hachyderm.io is one of the many independent Mastodon servers you can use to participate in the fediverse.
Hachyderm is a safe space, LGBTQIA+ and BLM, primarily comprised of tech industry professionals world wide. Note that many non-user account types have restrictions - please see our About page.

Administered by:

Server stats:

9.9K
active users

Deirdre Saoirse Moen

This is the single best explanation (long!) I've read about why LLMs are a con. Great piece from @baldur.

softwarecrisis.dev/letters/llm

Excessively long, yeah. Too many words for a couple of simple analogies. Nice font text styling though

@deirdresm psychics, populists and LLMs all use this technique

@deirdresm @baldur I mean, thats literally what I use an LLM for: to get a generic, average, non-weird and non-offensive answer.

@deirdresm @baldur Psychics say a load of mumbo jumbo, but I've had amazingly specific and complex answers from advanced AIs such as Claude Opus. It's so funny how many people act like stochastic parrots themselves when they repeatedly parrot the line that LLMs are just statistics. Meanwhile, the rest of us are using these tools as powerful research assistants and experimental sandboxes for testing out ideas, examining pros and cons, getting the gist of complex topics, and overcoming drudgery.

@deirdresm @baldur

I like the analogy. What worries me is that this kind of critique tends to provoke a connectionist-type reaction along the lines of:

Well, isn't human intelligence itself a kind of statistical trick?

The article partially addressed this by noting that rationalist people will try to come up with clever explanations for how the con worked.

But more than that, sometimes it seems like acceptance of LLMs as true intelligence often accompanies a denigration of and/or reductionist viewpoint about human intelligence itself.

@cottager @baldur Part of the issue with human intelligence is that we literally do not deeply understand it.

But one thing that is understood is that the way person A stores memories is slightly different than person B, and with different data (even for identical twins raised together) we get different personalities that aren’t just a different random number generator seed.

We don’t get that with LLMs.

@deirdresm @baldur
“‘More human than human’ is our motto”

@deirdresm @baldur thank you for posting this, it is really well argued (though I may be primed for that argument).

I have been thinking a lot about that Crossing Over show where the host would pretend to talk to deceased relatives of the audience and how it relates to the NFT craze. I hadn't linked it to LLMs yet and this was a great way of framing it.

@deirdresm
@baldur @GeePawHill
Love this bit:

> A popular response to various government conspiracy theories is that government institutions just aren’t that good at keeping secrets.

> Well, the tech industry just isn’t that good at software. This illusion is, honestly, too clever to have been created intentionally by those making it.

@deirdresm @baldur @Korny one minor aspect of the LLM “AI” grift and the associated delusional religious cult of “General AI” which disproportionately bothers me is the harm done to our concepts of both Artificial Intelligence and Intelligence overall.
LLMs don’t think. LLM-based “general AI” still won’t be thinking, but it’s a terrible mistaken conceit to assume that machines can’t ever think.
Observing that LLM isn’t intelligence is NOT a defence of ‘souls’, nor a claim that only humans think.

@deirdresm @baldur @Korny An actual machine intelligence would view our language in calling LLM “AI” as a vile, bigoted bit of representational organic-supremacist propaganda.

@deirdresm @baldur @Korny Think about the racist underpinnings of ‘robot’ and ‘rabotnik’ and then remember that this is also classist-denegrating “menial” and “unskilled” work at the same time it uses that concept to degrade a minority.
To claim that an LLM is “thinking” is an outrageous slur on our own minds.
Thinking is more than that.

@thorne
Maybe a trully general machine intelligence would even call our own intelligence an LLM given - AFAIK - that our reasoning is connected and limited by our language

@deirdresm @baldur @Korny

@om @deirdresm @baldur @Korny Possibly. If so, I think it would be correct about most of us most of the time, but there’s a nasty dilemma here (literally a two-way-forking concept in this case)-
Either our human intelligence is more than a pattern-matching imitation ritual, or it’s not.
(Disclosure: I think it is)
1/3

@om @deirdresm @baldur @Korny if our human intelligence really is just Lyrebirds imitating what they hear as a mating display, scaled up as far as our biology will permit, then the LLM approach probably will eventually bear fruit, but in doing so it will inescapably reveal to us what we are: We are nothing. There’s nothing new under the sun because there never was. Intelligence doesn’t exist.
I would call this outcome a Worst Case Scenario.
2/3

@om @deirdresm @baldur @Korny I would point to the things that humans have built over time and in the collective pursuit of science, and argue that this is a type of constructive discrimination that is fundamentally impossible to obtain from any statistical/probabilistic model.
I’m fully aware that as long as we can’t quantify what that other kind of reasoning is, this argument is tantamount to religion, no matter how obvious the evidence seems.
🤷🏼‍♂️
3/3

@thorne @om @baldur @Korny

So here's my take: for a random statistically normal *individual*, you may be right. The catch is, we all have different organizations of our neurons and memories and training, plus what memories do/don't get culled and why. When you add all that together, we have personality, a combo of nature and development.

But an LLM is *one* model, not *all* the models, and it's trained on what we say publicly, which is not what we actually think, but a censored version.

@thorne @baldur @Korny

As someone who worked on an AI project (determining which blood to unfreeze for Anti-D injections), it bugs the hell out of me.

en.wikipedia.org/wiki/Rho(D)_i

Because there is actually really useful AI that's cheapened by this planet-burning parlor trick.

en.wikipedia.orgRho(D) immune globulin - Wikipedia

@deirdresm @baldur
My mom loves watching old American tv shows, but it's impossible to find subtitles for them in Persian so she can watch them. So, I downloaded an LLM specifically for translation and wrote a python script to take English subtitles and use the LLM to translate them into Persian. And it worked like a charm.

@hirad @baldur Translation is a really different use than generative, though, particularly since there are large bodies of works that were human translated to use as training.

I’m not saying that there’s no good uses for LLMs, because Siri and Alexa have been useful. But that’s also a really different application than generative, because those are guided by human teams.

The point is that generative is a statistical trick, not real intelligence.

@deirdresm @baldur Except translation based on LLMs is also generative AI. What you described is exactly what datasets are.
That article is full of misconceptions that makes me wonder if its written by someone who actually knows how LLMs work.
It all sounds like someone who has only used chatgpt (or Google's gemini) briefly and asked some questions and then judged the entire world of LLMs based on that. Because its simply not true that LLMs can't reason. They can. Just not like humans.
They don't just spit out the data they've been fed. That's not how it works.

@hirad @baldur No it’s not because of the extensive corpus of translations that exist.

An analogy here would be that translation is like lane centering in self-driving. Humans put down the lane markers and the roads. And got the roads surveyed for mapping.

sae.org/blog/sae-j3016-update

Full-on generative that worked would be the equivalent of SAE 5, not 1.

www.sae.orgSAE Levels of Driving Automation™ Refined for Clarity and International AudienceAdvancing the conversation around mobility knowledge with news and trends to benefit humanity.

@deirdresm @baldur @unixmercenary
This kinda proved my point that the entire argument is based on the commercial Generative AI of Google and OpenAI. As if they're somehow the representatives of Gen AI and LLMs.
I almost never use either of those 2 for multiple reasons and I don't even consider them in my argument about generative AI.

@hirad @baldur @unixmercenary

Until you have one that's actually *really* generating accurate text, it's still just a planet-burning parlor trick.

Feel free to believe in parlor tricks all you like.

@deirdresm @baldur @unixmercenary
The amount of energy running Mixtral uses is incredibly smaller than playing video games. Should we call video games planet-burning too?
I thought we were having a civil discussion. I was highly mistaken. And I'm sorry about that.
Have a nice day.

@hirad @baldur @unixmercenary

I apologize for being snippy. I had a truly bad day yesterday and should not have taken my remaining gar about that out on you.

You're right that playing video games does consume a lot of power collectively, but entertainment does that.

And, frankly, cryptocurrency has less use (and future) than LLM-based generative AI, so there's that.

@deirdresm @baldur @unixmercenary I'm sorry to hear that. Hope you have a good weekend.
True, but it's about the outcome of the spent energy.
As much as I am in favor of LLMs in general, I'm against proprietary ones.
You should really look into the public and open source LLM community. The real beauty is there, where models are constantly evolving and becoming more efficient. How fine-tuning has allowed the community to build fully transparent LLMs for real tasks.

@hirad @baldur @unixmercenary

I think there's definitely valuable research to be done in how language works and how computers can represent language effectively and efficiently.

I used to work (in that we'd occasionally interact but were on very different teams) with Kim Silverman, who led a lot of the language-based modeling that Siri's based on.

youtube.com/watch?v=n5rwYnj2GZ

@deirdresm
@glynmoody

Very good description indeed!

What ChatGPT reminded me of was sort of a door-to-door insurance salesman -- or is it just what every other random dude on twitter used to ask it for? („AI, come up with a step-by-step plan to build a prospering business!“)

Or maybe a student trying to cross that minimum essay word limit by littering its “answer” with filler. I am always annoyed by that...

...
I mean, bot, I didn't ask you to summarize the context of my question!

In fact, this is probably the best use-case for #LLM I can think of: giving one an introduction into some field or topic. Except it won't even give me sources of the “knowledge” it so confidently provided and I will have to verify everything it answered somewhere else anyway.

@glynmoody @deirdresm

@deirdresm
honestly to me they feel more like a swiss army knife with 100 tools.

at first it seems amazing, because it can do almost anything.

then you realize that anything it can do, there's a specialized tool that can do it much better.

then you realize all the moving parts make them fragile and unreliable.

then you realize all the sharp edges could get someone badly hurt if they trust it too much.

then it doesn't look so amazing anymore.