With just about every piece of technology up until now, when it's not working correctly that is apparent to the user in some way.
The reason LLMs and today's versions of "AI" scare me is that users think they are working correctly even when they're putting out nonsense.
The reason LLMs and today's versions of "AI" scare me is that users think they are working correctly even when they're putting out nonsense.
Comments
Hot holding is 135F. Poultry is suppose to be cooked till it hits 160F internal temperature.
Fuckin... bro... you're talking to a trekkie... We were already talking...
I've had many instances where I
https://www.perplexity.ai/search/fix-a-short-board-c64-0cbACx2mSPaXwzPV8AP6tg#0
I checked the sources that it cites (chatgpt doesn't cite sources), and it looks like a legit way to diagnose and fix one.
It means the creators don't have a proper understanding how it works, and that there is no real 'intelligence' in the system.
The VC people are finally pulling the plug on it in all likelihood.
When Google returns results there's really no way to know if it returned the best or most accurate results. It's almost certainly wrong a lot of the time and you never know it.
But yeah, it's an important difference that Google linked out to web sites and you could evaluate how much you trusted those web sites.
Transformer models are a black box, they can't be debugged in the traditional sense. There's no deterministic algorithm. When something goes wrong, you can't "find the error."
Or, you can make a new black box and hope it works better.
That includes many highly intelligent people who should know better, but probably got too invested into it and somehow made it their cult-like worldview.
It’s feels like talking to flat-earthers.
👆 THIS!
If you need any more help with your social media posts just write me. I will help. 🙂
I've had to basically yell at people at work to not just read the summary
Best example is YouTube - a lot of new content is simply dull.
"When an entire reply consists of tokens selected from high-probability predictions, it indicates that:
• The model has strong statistical patterns backing each choice…" etc…
You are conflating linguistic correlation and truth.
I believe most users who are asking ChatGPT etc. for the answers to questions think there is some logic and reasoning behind the answers it gives. This is bolstered by how frequently they are mostly correct.
LLMs appear to "work" all the time, and the failure mode is only experienced later as real-life consequences.
Saw one example where they used tokens like "*SYSTEM" to separate system prompts from user input. Then removed all * from user input. (I don't recall what the actual character was.) Seems perfectly safe, right?
y'know. like real actual human experts do.
i wish we as a society fostered that curiosity more.
They produce a simulacrum.
But yes, LLMs and people's disturbing faith in their output are on a whole other level...
Which of course falls apart when it's being labelled as "intelligence".
The rumors that Siri will adopt an LLM next year do not instill confidence in me. Robotic yes-men, personalized for you.
An LLM Siri will always give you a response with confidence, even if it completely failed you.
It's like when the news reports on your industry and you're like, "they don't know the half of it." Spoiler: they (and LLMs) only scratch the surface of *everything.*
Dunning-Kruger makes stupidity invisible to the wearer, but can be detected through adverse consequences.
This comes out as an inconsistent, unsignaled failure, for example.
Just understanding, for instance, that the clicking sounds my stove makes are relays and then learning what the pattern of clicks it makes sounds like allows me to instantly have a troubleshooting first step.
I once diagnosed a faulty USB drive by the change in noise leaking from its data lines to the headphone jack nearby.
It really, really helped with "but how do we _know_ something to be true?" type of topics.
LLM
LL Cool J
Only one seems to have been a net benefit?
No one knows what it’s doing or how it’s working between garbage going in and garbage coming out.
As you said, they always ‘work’ but what a lot of people are expecting when they ask it questions with a factual answer is the correct response and they just assume it’s correct.
Seems they were right.
Like the same way the clicking relay sounds of last century's automotive turn signals are reproduced synthetically now so today's drivers know the turn signal is on.
(Don't worry, it's all low voltage/amperage DC stuff)
The idea of "programs being wrong" isn't new, it's just been perfected by AI like never before.
People don't need to do these kings of things anymore, so when it doesn't work, they dunno!
ChatGPT can’t even quote Wikipedia accurately and is incapable of providing an actual source.
As long as LLMs sound like they know what they are saying people will belive it, notice that LLMs don't just give you an answer, they make it while trying to sound as convincing as possible
Being a really good predictor based on probabilities.
Predictor not necessarily factually correct.
IMO the problem arises when the model is in fact imperfect and has to resolve to imperfect data in order to sound convincing.
The topic of hallucinations is still researched and debated but that's one of my (many) reasons for it that I've heard of.
I’ve supported ML in a large-scale environment (as devops), and it’s shocking how even the people who write and build these systems scarcely understand the specifics of how the output is arrived at.
It's like trying to use a toaster to play DVDs and saying the failure state is that your DVDs are melting.
WE are supposed to fact check the machines, not the other way around. Machines don’t know lies from truth.
don't you know they've already discovered tens of thousands of definitely viable new material configurations? material science is le saved!
/s
Don’t know enough on how AI works to know why something like that hasn’t been implemented.
Might we find that since an LLM is some sort of average of information taken together, that it's less wrong than many of other sources?
We’ll need to educate people to check the output, just like people were taught to check sources when using google search