I gave Groq a chance a few months back because of the speed and cost. Noticed they have aggressively quantized all their models to the point of not worth it. Seems like they are still doing it

I live in Bulgaria. I paid about $100 in property tax last year, or effectively around 0.03% market value. Might sound awesome but creates all kinds of issues with people refusing to sell their unused homes. Half of the apartments in Sofia are empty as it costs next to nothing to hodl

Absolutely wild that you can just prompt and get the system prompt. Maybe obfuscation worsens the usability?

I like it. If I have to have censorship, I'd prefer it to be in-your-face. Putting effort into making censorship more palatable is more repulsive in my mind.

How to decide what tooling is worth experimenting with? The learning curve is never-ending; a substantial amount of time (and money!) are needed for a fair comparison. How did you land on your current tooling?

Maybe the point isn't to directly compete with the likes of OpenAI and Anthropic. The marketing is... something different and that's an easier sell than having direct competition.

I'd guess that the generous free tier is a strategy to establish a bigger market share. Which is fine - developers should experiment and no cost while it lasts. But I doubt it'd stay free

Some of the best actionable advice around agents does seem to revolve around managing context rot. Glad to have a term for it

For me "jotting things down" is more about formalizing my thoughts; as my mind tends to wander off. In a sense, by writing something down, I commit my mind to a direction that could help me with the complex task at hand.

I'm not even sure what "solved" RAG is. Does that mean there is a proved best way of chunking, storing and retrieving text documents? In my mind those are too use-case specific to have a "best" way

Being anonymous means there's a chance of being a bad faith actor. Would the research in question be okay if it was a persuasion study with lies crafted by humans instead of LLM?

That's awesome! Very excited to see what people come up with

As far as I understand it, the main idea is that the test set is private and won't be released. You will be able to see your score at the end of the competition.

Sure, I agree.

Imagine you claim you've "built AGI" without a clear criteria for what AGI is - this would certainly seem out of place. Even if you've done something remarkable - not having a definition would hurt you since your competition will move the goalposts and diminish your achievement.