I know, it's seems extremely unlikely. But there is a person on Twitter claiming this and their tweets don’t trigger my personal “this person is obviously a scammer” or “this person is obviously crazy” alarms (maybe I’m too naive).
Their claim is that there’s a stealth startup working on a new neural network architecture 100x-1000x better than transformers in terms of speed on existing hardware (by speed I mean the amount of tokens generated per second). They will give public access to it in the same way OpenAI did with ChatGPT, and more details will be available in about two months. There are other claims as well, but let’s focus on these.
So this market will resolve YES if before June 1st, 2023 this startup goes out of stealth mode: some kind of public announcement is made and 1) it mentions at least 100x improvement in speed in comparison to transformers and 2) they position themselves as competitors to GPT-4 or any other state-of-the-art LLM at the time of announcement, and that claim should at least be somewhat reasonable based on the evidence provided (I'll be the judge on this).
I’d make the market more technical, but I don’t think I will be able to verify it, because I don’t expect that we’ll get any kind of access to such a model besides chat interface any time soon.
I'd bet YES just because the possibility is too exciting. But my confidence level is low of course.
Maybe transformers will be 100x faster than transformers. OpenAI spent a couple weeks optimizing and already got a 90% cost decrease(10x improvement).
"never get perfect performance" sounds like a red flag to me. what is "perfect performance" in machine learning?
@B I think he is referring to any of the popular benchmarks for LLMs. Perfect performance would be getting 100% on them.
@BairAiushin oh hmm i see. Maybe less of a red flag then but (not knowing the benchmarks) that still sounds very ambitious!
@B it is very ambitious indeed. But AI will get there, the questions is when. Seems quite unlikely that it will happen in 3 months.
I think it's plausible that an architecture 100X faster than transformers will be unveiled, but not from this stealth startup. It might be good to update the title to reflect this?
EG "Will the stealth startup referred to by this twitter user make available a SOTA-competitive LM 100x faster than transformers before June 1st, 2023?"
@NoaNabeshima fair point. But I don't think the probability of someone else coming up with superfast new architecture in 3 months is high enough to be a good reason for making the title this long :)
@BairAiushin I mean suppose that someone overtrains a pretty small mixture of experts model. That on its own might be sufficient.
@NoaNabeshima I doubt it. Anyway, what you're talking about will not resolve this market to YES, so your NO bet is safe :)
@BairAiushin I still think it's good to have titles that match descriptions
@NoaNabeshima if many people insist, I can make it match. I'm not sure people who already made their bets will like it though.
@BairAiushin But if both titles have the same resolution conditions, why would they not like it?
@NoaNabeshima sorry, I think this is too pushy or something
Can we add to constraint (2) to include that you should feel like they're competitors to a SOTA LLM as well? E.g. I can declare myself a super-fast competitor to GPT-4 with a bot that just responds "lol k" to every prompt. Nobody will take me seriously, but I can still do it.
@citrinitas sure, I'll add this.