Talkie-1930 launches as pre-1931-trained LLM for AI research
Updated
Updated · Decrypt · Apr 29
Talkie-1930 launches as pre-1931-trained LLM for AI research
6 articles · Updated · Decrypt · Apr 29
Led by Nick Levine, David Duvenaud and Alec Radford, the 13B open-weight model was trained on 260 billion tokens and is live at talkie-lm.com/chat.
The team says its hard 1931 cutoff prevents benchmark contamination by design, while Anthropic provided compute support and two Apache 2.0 checkpoints were released.
Researchers aim to scale the corpus beyond one trillion tokens and build a GPT-3-level vintage model by summer 2026 to study how non-web training shapes AI behaviour.
This vintage AI accidentally learned about the 1930s New Deal. What other future knowledge might be hiding in our historical data?
Since this AI predicted a utopian 2026, what does its failure reveal about the fundamental limits of AI forecasting?
Could an AI ignorant of modern times offer uniquely unbiased solutions to problems we face today?