These models are black boxes right now, but presumably we could open it up and look inside to see each and every function the model is running to produce the output. If we are then able to see what it is actually doing and fix things up so we can mathematically verify what it does will be correct, I think we would be able to use it for mission critical applications. I think a more advanced LLM likes this would be great for automatically managing systems and to do science+math research.

But yeah. For right now these things are mainly just toys for SUSSY roleplays, basic customer service, and generating boiler plate code. A verifiable LLM is still probably 2-4 years away.

I suspect that GPT4 started with a crazy parameter count (rumored 1.8 Trillion and 8x200B expert “sub-models”) and distilled those experts down to something below 100B. We’ve seen with Orca that a 13B model can perform at 88% the level of ChatGPT-3.5 (175B) when trained on high quality data, so there’s no reason to think that OpenAI haven’t explored this on their own and performed the same distillation techniques. OpenAI is probably also using quantization and speculative sampling to further reduce the burden, though I expect these to have less impact on real world performance.