I still find myself amazed and underwhelmed at the same time with AI, at least with the AI I have been exposed to via deepseek and chatgpt. It still does not demonstrate intelligence, and I am not sure how it will reach basic levels under its current design given the considerable resources already applied. The AI seems like a logistic curve of diminishing returns. In vital ways I can't see any progress over the last couple of years.
It was for example terrible at anagrams two years ago, but now is much better though still occasionally wrong - but if it was bad at something an 80s computer could undertake infallibly at such an advanced stage of its development, then how could it be expected to advance rapidly beyond such basic tasks to more complex problems? It just felt bad AI design, no matter how amazing it looked. Human beings make progress through reflecting on their work, not simply doing the work, stepping outside of it and subsequently reappraising and correcting - AI doesn't do this unless it is instructed to do so. But this is just another directed task rather than a reflection.
The other week I was playing with AI trying to replicate playing a card game - it was both impressive and deeply flawed. At one point it had produced two Ace of Spades then when pointed out went into its lying mode, saying it was removing duplicates - which makes sense from a real world explanation, but not of course for AI.
I noticed a lot of non-randomisation - repetition and failure to create certain patterns. So for exampe in 6 cards there would never be pairs, ever.
This was strange and bizarrely both chatgpt and deepseek made the same error. Inititally I wondered if the two platforms were a little more related then we were led to believe, but then considered they might be falling foul of the representative heuristic. So I undertook another test and sure enough it was.
Instruct with chatgpt or deepseek the following:
"produce 50 random 7 digit numbers"
A human will notice something quite quickly, there are no repetitions, each 7 digit number contains 7 distinct digits. This as is the AI's way, trying to prioritise satisfying the user rather than striving to produce an accurate answer. It has been well documented that humans have an idea of what randomness looks like, and what it doesn't - and so it has chosen numbers that look random over ones that don't appear random and at the end of deepseek in says the following:
"These numbers are randomly generated and can be used for simulations, testing, or any other non-sensitive purposes. Let me know if you'd like them in a different format!"
What it has done is anything but random it has filtered numbers satisifying a non-random criteria.
The other week I was able to instruct it to "generate" 50 7 digit random numbers and it produced numbers with repetition ofter outsourcing the task to a program - but that just failed as of writing.
Of course if you point out the lack of repetition in the randomness it will respond with a rtypical "My bad" and acknowledge the error - but it is incapable despite its incalculable resources of challenging itself - taking a basic definition of randomness and checking to see if it has fulfilled this criteria.
I find it thus difficult to trust it on matters I am unable to interrogate - unlike the randomness of 7 digit numbers. And since it doesn't self-interrogate I struggle to see how it is going to leap to intelligence on this design. When will it figure out it isn't producing random numbers? When we double, quadruple the number of chips? When looked upon using certain metrics, it is a very slow if not impossible progess towards intelligence using these models - there is a leap to intelligence needed that these models haven't made, not look likely to.
It is though going to dazzle us by making more and more pretty patterns.