs into prices.