3 ms·
They never put the parameter counts in their model names like other AI companies did, but back in the GPT3 era (i.e. before they had PR people sitting intermedi
by derefr 7mo ago
They never put the parameter counts in their model names like other AI companies did, but back in the GPT3 era (i.e. before they had PR people sitting intermediating all their comms channels), OpenAI engineers would disclose this kind of data in their whitepapers / system cards.
IIRC, GPT-3 itself was admitted to be a 175B model, and its reduced variants were disclosed to have parameter-counts like 1.3B, 6.7B, 13B, etc.
- technocrat8080 7mo agoWow, would love to see a source for this.
- derefr 7mo agoTable 2.1 in https://arxiv.org/pdf/2005.14165 https://arxiv.org/pdf/2005.14165.