FLAN 137B (Google Research) and GPT-4 (Mar 2023) (OpenAI) are both frontier AI models. FLAN 137B was published in September 2021 and GPT-4 (Mar 2023) in March 2023.
GPT-4 (Mar 2023) was trained on 2.1 x 10^25 FLOP, about 10.3x the compute of FLAN 137B at 2 x 10^24 FLOP. Training compute is the closest available proxy for how much was invested in a model, though it says nothing on its own about how well that compute was spent.
These two models share no benchmark on which both have been scored, so no direct performance comparison is possible here. The specification table below is a comparison of inputs, not of results.
| Field | FLAN 137B | GPT-4 (Mar 2023) |
|---|---|---|
| Organization | Google Research | OpenAI |
| Published | Sep 3, 2021 | Mar 15, 2023 |
| Training compute | 2 x 10^24 FLOP | 2.1 x 10^25 FLOP |
| Parameters | 137B | 1.8T |
| Dataset size | 2.5T | 5.4T |
| Training hardware | Google TPU v3 | NVIDIA A100 SXM4 40 GB |
| Chips used | — | 25,000 |
| Training time | — | 2.3K h |
| Training cost (2023 USD) | — | $37M |
| Training power draw | — | 19.9 MW |
| Accessibility | Unreleased | API access |
| Open weights | No | No |
| Country | United States | United States |
