A very pessimistic take by Marcus Hutchins on the current state of AI. The author touches upon a variety of topics which I have read independent of each other.
A logical problem I previously used to tests early LLMs was one called “The Wolf, The Goat, And The Cabbage”. The problem is simple. You’re walking with a wolf, a goat, and a cabbage. You come to a river which you need to cross. There is a small boat which only has enough space for you and one other item. If left unattended, the wolf will eat the goat, and the goat will eat the cabbage. How do you get all 3 safely across?
The correct answer is you take the goat across, leaving behind the wolf and the cabbage. You then return and fetch the cabbage, leaving the goat alone on the other side. Because the goat and cabbage cannot be left alone together, you take the goat back, leaving just the cabbage. Now, you can take the wolf across, leaving the wolf and the cabbage alone on the other side, finally returning to fetch the goat.
Any LLM could effortlessly answer this problem, because it has thousands of instances of the problem and the correct solution in its training data. But it was found that by simply swapping out one item but keeping the same constraints, the LLM would no longer be able to answer. Replacing the wolf with a lion, would result in the LLM going off the rails and just spewing a bunch of nonsense.
This made it clear the LLM was not actually thinking or reasoning through the problem, simply just regurgitating answers and explanations from its training data. Any human, knowing the answer to the original problem, could easily handle the wolf being swapped for a lion, or the cabbage for a lettuce. But LLMs, lacking reasoning, treated this as an entirely new problem.
Over time this issue was fixed. It could be that the LLM developers wrote algorithms to identify variants of the problem. It’s also possible that people posting different variants of the problem allowed the LLM to detect the core pattern, which all variants follow, allowing it to substitute words where needed.
This is when someone found you could just break the problem, and the LLM’s pattern matching along with it. Either by making it so none of the objects could be left unattended, or all of them could. In some variants there was no reason to cross the river, the boat doesn’t fit anyone, was actually a car, or has enough space to carry all the items at once. Humans, having actual logic and reasoning abilities could easily identify the broken versions of the problems and answer accordingly, but the LLMs would just output incoherent gibberish.
But of course, as more and more ways to disprove LLM reasoning were found, the developers just found ways to fix them. I strongly suspect these issues are not being fixed by any introduction of actual logic or reasoning, but by sub-models built to address specific problems. If this is the case, I’d argue we’re moving away from AGI and back towards building problem specific ML models, which is how “AI” has worked for decades.
Bonus: Check the wikipedia page of Marcus Hutchins.
naveegator.in 