A shocking story was promoted on the “front page” or main feed of Elon Musk’s X on Thursday:
“Iran Strikes Tel Aviv with Heavy Missiles,” read the headline.
This would certainly be a worrying world news development. Earlier that week, Israel had conducted an airstrike on Iran’s embassy in Syria, killing two generals as well as other officers. Retaliation from Iran seemed like a plausible occurrence.
But, there was one major problem: Iran did not attack Israel. The headline was fake.
Even more concerning, the fake headline was apparently generated by X’s own official AI chatbot, Grok, and then promoted by X’s trending news product, Explore, on the very first day of an updated version of the feature.
Yep. To add on, this is exactly what all the “AI haters” (myself included) are going on about when they say stuff like there isn’t any logic or understanding behind LLMs, or when they say they are stochastic parrots.
LLMs are incredibly good at generating text that works grammatically and reads like it was put together by someone knowledgable and confident, but they have no concept of “truth” or reality. They just have a ton of absurdly complicated technical data about how words/phrases/sentences are related to each other on a structural basis. It’s all just really complicated math about how text is put together. It’s absolutely amazing, but it is also literally and technologically impossible for that to spontaneously coelesce into reason/logic/sentience.
Turns out that if you get enough of that data together, it makes a very convincing appearance of logic and reason. But it’s only an appearance.
You can’t duct tape enough speak and spells together to rival the mass of the Sun and have it somehow just become something that outputs a believable human voice.
For an incredibly long time, ChatGPT would fail questions along the lines of “What’s heavier, a pound of feathers or three pounds of steel?” because it had seen the normal variation of the riddle with equal weights so many times. It has no concept of one being smaller than three. It just “knows” the pattern of the “correct” response.
It no longer fails that “trick”, but there’s significant evidence that OpenAI has set up custom handling for that riddle over top of the actual LLM, as it doesn’t take much work to find similar ways to trip it up by using slightly modified versions of classic riddles.
A lot of supporters will counter “Well I just ask it to tell the truth, or tell it that it’s wrong, and it corrects itself”, but I’ve seen plenty of anecdotes in the opposite direction, with ChatGPT insisting that it’s hallucination was fact. It doesn’t have any concept of true or false.
The shame of it is that despite this limitation LLMs have very real practical uses that, much like cryptocurrencies and NFTs did to blockchain, are being undercut by hucksters.
Tesla has done the same thing with autonomous driving too. They claimed to be something they’re not (fanboys don’t @ me about semantics) and made the REAL thing less trusted and take even longer to come to market.
Drives me crazy.
Yup, and I hate that.
I really would like to one day just take road trips everywhere without having to actually drive.
I love that example. Microsoft’s Copilot (based on GTP-4) immediately doesn’t disappoint:
It’s annoying that for many things, like basic programming tasks, it manages to generate reasonable output that is good enough to goat people into trusting it, yet hallucinates very obviously wrong stuff or follows completely insane approaches on anything off the beaten path. Every other day, I have to spend an hour to justify to a coworker why I wrote code this way when the AI has given him another “great” suggestion, like opening a hidden window with an UI control to query a database instead of going through our ORM.