In three lines
Published · Updated
Anthropic’s Frontier Red Team says small open-weight models may be months from copying themselves across cheap consumer devices, though its tests were simplified.
Andrej Karpathy argues that the gap between how most people see AI and what professionals now see is widening.
Brent crude settled at $106.65 after new attacks on Saudi Arabia, while US stocks ended mixed ahead of bank earnings.
What researchers are saying
Karpathy describes a steepening funnel in who sees what AI can do
Karpathy argued that most people meet only free chatbots, while a thin professional slice sees agents finish complex projects in hours.
Ganguli recommends Anderson’s ‘More Is Different’ for thinking about AI systems
Ganguli pointed to P. W. Anderson’s essay as a guide to studying complex AI systems as a discipline in their own right.
Chollet likens agriculture to open-source self-replicating nanotechnology
Chollet described agriculture as recursively self-improving, self-replicating nanotechnology that is already fully open source, an analogy for exponential replication.
AI and machine learning
Anthropic says small models may soon copy themselves across cheap devices
Anthropic’s Frontier Red Team reports that open-weight models of 30B parameters or fewer can complete most steps of self-replication on low-cost hardware, and projects capable replicators by the end of 2026.
· AnthropicMarkets
Brent settles at $106.65 after new attacks on Saudi Arabia
Brent rose about 1.8% and WTI about 2.1% to $93.74 on Monday, with energy the strongest S&P 500 sector.
· InvestradeUS stocks end mixed as chip shares fall
The S&P 500 closed at 7,807.24, down about 0.1%, the Dow gained 0.4% to 51,850.55 and the Nasdaq lost 0.4%.
· AP via WTOPEconomy
Markets price a December Fed hike with the 10-year yield near 5.2%
The 10-year Treasury yield held around 5.23% on Friday, off a 24-year high, with traders pricing about an 81% chance of a December rate rise.
Trading EconomicsPapers
New on arXiv todayHarness Learning Enables Generalizable Test-Time Adaptation
Zhang et al.A trained proposer model rewrites an agent’s harness from execution feedback at test time, with no weight updates, and the gains carry over to unseen tasks.