Will there be a gpt-4 quality LLM with distributed inference by the end of 2024?
14
109
1k
2025
31%
chance

  • The model has a Elo greater than 1190 on ChatbotArena (or if ChatbotArena is no longer available/updating, achieves GPT 4 (03.14) equivalent or greater performance on both MMLU and MT-Bench)

  • When running inference in a geographically distributed fashion (the computational hardware is not colocated, and is networked over typical consumer equipment)

  • on heterogeneous hardware (the computational hardware is varied in type, e.g. different GPU models)

  • without the act of distributed inference causing the model to require 2 OOM more energy usage (e.g. if doing so Is incredibly lossy and inefficient, it does not count. The burden of proof lies on anyone claiming this clause should be activated)

Note: if (or, when) an edge case is presented, it's applicability to this question will be evaluated in mine + Robert's understanding of the spirit of the question.

Get Ṁ600 play money
Sort by:
predicts YES

@firstuserhere I'm curious what updated your prediction from 50% to 30%?

bought Ṁ15 of NO

Useful markets for betting:

bought Ṁ15 of YES

I'd say rough draft could be (with GPT-4 performance as a baseline):

  1. The model has a Elo greater than 1190 on ChatbotArena (or if ChatbotArena is no longer available/updating, achieves GPT 4 (03.14) equivalent or greater performance on both MMLU and MT-Bench)

  2. when running inference in a geographically distributed fashion (the computational hardware is not colocated, and is networked over typical consumer equipment)

  3. on heterogeneous hardware (the computational hardware is varied in type, e.g. different GPU models).

  4. without the act of distributed inference causing the model to require 2 OOM more energy usage (e.g. if doing so Is incredibly lossy and inefficient, it does not count. The burden of proof lies on anyone claiming this clause should be activated).

  5. Note: if (or, when) an edge case is presented, it's applicability to this question will be evaluated in my understanding of the spirit of the question (is it possible to run inference on a SOTA LLM using a bunch of different peoples computers?).

Note: @firstuserhere my knowledge of LLM's is definitely in the midwit realm - if any of the above clauses (benchmarks?) seem bad feel free to improve. I do suggest we have this question evaluate GPT-4 or equivalent models though as from my read through of the paper, they already have LLAMA 70b running in a distributed fashion which is pretty close to GPT-3.5.

[DRAFT]?

@mattyb Will remove it once the resolution criteria is written

More related questions