Hacker News front pageApr 28, 2026, 05:55 UTC+8

Talkie: a 13B vintage language model from 1930

Nick Levine, David Duvenaud, and Alec Radford released Talkie, a 13B vintage LM trained only on pre-1931 text. The post shows a 24/7 Claude Sonnet 4.6 chat feed and tests surprise on nearly 5,000 NYT historical event descriptions. The key angle is temporal cutoff training as a probe of prediction, bias, and knowledge limits.

Why it matters: HKR-H/K/R all pass: the vintage-1930 framing is memorable, and the pre-1931 corpus plus ~5,000 NYT tests provide concrete substance. This is a strong research release, not a major frontier-model capability update, so it stays in 78–84.

Read the original ↗Export Markdown

Take

Talkie’s useful move is turning knowledge cutoff into a controlled variable. A 13B model trained only on pre-1931 text, then scored on surprise across nearly 5,000 NYT event descriptions, gives researchers a clean way to separate memorized history, extrapolation, and period bias. Modern post-trained models leak future knowledge everywhere, so they are bad instruments for this question.

I like the research shape, but I would not confuse it with a capability release. The article’s own Claude Sonnet 4.6 live feed shows solid answers on the Russian Revolution and Dickens, then repeated outages and a Shakespeare prompt that returns only “~or.” That smells like a cognitive archaeology tool, not a model launch. Radford’s name raises the bar; the hard part is whether the corpus audit and cutoff discipline hold up.

Source: Hacker News front page · talkie-lm.com
Published: Apr 28, 2026, 05:55 UTC+8 · Apr 28, 2026