Max LeibmanOct 1
You don’t need to understand the math behind large language models in order to get that they are not minds.
Show threadMax LeibmanOct 1

The next time you meet someone who believes ChatGPT knows anything, experiences perception, has motivations, or forms opinions, show them this video.

Link to the original:

https://www.tiktok.com/t/ZTM6ytyLR/

TikTok - Make Your Day

Show threadWulfy—Speaker to the machinesOct 1

@maxleibman

AI ate whole of the internet.
Of course it "knows things"

Just like with web search, especially in its early days, you have to know how to get at that information.

Show threadMax LeibmanOct 2
@n_dimension Exactly right that you put “knows things” in quotes, though. It knows things the same way a dictionary knows things. I didn’t say there’s no information in there, but that isn’t the same as knowing.
Show threaderuweroOct 2
@maxleibman @n_dimension with the difference that the things in a dictionary are usually correct and not some random made up bullshit for some of the words
Show threadveetee
@eruwero @maxleibman @n_dimension also, you can use a dictionary to find out if a word exists. I've never had an LLM tell me something doesn't exist - it just makes up something plausible sounding
Oct 2 at 2:09amWeb
Show threadWulfy—Speaker to the machinesOct 2

@vt52 @eruwero @maxleibman

You're not asking the right questions.
My AI says "No" all the time...
... This is from this morning.
#imsosmart

But more seriously, I maintain using AI effectively is A LEARNED SKILL.

Show threadveeteeOct 2
@n_dimension @eruwero @maxleibman won't disagree that it's an acquired skill. perhaps I'm just working in problem domains that lack adequate training
Show threaderuweroOct 2
@n_dimension @vt52 @maxleibman that's a great marketing strategy. If it doesn't work, you're using it wrong. They claim this stuff has "PhD level intelligence". If you need to learn how to talk to a PhD so they can answer basic questions for you, and they might still just make stuff up, then that PhD is stupid and pretty much useless.
Show threadWulfy—Speaker to the machinesOct 2

@eruwero @vt52 @maxleibman

Every time you launch an #LLM session, it fires up "OS prompts"
Like "Don't help make meth"...
...what's a few more prompts to make it more effective tool?

It's a little bit like firing up a Linux console and saying "Whoa, I now have to issue commands to make it work!?"

I note that quite a few AI antagonists seem to have a very optimistic interpretation of the power of these things whilst eschewing them at the same time.

Show threaderuweroOct 16

@n_dimension that this is how the make LLMs "safe" is ridiculous btw.

Not really, commands are supposed to do what you tell them to, prompting is more like guessing which tokens will lead to less wrong output, but if you don't know much about the subject you'll never know if it's even remotely correct.

I'm not optimistic about them but I also don't claim that the technology is completely useless. It's stupid, useless, wasteful and dangerous how (most) people use it IMO

Show threadWulfy—Speaker to the machinesOct 16

@eruwero

One can not ever make #llm safe.
It's a fallacy.

It's mathematically impossible

Show threadsaxnot23h ago

@vt52 @eruwero @maxleibman @n_dimension

> it just makes up something plausible sounding

exactly this.
It would be much better technology if it would say "I don't know" or "I'm unsure about" when appropiate.

I deeply respect when someone says they don't know. It feels like everything else they said is thay much more trustworthy because that stuff passed the "do I really know it" filter.
Especially impressive when its a person who usually has answers to everything.

Show threadsaxnot23h ago

@vt52 @eruwero @maxleibman @n_dimension what I describe sound so much different from an LLM.
The guy I am thinking of just knows a lot and tells me where to read it and whips up small demos on his laptop to show what he means. He also has a home lab for trying out network configurations

Hearing "I don't know" from him really makes one trust that everything else he says has a proper foundation.

Show threadWulfy—Speaker to the machines17h ago

@saxnot @vt52 @eruwero @maxleibman

Here is three examples of "Don't know" from my chats.

👉"Simatic nonlinear resonance modes" — I said "I'm not immediately certain what you're referring to," "I'm not immediately familiar with what you're referring to as 'Qualia Institute,'" and "I'm not familiar with that specific work" (about Emilsson's 5-MeO-DMT work).

👉"Information Requested on ANZPAA's Diploma of Police Intelligence Practice" — I said "Unfortunately I don't have any specific information about the POL50119 - Diploma of Police Intelligence Practice."

👉When you asked about searching Claude chats, I said "I don't know the full details of how chat history and search functionality work in Claude's interface." Ironic, given I'm now using exactly that feature to find these chats.

"The Emilsson one is particularly notable — that's QRI/Andrés Emilsson's work, which I now know is a long-standing interest of yours. Earlier versions of me clearly didn't have the context (or the memory system) to connect those dots."

If you're not using tech that literally improves from week to week,
You just might reinforce your outdated biases.

#aisycophancy #promptengineering or just, sigh #prompt