gbasin

you're doing what now?

me too except codex writes the code and i only look at git diffs :/

I don't believe that's true with 100% grass fed beef

I still use this regularly :) thanks for building it. any interest in open sourcing? i can help

I see, thanks. Seems just a matter of cost to reclaim it. Maybe regulation is needed

pardon my ignorance, in what was does it "consume water"? water cooling is closed loop

the main bottleneck will be model depth... you can only do so much with N layers, and recurrence has proven to be way less efficient (for now)

~210 calories

yep i switched to taza

i eat half a bar of it per day, amazing brain fuel

yep it should be possible. check this out: https://terraformindustries.com/

this was my main takeaway from 20th century philosophy lol

Yes I was specifically thinking of the latter with larger models, I think many many shot ICL still tends to outperform but you're right it's worth trying both for your use case

I am building in the mortgage origination space and have these sorts of enhancements on the drawing board... you hit the nail on the head that the bottleneck will be in QC and legal review, as 3rd-parties (especially regulators) may want to manually see the data you used to reach the conclusions you did. Although I'm still digging. You'd be interested in what we're creating btw, it's a crypto-based mechanism that enables instant digital mortgage origination, we should chat!

This should not be down-voted

For large models that are well-trained with large context support, ICL seems to work about as well as FT

great comment

what would it mean to plan ahead? decoding strategies like beam search are popular and effectively predict many words ahead

Strictly for research purposes, what would some of these offshore sites be?

Today's revenue is not tomorrow's, and companies are valued based off of the future. Where did Amazon's revenue come from (and still primarily, afaik) for most of its life?

Your conclusion may be true but your examples aren't. You can definitely predict the stock market based on past prices, and I suspect you can with weather as well.

TikTok's success is primarily about the creator tools, not the viewing experience. Making video editing mobile-friendly and accessible to more people is what enabled the proliferation of short-form content.

here's one, just google around for "how to access jhanas": https://www.lionsroar.com/entering-the-jhanas/#:~:text=If%20....

Jhanas are achieved a particular way and should be readily accessible once your concentration is strong enough (yours likely is). You may have not experienced them if you haven't accidentally pulled the right levers, so to speak. There are a few guides online of what to do, I suspect you'd succeed quickly

Yes, insofar as awareness of it may influence your own behavior which gives it some predictive power

and it's probably not the one you think

the primary cost would be the OpenAI GPT-3.5-turbo API, which is ~ $0.002 per 1k tokens

it's not... it's competing in the market for profits. citadel makes plenty of money from trading, separate from fee-based business

The latencies and costs are changing so quickly...

you don't need access to the weights though, unless I missed something? this is akin to the various groups using GPT-3.5/4 to generate data for fine-tuning smaller models