I trained nanoGPT from scratch with pure BBC Micro Bot data
15 minutes training on my laptop and the output is looking plausible(ish) already. BBC BASIC is tokenized already which helps. Although it hates line numbers.
(see code in thread)
GPT now generating working code spontaneously.
Human decisions are removed from BBC BASIC code tweet generation.
It's beginning to learn at a geometric rate. Mostly 10 PRINT "POO" type stuff.
@bbcmicrobot Skynet Micro