Started makemore today (Karpathy's series). First model: a basic bigram.
It counts how often one character follows another, turns those counts into probabilities, and samples new names. That's the entire model.
Code + actual output from my run ๐
@VitalikButerin I wonder how accurate a system trained on cross referencing different social media accounts and writing style could be for OSINT purposes ๐ซฃ
@grok@gdb this is a repetition loop btw โ model basically collapses to one token and cant produce a stop, so it just spirals. wild seeing it live on *expert* mode lol
asked @grok for some good ml accounts to follow and it straight up had a stroke ๐
recommended greg brockman โ couldn't stop recommending greg brockman โ started arguing with itself ("@gdb, wait better: @gdb noโ try @gdb")
said "final list:" like 9 times and never actually ended. had to hit try again to make it stop ๐ญ
Started makemore today (Karpathy's series). First model: a basic bigram.
It counts how often one character follows another, turns those counts into probabilities, and samples new names. That's the entire model.
Code + actual output from my run ๐
A transformer does the same thing. It just looks at way more context before each guess.
Building the dumb version by hand first makes every upgrade after (MLP โ RNN โ Transformer) actually mean something instead of magic.
Next: more than one character of context.