Published
The Nasdaq fell 2.86% on Friday, its worst day since 5 June, as Intel lost 28.6% on a weak forecast and Amazon’s spending unsettled investors.
The FBI blamed North Korea for the $6.3bn theft from Binance, and the exchange reopened withdrawals on Sunday after replacing most of the stolen assets.
Randall Balestriero and Yann LeCun proved that predictive models trained separately on different modalities learn the same representation up to one rotation.
Andrej Karpathy, who joined Anthropic in May, says his only manual steps were an API key, a server and a domain name. The account comes from an employee of the company that makes the model.
Full storyAmazon expects to spend $220bn this year. Alphabet plans up to $205bn, Microsoft about $170bn, Meta up to $145bn and SpaceX, which includes xAI, up to $88bn. Most of it pays for data centres and chips.
LeCun said large language models need many parameters to hold declarative knowledge, while predictive world models can get by on less memory and data.
Chollet wrote that evolution optimises for generality and robustness, since overspecialised species become fragile when environments shift.
DeepSeek put Super-Agent on GitHub under an MIT licence. It takes coding tasks from team chat and answers with pull requests; the lab calls it a research prototype.
· The Next Web · BloombergRobert Haas, a PostgreSQL committer, blames AI-reported bugs for the late release. Back-patched commits rose by more than 60% between feature freeze and 25 September.
· Robert HaasIn an essay that set off a debate among researchers, Alexander says demanding that safety-minded labs quit the race alone destroys the backup plan if a global pause fails.
· LessWrong · Discussion on XJEPAs trained separately on any number of modalities learn the same representation up to one rotation when positive pairs are two views of one scene. Next-token prediction is shown to be a case with bad views.
A variant of UCBVI reaches expected regret of order \(\sqrt{H^3 S A K}\) for large \(K\), matching the minimax lower bound with no extra polylogarithmic factors for the first time.
Modelling reinforcement learning as an exponential tilt of a base policy, the paper proves that first-try gains cost success on other prompts and force a crossover at large sampling budgets \(k\).
When two tail forecasts are compared by a consistent score, exceedances are the effective sample size. With few of them, the Diebold-Mariano test favours the less conservative forecaster.
On Terence Tao’s blog, Frank Calegari argues that the subject is not finished, and that mathematicians should spend their effort on hiring, publishing and training.
· What’s new (Terence Tao’s blog)The Nasdaq closed at 25,437.00 on Friday 30 October, its worst day since 5 June, after Intel and Amazon deepened worries about the cost of AI.
The FBI named North Korea’s TraderTraitor unit on Saturday. Binance reopened withdrawals on Sunday and reported $3.6bn of net outflows in the first 9.5 hours.
· Al Jazeera · Cointelegraph: FBI · Cointelegraph: outflowsThe expiring December contract settled at $117.64 on Friday and January at $113.43. January fell about 2% in Asia on Monday as talks in Doha continued.
Americans vote on Tuesday and the October jobs report comes on Friday. The Reserve Bank of Australia and the Bank of England decide rates, and 31% of S&P 500 companies report.
· Yardeni Research · Tech Insider