Home › Hard problems › Inside the model

Inside the model: the core of LLMs

Running out of human text; synthetic data collapse

Solo: possibleOne person can test a narrow slice of it.

good public textnearly used uphuman textgeneration 1generation 2generation 3each model trains on the last one's writingfewer kinds of answer: rare things vanish
Good public text is nearly used up, and training on model-written text can narrow the model each generation ("model collapse").

Why it is still open

Good public text is nearly used up. Training on model-written text can narrow the model ("model collapse").

An experiment one person can run in 8 weeks

Train small models over several generations on their own output, with and without filters. Measure diversity and accuracy loss per generation.

Verdict: Solo: possible

Small-scale results are interesting; large-scale claims need a lab.

The solo test behind the verdict

  1. Can it be measured? There is a number that shows you did better, such as an error rate, a catch rate or a cost per answer.
  2. Is it cheap to try? It runs on a laptop, free open models and open data, or under about ₹50,000 of rented GPU time.
  3. Is there a public yardstick? A benchmark or dataset others use, so your result can be compared and believed.
  4. Would a small win get noticed? A lab, a funder or a customer would care about a 10% improvement, not only a full solution.

Four yeses means Solo: strong. Two or three means Solo: possible. Fewer means Follow, don't build.

PreviousUnderstanding long videoNextEvaluation itself is broken

All 79 hard problems · Working on this, or know of a result we missed? basant@cleandataengineering.ai