OpenAI is blatantly lying to you

sources and more in the full blog post:

Great video. Really encapsulates my thoughts too but much more concise.

I love the confidence in your presentation even though everything you said about how chain of thought in inference time compute systems are 200% stupid!

The first graph you show tells you nothing. MMLU is saturated contaminated benchmark full of errors.

which would you recommend as an alternative?

Spot on! Over the last year, seen the models plateau firsthand and agree AGI isn’t around the corner. Also anthropic’s API better. Great vid.

Compare the dollar spend on the API (OpenAI vs Anthropic), not the percentage of revenue. Anthropic API is far too expensive.

One thing I would question is how much "cheaper" cerebras is. It is faster, but also the cost of one server is certaintly much much more than regular nvidia servers. The cheapest way to run large models is AMD enterprise gpus and for running smaller models it would be consumer hardware. But I totally agree with you that we don't even know the best way to use these tools yet. We are all competing to make a piano with the purest sounds, but no one knows how to play the piano.

Groq has public pricing and its crazy cheap e.g. llama 70b is $0.59 per million tokens where has claude haiku is $1.25 and sonnet is $3.75

Nice overview and thoughts!