Lower-precision training speeds up Inkling 2 Small as Cursor details its extension incident
Thinking Machines Lab has published the training recipe and speed measurements for Inkling 2 Small. Cursor has updated its account of the tampered extension's reach, and the Federal Reserve's fall survey puts AI at the top of the potential shocks cited by respondents.
01
Research and engineering
1 story
2026-11-13Thinking Machines Lab
Inkling 2 Small uses four-bit forward passes in its expert layers
Thinking Machines Lab published the training recipe for Inkling 2 Small. Its mixture-of-experts layers use NVFP4 for forward passes and MXFP8 for the backward computations that produce gradients. The lab kept the same precision settings from pretraining through the final reinforcement-learning rollout.
In its largest jobs on NVIDIA GB300 hardware, the lab reported an 8% wall-clock speedup for pretraining and 20% for reinforcement learning versus full MXFP8 training. At the same loss, pretraining required about 2% more floating-point operations.
02
Security and platforms
1 story
2026-11-13Cursor
Cursor details exposure from the tampered extension
Cursor released its findings on the tampered Claude Code extension distributed on October 29. It said a Cursor-issued access key exposed by vendor Fireworks led to the exposure of historical coding session traces. Fewer than 0.05% of users outside Privacy Mode had traces exposed. Measured separately, fewer than 0.01% of customers had API credentials exposed.
Cursor estimates that up to 5,000 users who requested specific extension versions that day may have received the tampered build. The company contacted affected customers, rotated credentials, rebuilt its VS Code extension marketplace and moved inference traffic off Fireworks.
03
Industry and policy
3 stories
2026-11-13Federal Reserve
AI risk was cited by 87% of contacts in the New York Fed's fall survey
The Federal Reserve's November Financial Stability Report includes a New York Fed survey of 23 market contacts in September and October. Some 87% cited AI as a potential shock over the next 12 to 18 months, making it the most frequently cited risk. In the spring survey, 50% of 20 contacts mentioned AI.
The same report puts advanced-computing-related debt issuance above $680 billion by October 2026, roughly five times the amount for all of 2023. Such borrowing accounted for about 31% of 2026 nonfinancial investment-grade corporate debt issuance.
2026-11-13KBS, TechCrunch
South Korea says Nvidia reaffirmed a plan to supply 260,000 GPUs
KBS reported that South Korean Science and ICT Minister Bae Kyung-hoon said on November 13 that Nvidia had reaffirmed plans to supply 260,000 GPUs and continue existing cooperation projects. Of roughly 35,000 GPUs the government plans to obtain directly from Nvidia, about 21,000 have been secured, with the remainder expected by early next year.
Separately, TechCrunch reported that a petition calling for Huang's reinstatement as CEO had drawn more than 115,000 signatures by Saturday afternoon. Signers included Elon Musk and Replit CEO Amjad Masad.
2026-11-14The Decoder
Anthropic's CEO discusses buying compute one to two years ahead
In a November 13 pre-IPO conversation with JPMorgan Chase CEO Jamie Dimon, Anthropic CEO Dario Amodei said compute typically has to be bought one to two years ahead, before the company knows whether future revenue will justify the expense.
The Decoder's account of the conversation says Amodei described limiting upfront purchases and adding capacity on the spot market if demand runs ahead of expectations. He also cited building a cash buffer as a reason to go public, to help manage a possible timing gap between compute spending and revenue.