[IanCal]

Have you had a look at othello gpt? https://thegradient.pub/othello/

It's a nice constrained example of a transformer learning a world model, not just looking up responses.

> It's impressive that it knows the difference between "how many are 5 more apples than 10" compared to "how many percent are 5 apples of 10" (I don't know if it does, just assuming). But the first release also tried to reason why the weight of 1 pound of nails depends with the simple prompt "how much do 1 pound of nails weigh". That's most likely a perfect example of it mashing the classic "what weighs more, 1 pound of nails or 1 pound of feathers".

Is there a formulation here that would get to a point where you'd think it's not just mashing things together? Are there elements of a simple question that would be required?

Here's a slightly trickier one for it "Which weighs more, a pound of feathers or balloons made from one pound of rubber then filled with 100g of helium?"

https://chat.openai.com/share/b841c96f-e46c-4adf-8ec3-8778ff...

[worksonmine]

Very impressive, but is it any more original than classic search engines' old trick of regular expressions to figure out if I mean the currency or weight when I ask "1 pound =" with the contexts USD or kg after "="? Does it understand the input, or are there just enough discussions in the training data to make it look like it is? I'm not convinced it's not the latter.

It uses context to figure out we're trying to convert something to something else. Then it adds all those numbers up. Taking helium into consideration is no doubt interesting, but they've also polished that task since that was the common critique they got so very wrong with the first release (which I mentioned they had fixed). I'm not qualified to assess this part of the answer;

> "If the balloons displace more than 100g of air when filled with helium, then they would effectively weigh less than if they were left empty. If they displace exactly 100g of air, then the balloons would have the same weight as if they were left empty."

I don't know enough to understand how much 100g of helium is and how it behaves. And it doesn't try to explain it to me, it mentions it then takes the easy route assuming it's a trick question. What does that tell you? I guess there are similar discussions around and it gives me the summary. Why doesn't it tell me how much air it displaces under what circumstances? Temperature etc, it should be easy if it's not just a simple discussion on a random forum. A conversion regex could do it.

This comment[1] has a very impressive example. But anything I'm qualified to assess has mostly been meh. If the fix is better training data does that mean it's reasoning or regurgitating? The mistakes it makes are what tells me how it works, not when it tricks me that it's correct. To me it's a very well polished search engine summary.

[1]: https://news.ycombinator.com/item?id=37219351

[IanCal]

If you've not looked at it I really recommend othello gpt. That is an experiment explicitly designed to tackle this kind of question, has it just seen enough moves that it knows what should come next?

> Why doesn't it tell me how much air it displaces under what circumstances?

You can ask it and it'll answer.

> If the fix is better training data does that mean it's reasoning or regurgitating?

More training data helps with things you can just bring to the fore, same as a lot of learning. More useful training data though can also help reasoning, which makes sense - deliberate training of people helps improve their logical reasoning capabilities. I know that doesn't guarantee that's what LLMs are doing but humans benefit significantly from both more teaching and better teaching.

> Very impressive, but is it any more original than classic search engines' old trick of regular expressions to figure out if I mean the currency or weight when I ask "1 pound =" with the contexts USD or kg after "="? Does it understand the input, or are there just enough discussions in the training data to make it look like it is?

I'd be interested to know any requirements around this to clearly show the difference. I tried asking what if I filled a balloon at a childrens party with a gas made of atoms that have 1 proton and 100 neutrons: https://chat.openai.com/share/71224df4-5c6c-45f7-88fd-eec316...

(tl;dr: "In the context of a child's birthday party, introducing such a balloon would be a grave mistake.")

It identifies:

* Whether it would float or not and why * That it would be radioactive, and likely types of radiation from it * What that would mean to the balloon * How people would react and the likely consequences of releasing it in a room of children

This is an element that does not exist, in a setting where nothing like this has happened before, with details ranging from types of decay, consequences and human emotional reactions to something like this. Yes, there are real things you can use as a base (e.g. how do people respond to events that kill people), but I feel it's an example of where it's beyond a search engine summary.

[worksonmine]

> If you've not looked at it I really recommend othello gpt.

I skimmed it and read the conclusion, and it looks interesting, will take a closer look when I have time.

The prompt covers a subject that goes completely over my head so I can't tell how well it reasons. I don't know what 1 proton to 100 neutrons means, but I gather it's radioactive. I don't think it's far fetched that it draws the same conclusion from the training set because to you it seems obvious, and is probably well known to anyone who knows the subject. Kind of like it would understand that "hotter than the sun" is super hot, can correlate to different melting points. But I wouldn't say it understands the concept of temperature. Given the right prompt it might give you the impression it does.

The feelings of the scenario reads like any PR comment after a tragedy. "We feel shock and disbelief" and so on. The scenario being hypothetical doesn't change that since it's probabilities. It acts just like you'd think it would. The earlier example with the helium balloon is similar, it assumes a human context and not the form, and environment the helium is in. True intelligence might not even consider the presence of atmosphere as the norm. "It has no weight outside of your human constraints" would be novel.

Lets say it has odd numbers between 1-9 in the database. Given the prompt 2 and 8 you will get back 1,3,7,9, sprinkled with some natural language and we get the impression it's intelligent.

Are you saying it understands the effect the neutron to proton ratio has, as opposed to just comparing the vectors closest to your prompt that it builds the answer from? Being tested on new and hypothetical examples only means it will be further from the vectors but still close enough to give us the impression it understands the subject. If the training data didn't include the words neutron or proton it would have no idea where to begin.

In my first comment that started this chain I said:

> I don't see why not. It's not taking a single answer from a database no, it's taking several based on probability and merging them into what it thinks we're looking for.

I don't think even this latest answer is any proof of anything other than that. Are you claiming there is? And what are you claiming is happening?

[IanCal]

> I don't think even this latest answer is any proof of anything other than that. Are you claiming there is? And what are you claiming is happening?

I'm claiming that it's reasoning through the problem.

> True intelligence might not even consider the presence of atmosphere as the norm

I hugely disagree, that's not how an intelligent human would answer the question, and if it did this people would be complaining that it clearly doesn't understand the human context in which the question was likely asked.

> I don't know what 1 proton to 100 neutrons means, but I gather it's radioactive.

It would be hydrogen, but a type of hydrogen that doesn't appear in nature. It's a deliberately absurd example so that it's not in the training set. Its answers are different if the question involves tritium which while radioactive has a moderate half-life and wouldn't immediately pop the balloon.

> I don't think it's far fetched that it draws the same conclusion from the training set because to you it seems obvious,

Only because I can reason through what would happen, not because it's something I've seen talked about before.

To figure out what it would do, it cannot rely on an explanation elsewhere, it needs to first identify that the ratio of protons to neutrons is extreme. Then it needs to understand that this typically results in particular kinds of radiation.

It has to then use that information to consider how that would interact with the material of the balloon (and that this is important).

It has to use that information to consider how it would affect people, and what their reactions would be both before and after it explodes/pops.

This is multi-step reasoning through an issue that involves pulling together common expectations, physics and how humans react.

Here's a statement in it that shows to me more than just pulling a few answers together

> Balloon Behavior: Instead of floating up like a helium-filled balloon, this balloon would drop to the ground because the gas inside is denser than air. This might surprise the attendees, and curious children might approach or pick up the balloon, further exposing themselves to radiation.

-

> The feelings of the scenario reads like any PR comment after a tragedy. "We feel shock and disbelief" and so on.

Those are typical things, which is not surprising, but it is also clearly linked with the question. You have to understand how out of context this would be.

> If the training data didn't include the words neutron or proton it would have no idea where to begin.

Fully rediscovering what took humans many years to do off-the-cuff is an outrageously high bar.

What features of a question would you look for to identify whether it's "taking several answers from a database and merging them together" or performing some reasoning? I've asked a few times but don't understand what you're expecting.

[worksonmine]

> Fully rediscovering what took humans many years to do off-the-cuff is an outrageously high bar.

> What features of a question would you look for to identify whether it's "taking several answers from a database and merging them together" or performing some reasoning? I've asked a few times but don't understand what you're expecting.

You're misunderstanding me, I'm setting no bars, and I have no threshold where this changes. We humans are also just looking things up in our database and doing deductions. We do some computing on urgency as well, like how when we hear a bang our mind goes for danger first before realizing it was harmless, but very similar to what these AIs do. Probabilities and experience. Fresh and novel ideas are very rare in humans as well, and not something I demand before I would consider someone a human.

I did however give you an example that would surprise me, if it considered mass and environment in a way that proves that it understands the problem for what it is. If it told me weight is a human construct and requires gravity/movement and how it depends. An intelligent human doesn't necessarily answer the question it is asked in the way it is phrased. It identifies and irons out misunderstandings, assumptions and other details important to correctly understand the problem, and may even rephrase the question to give a proper response. That would show me a deep understanding of the problem and maybe freak me out a little, but only if the hallucinations are gone and those can be difficult to spot.

This is, just like us, performing calculations and database look-ups. It may feel like it's doing something else but it's not. What would happen if we leave the weights as they are but switch the words? It would give us complete gibberish, but it's no less correct than it was before and it's not even giving us different answers, only the translations to language get distorted. Most people would call it stupid and pointless even if the only change is our interpretation of the answers.

I'm sure Hiroshima, Fukushima and other dangers of radiation is in the training set, as are all of the other steps you mention, it goes round and round testing the numbers based on training. Remember how this chain started, you claimed:

> They're not just retrieving stored text like pulling the most relevant passage from a database. If they were they'd not be able to deal with things outside the training set.

To which I simply replied:

> It's not taking a single answer from a database no, it's taking several based on probability and merging them into what it thinks we're looking for.

I read you (correct me if I'm wrong) as giving this way to much agency. To change my mind that it's doing something unexpected I would ask for logs on the calculations it does, and be able to correlate that to the training set. I have to be able to falsify the conclusions I'm asked to make. I know some people claim we don't understand these algorithms but I assume that's just hyperbole and with the correct measures we could follow every step.

If there are things there which I can not trace I would be very impressed, and honestly a little afraid. They are not trained for every single task, but approximations based on similarities have proven to be very capable even when we think we're out of context (we're not, it doesn't understand context and doesn't care, but neither do most humans).

[IanCal]

Perhaps we've been talking past each other then. I've been trying to show that these things can do reasoning, and my upper benchmark is "like a human". If you're starting from "these things might be doing what humans are doing / capable of performing similar tasks" then we're largely aligned.

The further question I still find interesting though.

> I read you (correct me if I'm wrong) as giving this way to much agency. To change my mind that it's doing something unexpected I would ask for logs on the calculations it does, and be able to correlate that to the training set. I have to be able to falsify the conclusions I'm asked to make. I know some people claim we don't understand these algorithms but I assume that's just hyperbole and with the correct measures we could follow every step.

This one is tricky. We know exactly what they do. Interpreting that is very hard though, they've a big pile of mathematical operations with billions of magical constants and it... works. We can see exactly what they do but if I could see every synapse firing in your brain I'd still not be able to understand how it works in a useful manner. So we understand them obviously, but at another level we really don't.

> I'm sure Hiroshima, Fukushima and other dangers of radiation is in the training set, as are all of the other steps you mention, it goes round and round testing the numbers based on training.

Just to be clear here, there is no recursion other than when you add more text. There is not an algorithm saying "identify parts X, then look in database Y, now summarise...". They're trained essentially to just predict the next word given some text. There's some later training to make them more conversational. The capabilities you see are just a consequence of that.

Othello GPT is show just moves. It ends up building an internal model of a board.

> I did however give you an example that would surprise me, if it considered mass and environment in a way that proves that it understands the problem for what it is. If it told me weight is a human construct and requires gravity/movement and how it depends. An intelligent human doesn't necessarily answer the question it is asked in the way it is phrased. It identifies and irons out misunderstandings, assumptions and other details important to correctly understand the problem, and may even rephrase the question to give a proper response. That would show me a deep understanding of the problem and maybe freak me out a little, but only if the hallucinations are gone and those can be difficult to spot.

Let's try and investigate that then, that sounds interesting. I'm not sure I understand myself what you mean that weight is a human construct (it explains the difference between weight and effective weight in the answers to me off the bat, that's the only real difference). Perhaps this is too simple, the answer is quite straightforward.

> If it told me weight is a human construct and requires gravity/movement and how it depends

I asked "Which is the most, a pound of feathers or a pound of helium?" with largely just your statement as the system message and got

>The question seems to contain an intrinsic confusion. When we discuss weight, both a pound of feathers and a pound of helium would weigh the same - a pound. The difference, though, comes in their volume and density. A pound of helium would take up a lot more space than a pound of feathers considering the density of helium is lower than the density of feathers. If you were implying which would be more in terms of volume, then a pound of helium would be significantly more than a pound of feathers.

> However, I might be wrong if we take into account that helium, being a gas, is usually measured in terms of its volume at standard temperature and pressure, rather than by weight like solid or loose materials such as feathers. Also, the weight of a pound can vary slightly depending on where on Earth it is measured due to differences in gravity. However, these factors don't fundamentally change the answer to the question as it was posited.

Perhaps instead you could give me a short question and the kind of answer that would surprise you? I know this thread has gone on some time, but personally this is interesting to me. If you wanted to shift off from hn, feel free to drop me an email, I have a vested interest in understanding how people view LLMs.