🏅 Top traders

It's been multiple days, resolving Yes based on the discussion below.

I reviewed the Yes claim from Jonathan for Gemini 1.5 8b and Gemma 2 9b.

I found a website called docsbot.ai which has a comparison feature between models that includes a benchmark comparison table:

First, it looks like gemma-2-9b comes up short in ALL benchmarks, I don't think this reaches "match or outperform":

https://docsbot.ai/models/compare/gemma-2-9b/gpt-4-0314

Next, gemini-1.5-flash-8b only has one benchmark in common with gpt-4-0314, but it does win on that benchmark:

https://docsbot.ai/models/compare/gemini-1-5-flash-8b/gpt-4-0314

I think this is enough to resolve Yes because it matched or outperformed on a benchmark. If I have terribly misunderstood things, maybe you can make an argument the other way. No one made any progress on this when Nathan Young looked at it 4 months ago and I am eager to resolve it.

I'll resolve this in 24-48 hours if no one has any complaints.

@traders Please review the above post, if there is something you disagree with, now is the time to share. I am trying to fulfill the creator's objective from this question. So far @jonathan21m has shown a comparison called LMSYS and I found a site with "Benchmarks" on it, but if this is not appropriate evidence, explain why.

@Eliza I sold my yes a loss a while ago because there were some better end-of-year market opportunities but I still think this resolves yes.

Obviously I agree tons of 15b models surpass the original GPT4 on blind voting. Furthermore the Llama 3.2 11b Vision model gets close to GPT4 on the MMLU even though GPT4 is 5 shot and Llama is 0 shot. Llama 3.2 11b trounces GPT4 on the MMMU.

TLDR; many 15b models surpass GPT4 on multiple benchmarks.

@mods can this resolve yes? According to LMSYS, the March 2024 version of GPT-4 is currently ranked 40th and is outclassed by Gemini 1.5 8b in 31st and Gemma 2 9b in 25.

@JonathanMilligan https://manifold.markets/BoltonBailey/will-we-have-an-opensource-model-be?r=Sm9uYXRoYW5NaWxsaWdhbg

@JonathanMilligan Closing while I try and understand.

@JonathanMilligan Wait, why should LMSYS count? That isn't a benchmark, right?

@NathanpmYoung Yeah LMSYS is a benchmark of how people rank the output of LLM models.

@JonathanMilligan Currently it seems to be the best benchmark of real “in the wild” use.

@JonathanMilligan Okay but it isn't a "benchmark" see "GPT-4's benchmark results as of its release in march 2023." Right?

@mods what do you think?

@NathanpmYoung If you look at the original paper where they introduce LLM arena they refer to it as a “benchmark” https://lmsys.org/blog/2023-05-03-arena/ and on the homepage they also refer to it as an eval benchmark. https://lmarena.ai/

Chatbot Arena: Benchmarking LLMs in the Wild with Elo Ratings | LMSYS Org

<p>We present Chatbot Arena, a benchmark platform for large language models (LLMs) that features anonymous, randomized battles in a crowdsourced manner. In t...

Sebastian Bubeck (Microsoft research) recently said that he thinks a 13 billion parameter model could do this. There've been reports of others working at similar sized models for similar performance.

@firstuserhere Source? Interesting speculation there

@CarterHinsley https://www.youtube.com/watch?v=K0XZ_ShxWkI