Evgiz

maximally curious software engineer

[Evgiz @Evgiz]

@piercefreeman People are not used to this very weird shape of intelligence it seems. Surprises me too really. One day I'm like omg yeah this is it. Then next day opus does the most stupid thing imaginable.

[Evgiz @Evgiz]

Google AI overview is so bad it's undermining the efforts to make people take AI safety seriously

[Evgiz @Evgiz]

It's scaled tremendously the last year alone. Another "mind is a word reserved for humans and does not match my own narrow definition of reasoning" argument. Look at what these models can do. They are certainly still very stupid in some areas, but yet vastly superhuman in an increasing number of domains.

[Evgiz @Evgiz]

@gmkurtzer Why can't computers be concious?

[Evgiz @Evgiz]

@neuroswish Because of the target audience

[Evgiz @Evgiz]

We were too productive guys, the reward is government assistance for life

[Dissident West @dissidentwest]

The most insane part about what’s about to happen to the job market due to AI is that all of the jobs that were full of people that were actually productive are about to get replaced and the only ones that will remain are the useless adult daycare jobs.

[Evgiz @Evgiz]

Programmers have always worked to automate everyone else. Now that we're getting around to ourselves we can't be mad

[kache @yacineMTB]

It's funny that the only people who seemingly don't hate AI are computer programmers, and they are really the only ones who are getting completely replaced

I guess programming really is that painful

[Evgiz @Evgiz]

RIP jev btw

[Evgiz @Evgiz]

Yann LeCun is making a severe categorical error in his reasoning

[Big Brain AI @realBigBrainAI]

Yann LeCun, Executive Chairman of AMI Labs, explains why LLMs are mostly retrieving human knowledge rather than thinking for themselves:

LeCun starts with why so many people misread what these systems are doing:

"I think there's a lot of confusion, really, because we tend to anthropomorphize systems that can reproduce certain human functions."

When an AI writes fluent answers, we assume there's a mind behind them. LeCun's view is that most of what we're seeing is something more familiar:

"LLMs, to some extent, except for a few domains, are mostly information retrieval systems. They can compress a lot of factual knowledge that has been previously produced by humans and can give easy access to it."

The key words are "previously produced by humans."

The knowledge in an LLM came from people. What the system adds is compression and easy access.

That puts LLMs in a long historical line:

"In a way, it's kind of a natural evolution of the printing press, the libraries, the Internet, and search engines. Right. It's just a more efficient way to access information."

@ylecun is clear about their value:

"LLMs are incredibly useful, there's no question about that. And they do amplify human intelligence, like computer technology going back to the 1940s."

He also allows that in a few areas, such as generating code and some types of mathematics, the capabilities seem to go beyond retrieval. But he notes what those areas share:

"It's still, to a large extent, domains where reasoning has to do with manipulating symbols."

That's where the retrieval framing shows its limits. If these systems were truly thinking for themselves, you'd expect that ability to carry over into the physical world. It hasn't:

"The problem is that why do we have systems that can pass the bar exam and win mathematics Olympiads, but we don't have domestic robots, we don't even have self driving cars."

Then comes his sharpest comparison:

"And we certainly do not have self driving cars that can teach themselves to drive in 20 hours of practice like any 17 year old. So we're missing something big still."

[Evgiz @Evgiz]

@IAmArcIvanov @misraetel So if autoregression is not sufficient for a "self-introspection loop", what is? Would some recurrent state the model controls be sufficient? What about feeding the model's own activations into itself and let it observe them?

[Evgiz @Evgiz]

Why did they have to give them normal human clothes though

[Nicolas Bustamante @nicbstme]

Today I saw the future of robotics, and I’m convinced: robots will clean and cook for us and I can’t waaait!!

Some people call that a useless luxury. My grandmother thought the same thing about washing machines.

The future is cool. Give people their time back and see what they do with it.

[Evgiz @Evgiz]

So... do we all agree that 1 – p(doom) ≈ p(utopia)?

[Evgiz @Evgiz]

@dioscuri I'm just some random engineer, but I was very wrong about CoT/harnesses. I figured there was no way that'd scale well, I figured it had to converge to junk. My intuition is still this, actually, and I remain very surprised by how it actually does work.

[Evgiz @Evgiz]

Introducing humanity's last invention. Global release fall 2027. Terms and conditions apply.

[Evgiz @Evgiz]

@a_musingcat Chill dude its all fine dont worry about it

[Evgiz @Evgiz]

@original_ngv Wdyt about this one?

[Evgiz @Evgiz]

@robbensinger @TravisMcDonald It feels obvious that 1-p(doom) ≈ p(utopia) no? Anyway it's all gonna be fine:

[Evgiz @Evgiz]

@krherr I'm sure we'll solve alignment any day now... Though we don't even know what that could even look like.

[Evgiz @Evgiz]

@Sauers_ Here's mine