In the abstract sense, I do agree.
But like any matrix based problem (no matter how large), they use SVD (Singular Value Decomposition) to arrive at the optimised result (at each neuron level).
I am just not convinced this process is optimal (you can do all sorts of adjustments for over-fitting but that just makes the whole thing even more problematic).
Part of my statement that I was trying to imply is that it takes several years for the human brain to get things calibrated and sorted out. Add the couple of decades (more or less) for them to figure out all of the varied social nuances . . . and that’s with additional outside influence to help with the “training” of the human brain-model.
Why would we expect for AI to do it in a much shorter time frame?
Remember that machine learning models were developed by engineers trying to program pattern recognition. They were programmed for action instead of inference.
I think machine learning models tend to be good when you have lots of data, and are fitting for a signal that is obvious to humans. For example, object recognition using computer vision, or sentence completion (which ChatGPT is really an example of.)
I am more skeptical about its use when fitting “latent” variables that cannot be directly observed, such as propensity for loss or credit default, say. That depends on the machine learning model though.
In my mind, the larger problem if we want “artificial intelligence” is that there does not appear to be a universal “optimal.” A predictive or learning model can be optimal for a particular environment, but as soon as that environment changes, it is not optimal anymore.
This is pure speculation, but I wonder whether there are physical limits to intelligence that we don’t understand yet. Perhaps we are close to “optimal” in certain senses.
As an example, I am thinking of the connection between entropy from physics and entropy from information theory. This implies certain thermodynamic limits on computation. I’m also thinking of the variance-bias trade off, which implies that the flexibility of an effective model must depend in an important way on the noise of the environment. Perhaps there are other limits we do not know about.
I would also wonder what GPT-4 had for study materials? Study guides? Past exams? The AO? Or was it working with just texts?
But mostly just terrible at math. I recall previous LLMs could be improved a lot by explitly telling them to think about it strp by step, and try different solutions, and so forth. I wonder if the same will help gpt-4.
Oh, another advantage that human actuaries get is scratch paper and a calculator.
Let me know when chatgpt can start asking the right questions, until them I’m not worried about my job.
whatever Google told it
==snort== awesome typo, there 
All we need to do is make all our study materials ungooglable!
Job Security by Job Obscurity!
Quick, shut down the Actuarial Outpost!
Yeah. I think until it masters high school math you’re safe.
I wonder if it has the actuarial outpost saved in its training data. it uses an older snapshot of the internet as i understand it.
And yes, this is my thinking. Probably outdated, but iirc Piaget said the average human brain can’t handle Abstract reasoning until teenagerhood.
In that way it could be like an absurdly well educated 11 year old.
So the best “Are You Smarter Than a 5th Grader?” player ever
Yeah I don’t know. What about the fact that it passed the Bar exam?
Does that qualify it to give legal advice? Definitely not.
Will I try to ask it for legal advice anyway? Of course.
The good news is that it’s remarkably good at explaining its thoughts. So it’s not really a black box.
The bad news is that it’s remarkably good at bullshitting explanations with confidence. So it’s even more of a black box.
Prompted:
In a chess game represented by the FEN notation: “r5rk/ppp1n2p/4qp2/3p1pRQ/3P3P/2N1P3/PP3P2/1K4R1 w”, what is white’s best move, and why?
Reply:
In the given FEN position: “r5rk/ppp1n2p/4qp2/3p1pRQ/3P3P/2N1P3/PP3P2/1K4R1 w”, it is White’s turn to move. The material is relatively balanced, with White having a slightly better position due to the active pieces and attacking chances.
White’s best move in this position is Rg5xg7+. This move captures the pawn on g7 with the rook while putting the black king in check. The idea behind this move is to take advantage of the exposed black king and the pinned knight on e7 (pinned to the queen on f6).
The position after Rg5xg7+ would be: “r5rk/ppp1n1Rp/4qp2/3p1p1Q/3P3P/2N1P3/PP3P2/1K4R1 b”
Black has only two legal moves in response to the check: Kh8 or Ng6.
If Black plays Kh8, White can play Rg7xh7+ followed by Qh5xh7# (a discovered checkmate). If Black plays Ng6, White can capture the queen with Rg7xf6, gaining a significant material advantage. In either case, White will have a decisive advantage, making Rg5xg7+ the best move in this position.
All of this sounds rather impressive, until you actually look at the board…
Not sure just how much we should anthropomorphize the AI.
That is, what characteristics are you looking for to determine that it lacks confidence? I would say that normally, we humans use social interactions and non-verbal cues to make this particular determination; and ChatGPT is incapable of the latter.
From the math examples above.
User: geometry problem.
GPT4: makes a mistake.
User follows up:
Note that GPT4 understands “geometry is not your strong point” to mean “you made a mistake”. It also translates the problem into askii using CSS script? which is pretty amazing.
User follows up, hoping that GPT4 can now literally see its error.
You could of course take that as “anthropomorphizing” but it totally does it itself.
Also I don’t know. GPT-4’s social skills can be impressive. Much better than Chat.
You could of course argue that they are entirely fake, but they’re still good.
See pages 56-59 here.
I don’t suppose you have a picture of the board?
And then…Board editor • lichess.org
Is the crux of the issue that Chatgpt says the move results in check when it obviously does not?
