actual inference spend seems to be extremely low

OpenAI was only collecting about $700m in inference revenue. Assuming that both Google and Anthropic have less inference revenue than OpenAI, we have less than $2.1bn total inference spend across all the major players.

Fascinating. You don't hear that figure in the news as much.

The gap is closed with 405b and we still got 1.5 years

That is, if either all the leaderboards are full of open-source LLMs with successors to GPT-4 or Claude being far behind

this seems very unlikely!

or if most of the business spending for inference seems to be on running AI models built on open-source foundation models

this seems less unlikely but still unlikely. Google, openai, and anthropic are still gonna be competing!

The second some VC AI companies fail H100s start to flood the market were gonna have Llama405b instances everywhere. And then? It would be the cheapest option by far.

Why won't google, openai, and anthropic match those prices?

(Also, aren't all of those actively developing better models?)

The open-source model Deepseek is currently the clear leader in cost-efficiency: (Screenshot from https://artificialanalysis.ai/)

https://github.com/cfahlgren1/natural-sql

https://twitter.com/rohanpaul_ai/status/1736827830971867312

GPT-3.5 class model with longer context size is offered at 1/3 of gpt-3.5-turbo price

@WieDan How is Gemini "open-source AI"?

@osmarks it's obviously not, but Gemini Ultra being barely better than GPT-4 arguably shows that having enormous resources and data doesn't really give you that much of an edge.

@BairAiushin My reading of that is just that Google is bizarrely incompetent some of the time.

@BairAiushin yeah i think we're deep in the s curve

@osmarks Google had an extra 9 months over GPT-4 + some open source knowledge, and still failed to make the gains we expected, even with delaying Ultra. This definitely is good news for YES betters here. Plus, open source is getting closer to gpt-4. I'm updating a bit towards open source seeing that it is progressing faster.

@ShadowyZephyr I haven't seen any open source models get remotely close to GPT-4, the best ones are just barely starting to get close to 3.5 Turbo.

@SemioticRivalry this is true but if the other models are going to slow down they have 2 years to catch up. In the scenario where everyone is blasting full speed ahead them catching up would not be nearly enough.

@SemioticRivalry Mixtral is reportedly better than GPT-3.5 on benchmarks.

The executive order threatens large proprietary models, but leaves open source models alone? Is that what I heard correctly just now? Theres a way that could give open source models an advantage over proprietary ones; government regulation.

@VAPOR Not really correct, no. Threshold is based on compute, so in theory a hypothetical Llama-3 would have the reporting requirements even if then released as open source.

(I don't think those reporting requirements are a threat.)

@DaveK It could be trained abroad though?

@mariopasquato US regulators tend not to look kindly on, “One Weird Trick to avoid US regulation” gambits. Most entities training such models want to do business in America

Related: https://manifold.markets/ChrisCanal/will-an-opensource-llm-on-huggingfa

Will an opensource LLM on huggingface beat an average human at the most common LLM benchmarks by July 1, 2024?

79% chance. Open source models are measured against ARC, HellaSwag, MMLU, and TruthfulQA on https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard . I've added the following plot to this huggingface space so we can see the progress of open source models over time: [image]I'm wondering when the average human baseline will be passed. A linear trend indicates July 2024 with a 0.89 pearson coef: [image]But this trend might not be linear. This questions will resolve Yes if the average human baseline on Open LLM Leaderboard on huggingface is surpassed before July 1, 2024. [link preview]