Incredible. Apple Intelligence summarized BBC news to claim Luigi had shot himself. This not only had not happened but was not something BBC reported.
AI news summaries are a terrible idea because "just making up shit" is basically an unsolvable problem in LLMs
https://bbc.com/news/articles/cd0elzk24dno
@vegetablegremlin I dunno. My first job was writing periodical abstracts, and it sort of seems like if text-based GenAI has *any* valid purpose (aside from maybe filling out health forms), it would be summarizing articles
If it can’t do that well … 😬
@vegetablegremlin @peterbutler
That is not the purpose of LLMs, though.
They’re glorified predictive text, that uses a token system to identify and replicate patterns in naturalistic language.
They shouldn’t be used for any information retrieval or summarisation task because they don’t *understand* the information — ideas are just clusters of word patterns.
@vegetablegremlin @peterbutler
This article is actually an excellent case study.
If you think of an idea as a shape — a square — represented by the way the words — the triangles — cluster around it, without a capacity for meaning-making, one square made out of triangles looks much like another.
LLMs find patterns, not content.
They follow pattern rules of naturalistic language, they don’t recreate meaning.
All the triangles are there, and they make a square.
@peterbutler @vegetablegremlin
LLMs have valid purposes which are often lost on people who write professionally — because they’re assistive aids for tasks that come easily for y’all personally. You don’t, personally, need them.
What they’re *for* is arranging language in conventional and thus easily legible patterns.
For people who struggle with written expression for any number of reasons, that’s valuable.
But it’s a limited use-case.
@RaphJ
A LLM very much *does not* “understand things and built an inner representation of the world”.
They’re sophisticated predictive text bots, which replicate patterns in language.
That can *feel* like “understanding and building an inner representation” to humans, because it’s what we do with language, and because one of our worst cognitive biases is a propensity for projection.
Marketing feeds that bias with language like “intelligence” and “hallucination”, but it’s misleading.
@justanotheramy
It does way beyond than this.
LLMs are just trained to predict the next word indeed.
But to achieve this task well, they compress human knowledge into an organized structure : they make "sense" of the world.
Read this paper in which GPT is feed with chess parties and learn the rules himself. It's than able to play random games it never seen before.
https://adamkarvonen.github.io/machine_learning/2024/01/03/chess-world-models.html
@RaphJ
You are being wildly mislead by language like “Internal World Model” and “brain”.
Turing’s ghost would like a word.
The algorithm that detects and replicates patterns detects and replicated patterns.
Domain-specific models and RAG models exist to ensure the specificity and relevance of those patterns to context, but that’s not understanding.
@RaphJ
You’re talking about something that doesn’t even have a logic module. Come on.
This is sophomoric stuff. Go read Turing, think about the number of Rs in strawberry, and stop wasting my time.
This is an evolutionary moment. Not for AI - for us. Humanity.
We're being presented with something that looks an awful lot like something that's intelligent. But - it's not. The lesson we need to learn is that you need to not trust something just because it sounds like a person trying to help.
It's not just AI. It's phishing. It's 911 scams. It's the latest hot cyber coin, the hot insider stock tip, the "your computer has a virus give me remote access to fix it". It's a new attack vector the human race has very little experience dealing with.
Remember how we win in the movie "ID4"? Same deal - the aliens never had NFTs to teach them a valuable lesson about trust.
Some of us are immunized - we need to try to help inoculate others. Some of us will never learn - we need to try to make sure when they crater they don't take out bystanders.
Merry Christmas, kids.
PS. AI isn't entirely useless. It's just not trustworthy, and probably will never be, at least using current technology. Once society groks in fullness what it really is, there are places it dan be very helpful. Just not in "summarize this thing where accuracy matters"
Here it starts… AI will be a problem, how to validate its actions?
Where is truth in the continuous dreams and nightmares it can produce?
How long till people say they attended UMSU? The University of Making Shit Up? Or are we already all enrolled and I, as usual, am far behind reality?
@vegetablegremlin Damn!
BBC complaining to Apple AI about making shit up instead of leaving it to their highly experienced reporters (ahem!) to make shit up.
#PotKettleBlack.
Vegetable Gremlin ⍼👻
Peter Butler
Just Another Amy
Raph J 🇺🇦🇪🇺🇵🇸 #HelloQuitteX
Josh the Blockaded
Григорий Клюшников
pjɛʁ-səlim 
SueDiOh
Tom Bortels
CleoMatra
Ben Rosengart
abadidea
Metafrastis
Sören
JF 
Patrick H. Lauke
Gokul Das
xs4me2
David Cantrell 🏏
B Kahn
Andrew_TighF 🏴
Julian L