| >Some experts, and lots of people who have not themselves built LLMs, such as Elon Musk, Yuval Noah Harari, Steve Wozniak. Never denied this. I did not intentionally hide this. I will say your statement did not mention that one of the signers is the premier expert. The person who reignited interest in neural networks, Geoffrey Hinton. It's not "some" experts. It's many many preeminent experts. It's more accurate to say, not "all" experts. That is the more fair characterization. > People drive cars every day and can't explain the first thing about the technology. I think you grossly overestimate how deeply people understand technology they use every day. Just ask a cell phone user how cellular communications works. Nobody fully understands these networks. Not even the people who build them. The "experts" admit this. What everyone understands is that these networks hallucinate. The hallucinations aren't deep things to understand, it's on the level of driving cars. You misunderstand how trivial it is to use this technology and figure out the limitations just like you figure out how to drive a car. We are all talking about user level phenomenons. You don't even have to describe what a neuron is or what backpropagation is to understand "hallucinations". Hallucinations aren't technical level concepts. >ChatGPT is already plugged into the Bing search engine, and into Windows, MS Office, and other products. New integrations and plugins get announced every day. Yeah pretty much everyone I know has used LLMs in some form or manner. And none of them were fooled. I'm sure the same goes for the people you know. Yeah you characterized how popular LLMs are. What you failed to characterize is the proportion of people who are using the tool and getting fooled by it. It is fact that the tool is popular and getting more popular but this says nothing about the topic at hand. How many people are getting fooled by chatGPT? Your just speculating the outcome based on how popular the tool is. Speculations prove nothing. I'm still waiting for that one guy to come to this thread and tell everyone chatGPT is sentient. If one guy claims this then you're right I completely misunderstood how dumb people can get. I mean don't even focus on this thread under my post. Let's focus on the entire thread under the main article. Is there even one person who completely thinks the AI is sentient? Find one guy on this entire thread who is honestly fooled by chatGPT and I'll concede to your argument. If you can't find even one guy... Then if you're rational you'll see that my argument is true: nearly no one is fooled by chatGPT. This is literally a data driven conclusion. Let's use some amateur science to get to the bottom of this. |
That's repeated a lot but it's not entirely accurate. No one can explain how an LLM gives the answers it does (not even the LLM). LLMs have a vast search space of tokens and use probabilities to make their responses non-deterministic. But the people who build and train LLMs do know how they work -- obviously since quite a few people know how to make one. By analogy, if I give a pile of Legos to a six-year-old I don't know what they will make, though I do know the constraints and limits (imposed by how Legos work and what was in the pile). It's not correct to say "I don't understand how Legos work" when I really mean "I can't predict what a six-year-old will make from a pile of Legos."
> You misunderstand how trivial it is to use this technology and figure out the limitations ... Hallucinations aren't technical level concepts.
I get that. But when I have discussed ChatGPT hallucinations, with examples from my own chats, I'm surprised when people don't even recognize the hallucinations until I point them out. Anecdotally people seem to defend the "AI" by accusing me of misleading it, or giving unclear information. They have anthropomorphized and then want to impose human notions of fairness, give the "AI" the benefit of the doubt even when they know that a person would have not made the mistake, or answered "I don't know" rather than confidently make stuff up like ChatGPT will. I think people don't believe computers can lie or make a mistake -- they imagine an intelligence like Mr. Spock at the other end, not a stochastic parrot.
I got ChatGPT to tell me -- in its confident and authoritative tone -- that no even number is also evenly divisible by 3. When I gave it contradictory examples -- 6, 12, 24 -- it then apologized but maintained that no even number was divisible by both 3 and 5 (um, 30, 60, 90...). I was just trying to get it to solve FizzBuzz with some variations. I could feed my younger children that same misinformation and they wouldn't question it. I could tell my parents that their Alexa listens to everything they say and records it forever on a big disk on a satellite in orbit, and they would believe me. Elon Musk can tell the world Teslas can drive from SF to New York without human intervention and get a whole TEDTalk audience and media "experts" to believe him. P.T. Barnum had some quips about that tendency.
> Find one guy on this entire thread who is honestly fooled by chatGPT and I'll concede to your argument. If you can't find even one guy... Then if you're rational you'll see that my argument is true: nearly no one is fooled by chatGPT.
That's not proof of anything. Maybe no one will chime in on this thread but you can easily find posts and comments on HN with people claiming LLMs are sentient. A Google researcher said that publicly (he got fired), and much discussion took place here. If I was more interested in the topic I could poll people, but I'm pretty sure I would find that quite a lot of people think current "AI" (LLMs like ChatGPT) are sentient, or they will be in the next couple of years. And then I could ask them what "sentient" means and never stop face-palming.