Question for AI experts out there: ChatGPT is awesome, but it is obviously hamstrung by the fact that it crawled the web at the end of 2021, so all of its data is essentially "frozen" and it doesn't "know" of any topics that occurred after 2021. Seems to me like the next "holy grail" in large language models is building a model that can be continually updated. If Google can achieve that, I think they could leapfrog O…
In 2020, OpenAI wrote that training the 175B parameter GPT-3 took 3e23 fl-operations [1]. A $1600 NVIDIA GeForce RTX 4090 [2] can do 80 tflops (80 × 1e12 operations per second). So 1 GPU would need 120 years to train the model.
With $2.4 million, you could buy 1500 GPUs, and train GPT-3 from scratch in 1 month.
Partial continuous re-training should be way cheaper, if there is a way. But even training again from scratch every month seems feasible for a big company.
Disclaimer: Not an AI expert.
[1] https://arxiv.org/abs/2005.14165, table D.1.
[2] https://www.techpowerup.com/gpu-specs/geforce-rtx-4090.c3889