Strypey
I find it amazing how unwilling people are to accept that current "#AI" tech (including the "LLM" tech that I call #MOLE Training) is not intelligent, by any meaningful definition of the word. The nonsense arguments they use to wriggle out of this conclusion are nothing if not creative.
Most common is the consensus reality wriggle; more people talk as if they think Trained MOLEs are intelligent, therefore they are. So if most people think perpetual energy is possible, then it is? Nope.
(1/?)
Strypey
@worik
This video by Ryan George, posted today, pretty much sums up the unimpressive performance of current "AI" when it tries to report knowable facts;
Worik
>> That was before transformers became known
> Which ones?
Transformers are the scientific breakthrough that made these LLMs (and other useful neural networks) possible.
I have not studied them. The main paper, which I have been too busy to read, is called "Attention Is All You Need"
Strypey
@worik
> Attention Is All You Need
This one?
https://doi.org/10.48550/arXiv.1706.03762
I've read the abstract. It describes a better way to do MOLE Training, and has little or nothing to do with the discussion we're having. Which is about whether MOLE Training creates "intelligence", and how much we have to shave off an everyday understanding of what "intelligence" means to answer "yes" to that question.
A Trained MOLE is about as intelligent as a parrot that can say "Polly want a cracker".
Worik
That was before transformers became known
It is out of date
My point is, and you refuse to listen, that semantic hair splitting aside I now have a machine I can ask questions of, in natural English, it gets my meaning o r I can correct it, and get the information I want
This is a huge deal. It will change history. It happened I 2022
The scammers are hard at work using AI to bolster scams, but that does not detract from the fact that we are witnessing an industrial revolution
Strypey
@worik
> That was before transformers became known
Which ones?
> aside I now have a machine I can ask questions of, in natural English, it gets my meaning o r I can correct it, and get the information I want
This is not new. Search engines were doing this long before the ChatGPT hype train got started. What I'm disputing is whether elaborate autocomplete is a form of *intelligence*. You can heavily dilute the meaning of "intelligence" to include it, but I don't see any value in doing that.
Strypey
@worik
See also this 2020 paper by Ragnar Fjelland, published on Nature.com, entitled 'Why general artificial intelligence will not be realized';
https://doi.org/10.1057/s41599-020-0494-4
No significant changes have happened in any area of AI tech since then. Just more compute added to existing MOLE Training methods.
Worik
@strypey that is bad article
We have very little idea of how human intelligence (and what about my dog?) Works, so the entire premise is false.
For seventy years there has been a standard, amd a test, established by Turing (which is why his biographical film is The Imitation Game)
Now machines are closing in on passing that test, you are moving the goal posts.
Unfair
Whatever, LLMs (particularly transformers) are the greatest computer advance I have ever seen
Strypey
@worik You clearly really *want* machines that think to exist. I'm sorry the inherent limitations of the current language autocomplete algorithms disappoint you so much. But they are what they are.
I asked pi.ai, "are you intelligent?" It's output was;
"I'm not intelligent in the same way that humans are - I don't have consciousness or emotions, and I can't learn in the same way that humans do. I'm just a computer program designed to simulate intelligence!"
Worik
What is intelligence?
Is my dog intelligent? A child? A baby? A tree?
I can hold a conversation with an LLM, I can correct it, and it corrects me
Close enough?
Strypey
@worik
> I can hold a conversation with an LLM, I can correct it, and it corrects me
You are anthromorphising. Like saying a windscreen wiper is clearing the windscreen *for you*.
The algorithm is generating statistical guesses based on your inputs. Like a spell checker autocorrect. There is no conversation. It's not "correcting" you. These terms imply decisions it doesn't make. See;
https://betterwithout.ai/mind-like-AI
> Close enough?
Not even on the same planet, let alone in the same ballpark.
Strypey
Another wriggle I strike frequently is that although the current MOLE Training tech can't produce intelligence yet, future versions will. Which is equivalent to saying that although current fusion energy tech can't produce cold fusion, future versions obviously can. Nope.
The existence of one body of tech that didn't used to exist, with hard limitations, doesn't prove that tech without those limitations will automatically come into existence in the future.
So much magical thinking.
(2/2)
Strypey
Was that you @lain? : P
Strypey
"[_Iain] automated the creation of thousands of typo-squatted packages (many targeting crypto libraries) and even used ChatGPT to generate realistic-sounding variants of real package names at scale. He shared video tutorials walking others through the process, from publishing the packages to executing payloads on infected machines via a GUI."
#FerossAboukhadijeh, CEO, Socket, 2025
https://www.theregister.com/2025/04/12/ai_code_suggestions_sabotage_supply_chain/
(2/2)
Strypey
"What a world we live in: AI hallucinated packages are validated and rubber-stamped by another AI that is too eager to be helpful."
#FerossAboukhadijeh, CEO, Socket, 2025
https://www.theregister.com/2025/04/12/ai_code_suggestions_sabotage_supply_chain/
(1/2)
Strypey
"All that's required is to create a malicious software package under a hallucinated package name and then upload the bad package to a package registry or index like PyPI or npm for distribution. Thereafter, when an AI code assistant re-hallucinates the co-opted name, the code will run the malware."
#ThomasClaburn, 2025
https://www.theregister.com/2025/04/12/ai_code_suggestions_sabotage_supply_chain/
(2/2)
Worik
@strypey you'll like this: https://www.theregister.com/2025/04/12/ai_code_suggestions_sabotage_supply_chain/
Strypey
"... AI code assistants invent package names. In a recent study, researchers found that about 5.2 percent of package suggestions from commercial models didn't exist, compared to 21.7 percent from open source or openly available models.
Running that code should result in an error when importing a non-existent package. But miscreants have realized that they can hijack the hallucination for their own benefit."
#ThomasClaburn, 2025
https://www.theregister.com/2025/04/12/ai_code_suggestions_sabotage_supply_chain/
(1/2)
Worik
@strypey To summerise:
* You do not accept the Turing Test as a benchmark for AI. Ok, I disagree but it is semantics
* LLMs are very, unbelievably (to four years ago me) useful.
I think you are do not agree.
I point to my lived reality, and an expert in the field of computing, that they are amazing technology. A "steam engine" moment.
Where we are in total harmony is the scammers, grifters, and greed merchants are out in force
AI cannot do your job.
Your boss bought AI.
You're fired!
Strypey
@worik
That's a pretty good summary. Thanks for the discussion : )
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.