Can’t help hallucinating: Baby, LLM was born this way
There’s no better way for me to start than with what Dario Amodei (CEO of Anthropic) said:
It really depends how you measure it, but I suspect that AI models probably hallucinate less than humans.
Let's be clear what “hallucination” means first...
Before we go further, let’s forget the clinical definition of hallucination for a moment.
In the world of large language models (LLMs), hallucination doesn’t mean hearing voices or seeing things that aren’t there. It basically means making things up — without purposely intending to deceive.
An honest mistake, so to speak.
Humans do this as well.
Like when a friend recently insisted that I did tell him I have three brothers. I have two. So I was absolutely certain I hadn't said otherwise. He was in complete disbelief. Totally flabbergasted. Because he was so sure I had said three.
And now, let’s get back to comparing humans’ and LLMs’ hallucinations
In case if you’re now thinking, “Meh, so both humans and LLMs are equally unreliable, tough luck...”
Uh, I’d like to stop you there.
While both humans and LLMs can make these kinds of mistakes, the reason WHY THEY MAKE THEM and WHAT THEY DO ABOUT IT makes a world of difference, as shown in the infographic.
Though, for a more whipping effect, I like how Gary Marcus (who’s been warning about hallucinations in neural networks since 2001) puts it:
It is a fallacy to think that because GenAI outputs sound human-like that their computations are human-like. LLMs mimic the rough structure of human language, but 8 years and roughly half a trillion dollars after their introduction, they continue to lack a grasp of a reality. And they have such a superficial understanding of their own output that they can’t begin to fact check it.
TL;DR: Human’s and LLM’s hallucination are not the same. And the difference makes all the... well, difference.
LLMs going to cause mass human job loss?
Before diving into this topic of LLM hallucinating, I couldn’t help but believe that LLMs were going to cause mass unemployment when it came to non-physical jobs. They appeared to be cheaper, more efficient, and had less drama than human workers.
And with images like this…
…I mean, even though those looked annoying, I couldn’t disagree with them. I felt anxious. The LLM is already here. Just a matter of time when most companies subscribe to them and retrench the humans. And with governments hardly ever talking about the problem, let alone about the solutions, I imagined a bleak future.
But after learning about and experiencing LLM hallucinations, I now highly doubt they’re going to cause mass unemployment. Many jobs involve understanding context, problem-solving, and adjusting to new situations: things LLMs mimic but can’t robustly do because their very architecture can’t fundamentally support it! Relying on LLMs to do critical jobs would be an insanely high risk thing to do.
Many people understandably still believe LLMs will replace them at work, though.
At a recent family gathering, my eldest brother, being the super extrovert that he is, was holding court, telling us about how his lawyer friend was looking for other ways to earn money — because “ChatGPT is going to replace her and all other lawyers soon.”
Okay, first of all, that lawyer friend is a partner of a law firm, so it’s a bit of a stretch that’s going to happen to her. But even for the lower-level lawyers in her firm, I’m doubtful they’d be retrenched in bulk over ChatGPT or any other LLMs. Read further on why…
Circling back to Dario
Really, I didn’t plan to write this post with the intention to continuously dunk on Dario, but omg, he’s like a gift that just keeps giving.
A lawyer representing Dario’s company, Anthropic, actually had to apologize for using fake citation in her draft filings. And where did that fake citation come from? Claude — the LLM developed by Anthropic. I’m not making this up (a.k.a. not hallucinating).
It’s not an isolated case, btw. Feel free to eat your popcorn while reading this article about how lawyers have been questioned/disciplined for using LLM-made-up-fiction “in at least seven cases over the last two years”.
I wonder if Dario truly/still believes that AI could eliminate HALF of all entry-level white-collar jobs within 5 years.
Most of them are unaware that this is about to happen. It sounds crazy, and people just don’t believe it… We, as the producers of this technology, have a duty and an obligation to be honest about what is coming.
I mean, if we’re talking about LLMs getting even the citations wrong, is it realistic to expect the big shots in law firms to be checking each citations? Call me jaded, but it’s doubtful. They’d still need their underlings to do those while they wine and dine with their clients.
And those are just legal stuff. What about matters of life and death like medical matters?
Anyhow, at the end of the day, I wonder, would Dario Amodei replace his human lawyers with his billion-dollar AI, Claude?
Think of any "AI" as little more than a very expensive autocomplete.
Reminds me of OpenAI claiming all sorts of things while having all sorts of job positions open. And that‘s not only engineers but also for example in accounting.