this post was submitted on 13 Jul 2024
1 points (100.0% liked)

TechTakes

1385 readers
17 users here now

Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.

This is not debate club. Unless it’s amusing debate.

For actually-good tech, you want our NotAwfulTech community

founded 1 year ago
MODERATORS
top 24 comments
sorted by: hot top controversial new old
[–] [email protected] 0 points 3 months ago (2 children)

The document describes a project that uses Strawberry models with the aim of enabling the company’s AI to not just generate answers to queries but to plan ahead enough to navigate the internet autonomously

Navigate the Internet that their software has already filled with shit. And the point of that would be...?

[–] [email protected] 0 points 3 months ago

It can seamlessly navigate all the way up its own ass, an artisanal reverse-Ouroboros.

[–] [email protected] 0 points 3 months ago

If they could make something that could navigate past all the bullshit they are producing, that'd be a really shrewd business move in a completely sociopathic end-stage capitalism sort of way.

[–] [email protected] 0 points 3 months ago* (last edited 3 months ago) (8 children)

LLMs don’t do reasoning. Facts are not a data type in an LLM — the data is just tokens and the LLM tries to predict the next token. An LLM can look a bit like it’s reasoning as long as the correct answer is already in its training materials. [arXiv, PDF; arXiv, PDF; CustomGPT]

This is a weird kind of assertion. First of all. You could make facts a token value in an LLM if you had some pre-calculated truth value for your data set. That's not how it works now but it's a weird assertion to make about an unknown new generation of AI. As the author points out, facts kind of are a data type, it's just that the AI considers the most related words to the prompt to be the most correct, which of course, with a bad data set they are not.

Also, the current generation of ai, as admitted by the company, is not meant to be a tool for finding facts. It's a tool for generation, yes, a bit like an auto-complete but for natural language and with a much much wider scope.

What Strawberry apparently is, is a machine that reasons, which is NOT similar to what Open-AI ever claimed ChatGPT ever was. It's like a guy promised to bring a new animal to the village that will be able to pull the plow and the author is saying "this guy's full of shit! We have cats all over the village and even the biggest one could never pull a plow! They aren't designed for it! All animals are good for is catching mice!" And the guy brings in an Ox.

Edit: honestly my opinion of AI is lukewarm. I'm with a lot of people that the hype of it now being integrated into all sorts of nonsense is stupid. Its just that all of the bad arguments against it makes me tired.

[–] [email protected] 0 points 3 months ago* (last edited 3 months ago)

That’s not how it works now but it’s a weird assertion to make about an unknown new generation of AI.

Your quote is about LLMs, not "an unknown new generation of AI". Research on LLMs dates back at least to 1990s, they're neither new nor unknown.

What Strawberry apparently is, is a machine that reasons, which is NOT similar to what Open-AI ever claimed ChatGPT ever was.

Sammy boi claimed waaay wilder things than just "it reasons" about his grift so gtfo, this is some incredibly disingenuous apologetics for OpenAI.

Also, "apparently" is such a load-bearing word there.

[–] [email protected] 0 points 3 months ago* (last edited 3 months ago) (1 children)

I'd like to see any evidence of openai concretely saying what chatgpt is for.

[–] [email protected] 0 points 3 months ago
[–] [email protected] 0 points 3 months ago

First of all. You could make facts a token value in an LLM if you had some pre-calculated truth value for your data set.

An extra bit of labeling on your training data set really doesn't help you that much. LLMs already make up plausible looking citations and website links (and other data types) that are actually complete garbage even though their training data has valid citations and website links (and other data types). Labeling things as "fact" and forcing the LLM to output stuff with that "fact" label will get you output that looks (in terms of statistical structure) like valid labeled "facts" but have absolutely no guarantee of being true.

[–] [email protected] 0 points 3 months ago* (last edited 3 months ago)

It can figure anything out with perfect reasoning, provided that that it has an oracle to verify facts.

For my next trick, watch me collapse the polynomial hierarchy [1]

^[1]^ ^quadraticaly^ ^faster^ ^than^ ^the^ ^time^ ^your^ ^mom^ ^sat^ ^on^ ^it^

[–] [email protected] 0 points 3 months ago

@Donkter @dgerard have you bumped your head?

[–] [email protected] 0 points 3 months ago (1 children)

First of all. You could make facts a token value in an LLM if you had some pre-calculated truth value for your data set.

Why stop there? Just make consciousness a token value. See, AI is such an easy problem that me, a non-expert, is able to solve with just one sentence.

[–] [email protected] 0 points 3 months ago

First of all. You could simply prompt the LLM to become conscious, and I bet none of you so-called AI skeptics have noticed that Open-AI has NEVER included text like that in any of their system prompts.

[–] [email protected] 0 points 3 months ago

This is a weird kind of assertion. First of all. You could make facts a token value in an LLM if you had some pre-calculated truth value for your data set. That’s not how it works now but it’s a weird assertion to make about an unknown new generation of AI.

fuck almighty where do you people pull this absolute horseshit from

What Strawberry apparently is, is a machine that reasons, which is NOT similar to what Open-AI ever claimed ChatGPT ever was.

well shit, you’ve got a long list of supposed AI researchers to tell that to. here, I’ll make sure you’ve got plenty of time to read their back catalog of utterly fucking stupid claims!

[–] [email protected] 0 points 3 months ago

Reading and listening to fans of AI has actually managed to convince me that machine learning algorithms really are better at reasoning than some humans.

[–] [email protected] 0 points 3 months ago (4 children)
[–] [email protected] 0 points 3 months ago (1 children)
[–] [email protected] 0 points 3 months ago

Epistemic Status: rekt

[–] [email protected] 0 points 3 months ago* (last edited 3 months ago)

They also said that the Earth isn't a perfect cube with four equal sides!

[–] [email protected] 0 points 3 months ago (1 children)

The Wikipedia editors who hang out here might thank you for calling to their attention pages that clearly need to be fixed, since they cite non-peer-reviewed preprints on the arXiv, the shit journal Entropy, and the fucking LessWrong blog.

Me, I'm going to block you anyway. Bye!

[–] [email protected] 0 points 3 months ago (1 children)

don't you know, the field moves so fast that preprints are just as good as articles and requiring actual scrutiny just stifles SV innovation

[–] [email protected] 0 points 3 months ago

(i vaguely remember that somebody allegedly important in machine learning started doing just that, that is outputting multiple preprints and leaving it at that stage instead of pushing normal papers, but can't remember who)

[–] [email protected] 0 points 3 months ago (1 children)
[–] [email protected] 0 points 3 months ago

'planes could not fly' is some sort of weird pro AI argument. It is the 'we are still early' or 'people were also against the internet when it still arrived' made famous by cryptocurrencies.