> They saw ratings hover around 60% with their original, in-house tech — this improved by 7-8% with GPT-2 — and is now in the 80-90% range with the API.
> The F1 score of its crisis classifier went up from .76 to .86, and the accuracy went up to 96%.
> With OpenAI, Algolia was able to answer complex natural language questions accurately 4x as often as it was using BERT.
I think the most informative are the first two, but the most _important_ is the final comparison with BERT (a Google model). I am, uh, a little worried about how fast things will progress if language models go from a fun lil research problem to a killer app for your cloud platform. $10m per training run isn't much in the face of a $100bn gigatech R&D budget.
$10m per training run gets me a lot of engineering time to build our own version of this system and lease it to other customers. Just skip one training run and I've got a pretty good team.
Putting aside the question of whether it would ever be a choice between spending $10M on a training run and hiring a team for $10M, GPT transformers were the end result of decades of language research and innovations. You’re making it sound as though you can build the next iteration past GPT-3 for $10M, which I don’t think is the case.
> They saw ratings hover around 60% with their original, in-house tech — this improved by 7-8% with GPT-2 — and is now in the 80-90% range with the API.
> The F1 score of its crisis classifier went up from .76 to .86, and the accuracy went up to 96%.
> With OpenAI, Algolia was able to answer complex natural language questions accurately 4x as often as it was using BERT.
I think the most informative are the first two, but the most _important_ is the final comparison with BERT (a Google model). I am, uh, a little worried about how fast things will progress if language models go from a fun lil research problem to a killer app for your cloud platform. $10m per training run isn't much in the face of a $100bn gigatech R&D budget.