Amir Efrati

GPT-5 is good. But model performance gains are still slower than in past years and this year has been a technically challenging one for OpenAI researchers. The inside story here...

@santiago__lisa is not a startup

@amir Notably excludes Gemini lol

Anthropic & OpenAI’s share of AI model and application revenues generated by 34 top AI startups is going ⬆️

@edzitron It says right there?

@amir Is this GitHub or just GitHub copilot

GitHub revenue is smaller than many folks assumed

@lukaemon they got millions of AI users before anyone else, and they're getting tons of AI traffic to the core service

@amir since when github is in the lead?? eroding what?

theinformation.com/articles/micro… @aaronpholmes

theinformation.com/articles/anthr… @steph_palazzolo

theinformation.com/articles/cereb… @coryweinberg @julia_hornstein @Katie_Roof

however hard things get, just keep going. hashtag cerebras

@ThinkFinance999 would suggest reading the article :)

@amir What they're saying is something different. It's about token usage being unpredictable (leading to unpredictable costs).

something interesting happening as Anthropic raises prices & its models cost more to use: customers are just eating it.

here's where the Microsoft-OpenAI economics currently stand, incl capex

@_Reeesh people have been saying this for a while. not much sign of it so long as frontier keeps getting better.

@amir I wonder how long this can continue. As older models get better/cheaper and frontier models get more expensive, I can see CIOs routing most of their compute to older/open source models and reserving frontier models for R&D teams

theinformation.com/articles/anthr… @LauraBratton5 @aaronpholmes @KevKubernetes

Remarkable.

theinformation.com/articles/micro… @srimuppidi @aaronpholmes

for all the talk of circular AI deals microsoft sure is getting its money's worth from OpenAI. already 2.5x its investment in OpenAI in new revenue.

@BrennanWoodruff I have no dog in fight but mistral had very different approach/goals

@amir lol gonna fail so horribly. Look at mistral.

deepseek could be valued at $50 b after this megaround 🫰

@Craig_Hughes_ @eric_seufert it was actually true at the time and then the labs figured out new solutions

@eric_seufert ha, also fair!

@andrewarruda fe-static.deepseek.com/chat/transpare…

what data did deepseek use in it's training?