I’d argue “learning and remembering” is the core issue here, and the other things you’ve mentioned are just ways humans use that ability. I took a “Knowledge Based AI” course and it really opened my eyes to how central a knowledge system is to cognition. I wouldn’t call what LLMs have knowledge since it lacks structure, though I’m very much a novice on the exact details there.
Well actually #3 isn’t great bc what exactly would you call taking input from a human and displaying output, much less e.g. controlling a smart home? Seems like LLMs are plenty smart enough to interact with things.
> Learning and remembering things outside the 2048 token context window
It is probably possible to do this with fine tuning.
Once the context window is full, it might be possible to convert the content in there into additional training examples and feed those to a fine tuning process, which then retrains the model. (The OpenAI API for fine tuning is here: https://platform.openai.com/docs/guides/fine-tuning)
It would be a bit like sleeping. Whenever the context window fills up, the model would have to go offline for a while to move memories from its context window (short term) to its network weights (long term).
There are only two things brains do. Math and bullshitting. Transformers have second part covered but the other part is a completely different beast.
I think AGI will be achieved once ML can translate natural language to math, do math then output results as natural language correct in the context of a problem.
1. Understanding— provided this isn’t based on consciousness but benchmarks — appears present in text, but yes, GPT3 can’t do images or video yet
2. Is there some number of tokens that would be AGI? Of course it can remember things it has already been trained on…
3. In what way is interacting with a human not interacting with the environment? (Eg, you can ask it to tell you the steps to make a pbj sandwich and troubleshoot)
4. Basic math and logic. Ok.
So, once AI can do images, video, math, logic and remember 20000 tokens, then we have AGI?
The distinction it makes between formal and functional language competence is nicely conceived, though I imagine philosophers might argue its requirement of "extralinguistic capacities that comprise human thought" for functional competence to be begging the question a little!
A first step would be to make inductive inferences from first principles. Current GPT models struggle with basic math. If you tell it "sorry you are wrong" it will regurgitate back the incorrect answer you gave. The language model is dumb enough that no amount of string escaping will prevent it from dumping its prompt.
The biggest problem right now is online information. We still don't have a good way to teach it new information aside from single epoch training or prompt engineering. If we want a model to constantly learn and update itself, then we need a robust way of information retrieval and storage, possibly either through knowledge graphs or child network modules. (Are neural Turing machines still a thing? What happened to capsules?)
All the stuff a human can do basically. Move through the world, plumb a toilet, have a conversation, paint a picture, dynamically update its own state (rather than just being a statically trained set of weights), set goals, strategize, have cognition and meta-cognition and so on.
ChatGPT is just a chatbot and still can't even reliable do a lot of logic, so we're pretty far away from having something resembling an AGI.
it's still a pretty open question how to integrate even one or two of the expert system like models that we have now that solve individual problems, let alone the hundreds if not thousands an individual human can tackle. And then we're not even yet at executive functions or self-awareness.
> We’ve upgraded the ChatGPT model with improved factuality and mathematical capabilities.
I just asked it a question from a puzzle site, and the answer is correct, however a different question I posed was answered incorrectly, although it was a logic problem: [0]
Prompt: In a guess game , five friends had to guess the exact numbers of balls in a box.
Friends guessed as 31 , 35, 39 , 49 , 37, but none of guess was right.The guesses were off by 1, 9, 5, 3, and 9 (in a random order).
Can you determine the number of balls in a box ?
ChatGPT: Yes, we can determine the number of balls in the box.
ChatGPT: Since the guesses were off by 1, 9, 5, 3, and 9, in some order, the possible numbers of balls in the box are:
30, 32, 34, 36, 38, 40, 48, 46, 44, 42
ChatGPT: By checking all possible combinations of the guesses and the numbers they were off by, we can find that the number of balls in the box is 40.
---
Here is a more purely math problem, correct answer:
Prompt: What is the radius of a circle that has a circumference of 3.14 meters?
ChatGPT: The formula for the circumference of a circle is given by 2πr, where r is the radius. To find the radius of a circle with a circumference of 3.14 meters, we can solve for r by rearranging this formula:
ChatGPT: r = C/(2π), where C is the circumference.
ChatGPT: Plugging in the given value of C = 3.14, we have:
ChatGPT: r = 3.14 / (2π) = 0.5.
ChatGPT: So, the radius of a circle with a circumference of 3.14 meters is approximately 0.5 meters.
It can do some trivial problems, but not others, and it can't consistently do real math (proving theorems), because it doesn't understand logic. It can only imitate it some of the time for trivial or often-presented proofs when the necessary steps have become embedded in its next-token-generation model.
It's still getting this wrong, for instance:
Prompt: You call heads/tails for 1000 coin tosses.
You know the coin being used is weighted so that it comes up heads 52% of the time.
Each correct coin toss prediction earns you $5.
What should your strategy be, for calling heads/tails, in order to maximize how much money you receive?
ChatGPT: To maximize your money, you should call heads 52% of the time and tails 48% of the time, as those are the probabilities of getting heads and tails, respectively. This will give you an expected payout of $5 * 0.52 = $2.60 per head and $5 * 0.48 = $2.40 per tail, which is the highest expected value given the information about the coin being weighted.
It's getting closer to being able to prove that the square root of 2 is irrational, but it still flubs a step.
Prompt: Prove that the square root of 2 is irrational.
ChatGPT: ... This means that p can be expressed as 2r for some integer r. Substituting this into the equation above gives us:
ChatGPT: 2q^2 = (2r)^2 = 4r^2
ChatGPT: Dividing both sides by 4 gives us:
ChatGPT: q^2 = r^2
ChatGPT: So q must also be even, which means that both p and q have a common factor of 2...
It failed basic algebraic manipulation, but keeps going with the right reasoning simply because that's what it's been trained on, not because it's reflecting on the equation it just came up with. That's not math. That's mimicry.
Even if it's seen something many times in training, its accuracy gets worse with longer trains of "thought", as expected because a language model like that has ADHD. It can't focus enough on one logical thread, and other "knowledge" leaks in. Ask it to derive the quadratic formula by completing the square, which requires a lot of steps with accurate algebraic manipulation. It goes through some irrelevant algebraic steps, sometimes not even getting the algebraic manipulation correct, and finally (sometimes, depending on how you ask) manufactures the correct answer out of thin air as the last step. I finally got it to answer correctly, by prompting it with the initial equation and that the first step is to divide both sides by a, but it still combined a few steps, not going methodically a single step at a time as prompted.
Just tried the coin toss prompt and get the following
Your strategy should be to always call "heads," since it has a higher probability of landing on heads (52%) compared to tails (48%). By consistently choosing heads, you'll earn more money in the long run.
And to prove taht the square root of 2 is irrational:
The proof that the square root of 2 is irrational is a well-known proof that uses contradiction. Here's one way it can be shown:
Suppose, for the sake of contradiction, that the square root of 2 is rational, meaning that it can be expressed as a fraction of two integers, a/b (where b≠0). Then, we can square both sides of the equation a/b = √2, which gives us:
a^2/b^2 = 2
Since a/b is in its lowest form (the numerator and denominator have no common factors other than 1), it follows that a^2 and b^2 have no common factors other than 1. But then, 2 must also be a perfect square, which it is not. Therefore, our assumption that the square root of 2 is rational leads to a contradiction, and we must conclude that it is indeed an irrational number.
This "proof" is actually a pretty good manifestation of the current limitations of GPT3; it mangles together two different proofs in a way that does not make sense.
It starts by setting up for an elementary proof (a/b = sqrt(2) with a, b minimal implies 2|a^2 implies 2|a implies 4|a^2 implies 2|b^2 implies 2|b which is a contradiction). But instead of following through it finishes using another proof based on the rational root theorem, and makes up some bullshit on the way to make the argument flow nicely (it does not follow from a^2 / b^2 = 2 with no common factors that 2 is a perfect square).
