skeskinen

Lead times are so long that there is a lot of risk the chips would be obsolete by the time they come out.

Also, it's hard to get fab capacity for any project. Let alone something so experimental.

Bitter Lesson is about exploration and learning from experience. So RL (Sutton's own field) and meta learning. Specialized models are fine from Bitter Lesson standpoint if the specialization mixture is meta learned / searched / dynamically learned&routed.

I love the Witness, like the Portal games and hate the Talos Principle. To each their own, I guess.

The premise of this study makes no sense. The goal of coffee brewing is to get a very specific extraction at around 20% that tastes good. If you go above the ideal, it's going to taste overextracted.

If you want to maximize extraction at all costs, you can just use immersion brewing with mixing. You don't need fancy pour-over technique.

That said, it's still interesting to learn more about the physics of extraction in pour-over coffee.

Thank you very much! Do report back if you have time to try it. Good and bad, both are welcome

Well I study Ukrainian and was just curious why it's missing, since it's on the supported list for both whisper and OpenAI TTS. I can't find a list of languages supported by GPT-4o anywhere in the release materials or API documents. It's not in GPT-4 technical report either.

I'm native in Finnish and would say that 4o writes sort of ok Finnish but definitely makes many mistakes. I'd imagine many of the languages on the list of supported ones are worse.

I'm not an expert on the subject, but this seems like a very insightful post. I especially liked the part about forecast skill analysis being impossible for tail risks.

Any ideas on how to reduce errors in low resource languages? How did you choose the list of languages to support?

No doubt Google has some awful incentives when it comes to user tracking and making money from ads. But tracking has real benefits for spam and fraud detection. So any attempt at reducing tracking can be siderailed into "personal privacy vs. public safety" discussion.

I don't know if perpetual free trial fits for your definition of genuinely free, but I'm trying to build a competition for Descript at https://smartmediacutter.com/

Hi, matcha.video looks very cool! I'm working on https://smartmediacutter.com which has some overlap in functionality. I'd love to have a chat about matcha and if you have any plans for commercialization, etc.

Thank you very much! Sorry to hear that your PhD wasn't what you wanted. These days it's much easier to provide value outside of academia than inside it.