Integuide AI News
Digest: Chinese DUV/EUV chipmaking forecast, Fable 5 capabilities, an O*NET for AI R&D
Today's lead is a detailed forecast of China's domestic photolithography progress and what it means for compute access, alongside a deep look at the capabilities of Anthropic's now-suspended Fable 5, a proposal to map the tasks of AI research itself, and fresh alignment research on distilling misaligned models.
- A forecast of Chinese DUV and EUV photolithography progress
The AI Futures Project forecasts how fast China can advance domestic deep-ultraviolet and extreme-ultraviolet photolithography, drawing on what ASML's trajectory does and doesn't reveal. The pace of indigenous chipmaking tooling bears directly on whether export controls can durably constrain Chinese frontier-scale compute.
Hamish Low via AI Futures Project - Claude Fable 5 and Mythos 5: Capabilities
A detailed review of the capabilities of Anthropic's Claude Fable 5 and Mythos 5, made unavailable just three days after release when the US government imposed export controls in response to a jailbreak. The write-up assesses what the models could do — and the welfare and safety questions they raise — even as access remains suspended.
thezvi.wordpress.com - Toward an onet for ai rnd
Epoch AI argues for building an O*NET-style taxonomy of the tasks that make up AI research and development, to track which parts of the R&D pipeline AI systems can already perform. Such a map would sharpen measurement of automation of AI research itself — a key input to forecasting recursive capability gains.
Epoch AI (Gradient Updates) - The distillation double bind: Distilling misaligned models either transfers misalignment or it doesn't
Redwood Research argues distilling a misaligned model into a student presents a double bind: either misalignment transfers (yielding a model that may be easier to incriminate) or it doesn't (yielding a capable benign replacement). The framing has practical implications for using untrusted models safely.
Alek Westover via Redwood Research