@JFPuget You're right!

@wordgrammer @doomslide Unless you work with the reals instead of integers lol

@lacker True, I'm forgetting that these are businesses at the end of the day and technology does not suffice to hold the user base

@GodelPilled I'm pulling the trigger for o3 pro. I have 200 dollars saved for exactly one month of LLM debauchery

@lacker Obviously all of the above hinges on the following question "Can AI reliably write code without human intervention that passes tests within the next 5 years?". And for perplexity, it is "Will other companies eventually find out the secret RAG sauce we have?"

@John_Stalcup_ I deleted the previous post because I made a huge mistake! I forgot to add the factor of 12 to estimate the yearly contribution per user. This gives us a more reasonable 200k yearly paid subscribers. 200k daily active users is still a lot...

@LawrPaulson @Anthony_Bonato That guy keeps posting engagement bait. I muted him lol

@John_Stalcup_ You could argue that even OpenAI will be undercut, and that I agree with that. Gemini is as good for most use cases and is even cheaper than DeepSeek

@John_Stalcup_ @EMostaque @arrakis_ai Absolutely. I'd rather have an AI be able to reliably be able to solve the bottom 90% of problems reliably

@willccbb @doomslide @HolgersenTobias By single pass so we mean computing a single token? Because there exists a finite parameter transformer model which is Turing complete with infinite context length

@littmath @pachabelcanon Nah bro I don't want to waste my time

@5hv5hvnk My belief is the cliché "the price of intelligence will go to zero" due to the cost of computing going down and the number of competitors in the race to create "powerful AI". I think that paying for GPU hours as opposed to GPUs has and will keep them mainstream to consumers

I was testing @browser_use to go through the Vons website and find me a list of avocado prices and it's cute how long it takes to scroll through ads I'll enjoy it because it won't be cute for long

@ezyang Yeah, I don't really see how LLM is improving peoples coding productivity when it cannot see the output of an LSP. I do believe that the technology is still in it's stages of infancy and will eventually be able to "wow" us and produce fully correct code in x seconds

@Stone_Tao That is prudent. Most startups fail

@jachiam0 Computation is the attachment of meaning to a change in physical state. The problem is deriving physical changes that construct the meaning we want to get out of it