this post was submitted on 10 Aug 2024
550 points (96.0% liked)
Technology
59495 readers
3135 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I thought this was fake or a bad result or something, but totally just duplicated it. Wow.
If you read the block of text…. It doesn’t make sense either.
I expect if you follow the references you'd find one of them to be one of those "if Earth was a grain of sand" analogies.
People like laughing at AI but usually these silly-sounding answers accurately reflect the information the search returned.
It's in the quote that they scaled it.
The point is that the entire alleged value is the ability to parse the reading material and extract the key points, but because it doesn't resemble intelligence in any way, it isn't actually capable of meaningfully doing so.
Yes, not being able to distinguish between the real answer and a "banana for scale" analogy is a big problem that shows how fucking useless the technology is.
Yes but they supposedly scaled it to "one meter per meter". A "scale where the distance from the Sun to Earth is 150 million km" is the actual distance.
lol I did miss that, but it was enough to make it not a guess that its source was scaling for comparison.
My whole point was the same as your OP, though. A condom that's 95% effective isn't worth shit. You can't let a toy without reading comprehension do your reading for you.
But the thing is condoms ARE 98% effective, and yet people still use them every single day.
Nothing is perfect, humans, AI/LLMs, etc, no matter what, absolutely nothing is.
Regardless, anything I say about AI/LLMs that isn't that it's terrible and useless and nobody should/would ever use it is going to be met with criticism.
*Dangerous! Don't forget how dangerous it is — considering all tech bros and corps are acting as though LLM's are on the verge of real intelligence, instead of being a stochastic parrot that's essentially a mathematical magic trick.
"Now watch as I, the great mathemagician, make a statistical algorithm appear to hold general intelligence!"
Our "intelligence" agencies already kill innocent people based entirely on metadata — because they simply live or work around areas that known terrorists occupy — now imagine if an AI was calling the shots. The more LLM's are integrated into our day to day lives, the more people will trust them and disregard their own logic, and the more dangerous they become.
So by your own scenario, intelligence agencies are already getting stuff wrong and making bad decisions using existing methodologies.
Why do you assume that new methodologies that involve LLMs will be worse at that? Why could they not be better? Presumably they're going to be evaluating their results when deciding whether to make extensive use of them.
"Mathematical magic tricks" can turn out to be extremely useful. That phrase can be used to describe all manner of existing techniques that are undeniably foundational to civilization.
Calling "AI" (I know it's not true AI but rather an LLM) useless is very dismissive and just not true at all.
I wrote ArigatouAnimeTracker nearly entirely using ChatGPT including the description, nearly all 600 commits entirely from ChatGPT generated code. It is very far from useless and I feel much more comfortable with my dev job knowing I am willing to and able to leverage these newer technologies. They are only going to get better and what they are already capable of is impressive. If I didn't use an LLM it would have easily taken me 5x as long to write that project.
Regardless, anything I say about AI/LLMs that isn't that it's terrible and useless and nobody should/would ever use it is going to be met with criticism.
Except it is capable of meaningfully doing so, just not in 100% of every conceivable situation. And those rare flubs are the ones that get spread around and laughed at, such as this example.
There's a nice phrase I commonly use, "don't let the perfect be the enemy of the good." These AIs are good enough at this point that I find them to be very useful. Not perfect, of course, but they don't have to be as long as you're prepared for those occasions, like this one, where they give a wrong result. Like any tool you have some responsibility to know how to use it and what its capabilities are.
No, it isn't.
You're allowing a simple tool with literally zero reading comprehension to do your reading for you. It's not surprising your understanding of what the tech is is lacking.
Your comment is simply counterfactual. I do indeed find LLMs to be useful. Saying "no you don't!" Is frankly ridiculous.
I'm a computer programmer. Not directly experienced with LLMs themselves, but I understand the technology around them and have written program that make use of them. I know what their capabilities and limitations are.
Your claim that it's capable of doing what it claims isn't just false.
It's an egregious, massively harmful lie, and repeating it is always extremely malicious and inexcusable behavior.
I have genuinely found LLMs to be useful in many contexts. I use them to brainstorm and flesh out ideas for tabletop roleplaying adventures, to write song lyrics, to write Python scripts to do various random tasks. I've talked with them to learn about stuff, and verified that they were correct by checking their references. LLMs are demonstrably capable of these things. I demonstrated it.
Go ahead and refrain from using them yourself if you really don't want to, for whatever reason. But exclaiming "no it doesn't!" In the face of them actually doing the things you say they don't is just silly.
They absolutely cannot reliably summarize the result of searches, like this post is about, and OP in and of itself proves conclusively.
Any meaningful rate of failures at all makes them massively, catastrophically damaging to humanity as a whole. "Just don't use them" absolutely does not prevent their harm. Pushing them as competent is extremely fucking unacceptable behavior.
And this is all completely ignoring the obscene energy costs associated with making web searches complete and utter dogshit.
The problem is that it did summarize the result of this search, the results of this search included one of those "if the Earth was the size of a grain of sand, Alpha Centauri would be X kilometers away" analogies. It did exactly the thing you're saying it can't do.
Nothing is perfect. Does that make everything a massive catastrophic threat to humanity? How have we managed to survive for this long?
You're ridiculously overblowing this. It's a "ha ha, looks like AI made a whoopsie because I didn't understand that I actually asked it to do" situation. It's not Skynet coming to convince us to eat cyanide.
Of course it's ignoring that. It's not real.
You realize that energy costs money? If each web search cost an "obscene" amount, how is Microsoft managing to pay for it all? Why are they paying for it? Do you think they'll continue paying for it indefinitely? It'd be a completely self-solving problem.
Summaries distinguish substance from nonsense. It cannot be described as a summary of a piece of content if it does not accurately portray the substance of that content.
LLMs aren't imperfect. They're dumpster fire misinformation machines with no redeeming qualities. Of course it's not Skynet. Skynet was intelligent. This isn't within 100 orders of magnitude of intelligence.
Companies burn obscene amounts of money on moonshots all the time, even ones that have no possibility of success. Willingness to lose billions burning energy to degrade every single search made is not an indication that it's not a nightmare for the environment (again, for literally no purpose because every single search with an LLM is worse than without it).
No, a summary is just a condensed version of some larger work. If the larger work contains bullshit then so can the summary, that doesn't stop it from being a summary. As you say, a summary accurately portrays the substance of that content. In this case there was content that said Alpha Centauri was 13 km from Earth, so the summary said that too.
This is really not complicated.
If you think it has no possibility of success, sit back and relax as AI goes away.
Yep. This is exactly it, and this is what people don't seem to understand. AI is not going away, because it is actually useful, it has actual uses and people are actively using it. It's not entirely fluff based pointless technology like blockchain etc, it is actually useful and real-world people use AI/LLMs.
This is just not true.
Just because you don't like LLMs doesn't mean that they have no purpose. If they were really entirely useless and served no purpose and never did anything, they would not be the talk of the town and OpenAI would not be a multi-billion dollar company. If they were useless, nobody would use them, but people absolutely do use them.
I literally use ChatGPT daily to automate writing code for me and it honestly does a good job. I literally used it to write an entire Laravel project called ArigatouAnimeTracker, over 600 commits including documentation all written using ChatGPT, and tbh my project is awesome. It easily would have taken me 5x as long to write it without ChatGPT and tbh it might not have ended up existing without ChatGPT because of how long it would have taken to write without LLMs doing the heavy lifting.
Sure, you have to verify the output, but you know what? That's going to be the case for any code that is written regardless, code review is essential and completely normal and existed long before LLMs did. That doesn't mean that LLMs don't have a purpose, or that nobody actually uses them. People do use them, it's a multi-billion dollar industry for a reason and people are going to continue to use them, even if you say they have no redeeming qualities. There are definitely ethical concerns about LLMs, but to say they have no redeeming qualities is just not correct.
Regardless, anything I say about AI/LLMs that isn’t that it’s terrible and useless and nobody should/would ever use it is going to be met with criticism.
I disagree with that view on them, and I think the fact that they fail is actually a good thing in terms of preventing damage to humanity.
If they are able to perfectly do all kinds of jobs without ever making mistakes and being better at it than any humans, that would be infinitely more damaging than having them make mistakes meaning their use is limited to having to have their output carefully reviewed yet still used when it is helpful and appropriate.
Regardless, anything I say about AI/LLMs that isn't that it's terrible and useless and nobody should/would ever use it is going to be met with criticism.
AIs are definitely not "good enough" to give correct answers to science questions. I've seen lots of other incorrect answers before seeing this one. While it was easy to spot that this answer is incorrect, how many incorrect answers are not obvious?
Then go ahead and put "science questions" into one of the areas that you don't use LLMs for. That doesn't make them useless in general.
I would say that a more precise and specific restriction would be "they're not good at questions involving numbers." That's narrower than "science questions" in general, they're still pretty good at dealing with the concepts involved. LLMs aren't good at math so don't use them for math.
AI doesn't seem to be good at anything in which there is a right answer and a wrong answer. It works best for things where there are no right/wrong answers.
I just tried and got "about 40,000 billion kilometers". Also the references are completely different from the ones in the post, so I guess it was a ranking issue
AI is just too unpredictable, hard to know what's accurate and you end up doing the work yourself anyways
the loaded die at the end that chooses one of the llm's answers happened to land on a good word