Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Jeff Bezos on AI (1998) [video]

news.ycombinator.com

Post reply on HN
Print view View original

91–100 of 141 posts

Re: Jeff Bezos on AI (1998) [video]

#91

Post by ben_w »

oldandtired wrote: post #84
ben_w wrote: post #66

Earlier quoted context omitted.

> All of the tasks that you have mentioned have been programmed that way. It has taken human ingenuity to work out how to do this programming. The successful Go AI were programmed to learn ; we still can't program a decent Go AI with rules humans come up with. > The literature is there Do you have a link? Two Minute Papers just had a video about an AI systematic finding ways to confound other AI, but I thought we'd p…

Your example of the Go AI being programmed to learn is not all that accurate for what has been achieved here. I didn't keep the link for the discussion on the confounding of the Go AI system. What the discussion covered though was that there were simple Go configurations that the GO AI failed abysmally on when playing a human - it didn't learn here. I have spent forty years dealing with all sorts of computer systems…

> Your example of the Go AI being programmed to learn is not all that accurate for what has been achieved here.

What do you mean?

AlphaZero was trained entirely on self-play, and is a generic reinforcement learning algorithm. All it starts with are the rules (Chess, Go, Shogi) and a few million games later it beats — so far as I can see from a quick Google — all the humans, and most matches against AlphaGo Zero which learned the same way and which in turn beat AlphaGo Lee in every match, and that (unlike the aforementioned) was trained on examples of human matches in addition to self-play… but still learning from those examples as there's no known useful[0] set of rules that even says if a Go game is over let alone which moves are good.

There are AI which can find and exploit its weaknesses, but I've not seen anyone else suggest humans can defeat it.

> I didn't keep the link for the discussion on the confounding of the Go AI system. What the discussion covered though was that there were simple Go configurations that the GO AI failed abysmally on when playing a human - it didn't learn here.

Do you remember the name of the AI?

A bit of rummaging got me KataGo, but the humans had to use another AI to discover the weaknesses of KataGo rather than figuring it out for themselves.

And yes, KataGo absolutely does learn. The fact you can trivially stop the learning process is a feature not a bug for AI, precisely because it means any safety testing of the sort you're calling for is actually possible (albeit rather different than formal logic).

[0] pathological cases are easy — "board empty == not finished" — but not helpful.

Re: Jeff Bezos on AI (1998) [video]

#92

Post by sacado2 »

nologic01 wrote: post #36
mholt wrote: post #5

The common sentiment around AI in the 90s and early 2000s was that it didn't work; it had its hype, it had its heyday, but it seemed like a dead-end for the most part. The Perceptron was merely a linear function approximator. And the Multi-layer Perceptron was a little more capable, but the many orders of magnitude it would have to scale up in order to be convincing just wasn't feasible back then (it finally was in t…

AI is not a parlor trick. AI is a branch of statistics. Nobody said that statistics must limit itself to quasi-linear models of numerical data. It was just a limitation of computational resources (initially "AI" was developed by human computers). The trick is to get people not to associate the dictum "lies, damn lies and statistics" with "hallucinations, damn hallucinations and AI".

AI is not only a branch of statistics. Symbolic AI has nothing to do with stats.

Re: Jeff Bezos on AI (1998) [video]

#93

Post by sacado2 »

mholt wrote: post #5

The common sentiment around AI in the 90s and early 2000s was that it didn't work; it had its hype, it had its heyday, but it seemed like a dead-end for the most part. The Perceptron was merely a linear function approximator. And the Multi-layer Perceptron was a little more capable, but the many orders of magnitude it would have to scale up in order to be convincing just wasn't feasible back then (it finally was in t…

You seem to be conflating AI (in general) and strong AI. They are not the same thing at all. There have been industrial uses of AI techniques (I'm going by the definition from the 1956 Dartmouth workshop) for decades. And what's said in that video is a good example. We take recommendation modules for granted nowadays. They are applied AI.

Disappointing for sure, if you're still waiting for a sentient robot, but they do something we thought was limited to humans back in the day: going to a bookstore, and having the owner tell you "hey, I know you enjoy mystery novels taking place in England, I might have something new for you".

Re: Jeff Bezos on AI (1998) [video]

#94

Post by worksonmine »

IanCal wrote: post #89
worksonmine wrote: post #79

Earlier quoted context omitted.

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. If you learn to multiply with code to perform one task, you can then apply that knowledge for a completely different task. It may look like solving a completely new problem but the LLM doesn't even see the difference. When you use the term "custom li…

> . It's still just looking up function to do x, function to do y and applying it to the output. I mean no, no it isn't. I'm giving it info on how to construct data models with a custom library, so interacting with that is not using anything previously stored, and then giving it businesses/tasks to model as simple human descriptions. If you tell me that something which * Takes a human description of a problem * Descr…

It's difficult for me to assess how original your library is without examples, maybe I could find the exact implementation on github within 30 minutes. But I've yet to see anything that isn't just mashing together stackoverflow and git repositories to save time. I get the same answers with less wordy fluff from a simple search, but I also know where to look.

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".

It IS just looking in a database, and mashing it with some fluff. I'm happy to be proven wrong but I need more than your word for it. My experience is that as the topic gets more niche (less data in the training set) the worse the answers I get and it starts making things up based on probability. It doesn't reason in the sense I assume you're expecting.

Re: Jeff Bezos on AI (1998) [video]

#95

Post by bondarchuk »

keiferski wrote: post #90
bondarchuk wrote: post #87

Earlier quoted context omitted.

> A lot of people seem to think that if an AI can simulate the appearance of a human being, that makes them equivalent to one. That is not what BoiledCabbage was saying. He was saying: " And what do you feel when we understand it well enough to realize we're the same type of parlor tricks? " > This is a positivistic argument and as I pointed out, positivism has a lot of issues. The best counter argument IMO being tha…

If you haven’t done the reading, I can’t explain it to you in a HN comment. I’m not trying to be snarky about it, but I genuinely don’t know what else to tell you. This is a pretty foundational ideal in the philosophy of science. What’s wrong with the speculation is that it’s a positivistic argument that is needlessly reductive. It’s reductive because it assumes that appearing human-like is equivalent to being human.…

>It’s reductive because it assumes that appearing human-like is equivalent to being human.

I don't read that assumption into BoiledCabbage's statement at all: "[..] when we understand it well enough to realize we're the same type of parlor tricks?" This clearly implies a (hypothetical) deeper understanding of processes in the brain and their specific qualities, rather than (as you seem to be implying) a mere comparison of the outputs.

Edit: anyway, the criticism section opens like this:

>Historically, positivism has been criticized for its reductionism, i.e., for contending that all "processes are reducible to physiological, physical or chemical events," "social processes are reducible to relationships between and actions of individuals," and that "biological organisms are reducible to physical systems."

This (at least the 1st and 3rd quoted item, while I think the 2nd one is just out of scope) is exemplary of the kind of things that are obviously true for anyone but a subset of philosophers clinging to magical and unprovable beliefs about the human mind. I asked you to elaborate your argument precisely because if it all boils down to simply rejecting physicalism (in philosophy of mind terms) there's nothing new to argue about. The recurring discussion about "AI can never be like humans" is only interesting when the participants do a little bit more than just staking out their own position in idealism vs dualism vs physicalism terms and regurgitating all the known debates between these camps.

Re: Jeff Bezos on AI (1998) [video]

#96

Post by IanCal »

worksonmine wrote: post #94
IanCal wrote: post #89

Earlier quoted context omitted.

> . It's still just looking up function to do x, function to do y and applying it to the output. I mean no, no it isn't. I'm giving it info on how to construct data models with a custom library, so interacting with that is not using anything previously stored, and then giving it businesses/tasks to model as simple human descriptions. If you tell me that something which * Takes a human description of a problem * Descr…

It's difficult for me to assess how original your library is without examples, maybe I could find the exact implementation on github within 30 minutes. But I've yet to see anything that isn't just mashing together stackoverflow and git repositories to save time. I get the same answers with less wordy fluff from a simple search, but I also know where to look. It's impressive that it knows the difference between "how m…

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...

Re: Jeff Bezos on AI (1998) [video]

#97

Post by keiferski »

bondarchuk wrote: post #95
keiferski wrote: post #90

Earlier quoted context omitted.

If you haven’t done the reading, I can’t explain it to you in a HN comment. I’m not trying to be snarky about it, but I genuinely don’t know what else to tell you. This is a pretty foundational ideal in the philosophy of science. What’s wrong with the speculation is that it’s a positivistic argument that is needlessly reductive. It’s reductive because it assumes that appearing human-like is equivalent to being human.…

> It’s reductive because it assumes that appearing human-like is equivalent to being human. I don't read that assumption into BoiledCabbage's statement at all: "[..] when we understand it well enough to realize we're the same type of parlor tricks?" This clearly implies a (hypothetical) deeper understanding of processes in the brain and their specific qualities, rather than (as you seem to be implying) a mere compari…

I don't read this statement as hypothetical at all.

And what do you feel when we understand it well enough to realize we're the same type of parlor tricks?

It seems pretty clearly when and not if. Not sure what you're reading there.

Re: Jeff Bezos on AI (1998) [video]

#98

Post by bondarchuk »

keiferski wrote: post #97
bondarchuk wrote: post #95

Earlier quoted context omitted.

> It’s reductive because it assumes that appearing human-like is equivalent to being human. I don't read that assumption into BoiledCabbage's statement at all: "[..] when we understand it well enough to realize we're the same type of parlor tricks?" This clearly implies a (hypothetical) deeper understanding of processes in the brain and their specific qualities, rather than (as you seem to be implying) a mere compari…

I don't read this statement as hypothetical at all. And what do you feel when we understand it well enough to realize we're the same type of parlor tricks? It seems pretty clearly when and not if. Not sure what you're reading there.

"When" is in the future so I don't get your issue. Unless your point is that we can in principle never hope to understand certain things about it.

Re: Jeff Bezos on AI (1998) [video]

#99

Post by keiferski »

bondarchuk wrote: post #98
keiferski wrote: post #97

Earlier quoted context omitted.

I don't read this statement as hypothetical at all. And what do you feel when we understand it well enough to realize we're the same type of parlor tricks? It seems pretty clearly when and not if. Not sure what you're reading there.

"When" is in the future so I don't get your issue. Unless your point is that we can in principle never hope to understand certain things about it.

When implies that the thing is going to happen. If implies that it might happen.

https://dictionary.cambridge.org/grammar/british-grammar/if-...

The comment is written in such a way that it assumes this is the nature of reality and at some point, we will learn that the human mind is no different from the parlor tricks of LLMs. This is what I was criticizing.

Re: Jeff Bezos on AI (1998) [video]

#100

Post by worksonmine »

IanCal wrote: post #96
worksonmine wrote: post #94

Earlier quoted context omitted.

It's difficult for me to assess how original your library is without examples, maybe I could find the exact implementation on github within 30 minutes. But I've yet to see anything that isn't just mashing together stackoverflow and git repositories to save time. I get the same answers with less wordy fluff from a simple search, but I also know where to look. It's impressive that it knows the difference between "how m…

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…

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

Post reply on HN