lostmsu

Based on

> [Before writing this, Astra thought for 1,034 tokens. That thinking is hidden. The API returns only its 5 headings: Reconstructing the active map / Mapping maze corridors / Preparing exact maze map / Reconstructing row groupings / Checking channel patterns]

This might just be an instance of "A working Nintendo in ~1GB of Qwen weights" from 3 days ago, human-made, than what everyone expects.

https://news.ycombinator.com/item?id=50306442

UPD. I just asked Astra (max) to attempt to do that for https://github.com/lostmsu/TurboGPT (a tiny transformer that trains in under 1min on RTX 5xxx series), and the produced weights were much much worse (bpb 4.47) than the trained ones (bpb 2.53).

Reasoning bit: "I’ll ... encode more specific rules directly in the weights." -- exactly the approach from above.

E.g. no recursive self-improvement here. Still impressive.

This sounds more like "A working Nintendo in ~1GB of Qwen weights" from 3 days ago, human-made, than what everyone expects.

https://news.ycombinator.com/item?id=50306442

But is it slower than Astra?

Nothing new, the models have to fit in RAM+VRAM contrary to the expectation you might have from the title.

> It's all political and ideological choices.

And we already know what choices will be made.

People complained about excessive AI posts. Be careful what you wish for!

MSFT alone is 4T and is probably at least 75% income-from-software.

If either of AI companies "solves software" (does not look improbable rn) ahead of all others, that would defend ~3T valuation right away.

And that's just software.

> That's not really a great investment thesis, unless you're drunk on the hype.

To be fair it is very visibly getting better basically nonstop in the last 2 years at least.

> Chinese models trained with millions

I see you are still under illusion that Chinese models are trained with millions.

The top models still require 1T+ of VRAM. Even if prices go down, that's still gonna be very unaffordable.

I expect people here to root for the future. In the future AI overtaking humans is inevitable.

The ordeal feels like 2yo who hasn't learned yet that the pee will come no matter if he wishes to play or not.

Maybe than Google Search, but not than the Internet.

My Titan Vs that were running non-stop since 2019 for 4 years and then on and off until mid this year exhibit no signs of any kind of issues. I just sold both because they don't have enough VRAM for LLMs.

4x3090s bought in 2021 also still work flawlessly (although they were permacapped at 250W to avoid fan noise).

Well duh, they don't have competitive AI accelerators.

More puzzling is the sale of other stocks that actually do.

msedit ftw

Honestly after VSCode ate a good chunk of vim and emacs users I am not sure why something with similar UX but in terminal is not the standard yet in every *nix. Talk about sunken cost fallacy.

Probably will be UTO

Not under the circumstances IMO

I referred to indirect ownership through owning stakes in the private investment fund(s) that owns their own practice.

Because the article is trash ragebait

> private equitization of dental offices

Owned in the end by the very same dentists, no?

The point it does "apply the same". And sometimes it doesn't apply, but it doesn't apply "the same" between Google and avg Elias - at least as far as I can see.

What's the cache input discount relative to the regular input?

You are misrepresenting what I said. I am talking about fairness of application of law. I am sure Finns are all law abiding citizens, but the point is that regulations are overcomplicated, and a lot of people knowingly or unknowingly violate some a lot of time. In construction in particular.

My claim is that Google is getting unfair (in their favor) treatment is incorrect, because average Elias will also get no or next to no negative consequences making a very similar mistake on average Elias's scale.

The fact that some rules are strictly enforced does not mean all are.

> It's about fairness.

It is a nothingburger from the fairness point as well. People knowingly and unknowingly violate a lot of permitting processes all the time and in many cases never suffer any consequences even if discovered, unless they have zealous neighbors.

So what are the "dubious assumptions"?

I stopped using Windows Terminal altogether because it somehow tends to crash the processes running inside its tabs. And I have a lot of terminals open.

An example issue: https://github.com/microsoft/terminal/issues/16363 . This exact error message with different codes happened to me across many machines at many different times. Never seen anything like that with the classic console host.

In terminal I would do

  gci -F<tab> *.txt -force

In this particular case I would type the word force out but with other parameters starting with `f` I would just use <tab> until I get the right one.

Presumably OpenAI CLI would handle OAuth