Submission + - pre-1931 vintage LLM can code Python (talkie-lm.com)
puzzled writes: Talkie, an LLM trained on text from no later than the end of 1930, can learn to code Python after seeing just a few examples. This puts an end to the thinking that models just memorize really well, instead of actually learning.
Vintage models open a whole range of experiments involving trying to reproduce science that was not discovered until after their training cutoff. This opens up the possibility of exploring just how the frontier models of today might be making new discoveries.
The Talkie model is a 13 billion parameter LLM available on HuggingFace. There's even a 4 bit quantization of it that will run on a 16GB Mac.
Vintage models open a whole range of experiments involving trying to reproduce science that was not discovered until after their training cutoff. This opens up the possibility of exploring just how the frontier models of today might be making new discoveries.
The Talkie model is a 13 billion parameter LLM available on HuggingFace. There's even a 4 bit quantization of it that will run on a 16GB Mac.
