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

The AI Battle

aifuture.substack.com

Post reply on HN
Print view View original

121–130 of 137 posts

Re: The AI Battle

#121

Post by bsenftner »

rytill wrote: post #113
bsenftner wrote: post #100

Earlier quoted context omitted.

One can not explain anything to an AI, it has no comprehension. You're talking to nothing, a brick. Lacking comprehension, using "intelligence" in the name of AI is a marketing joke. I write the stuff, at a very high level. All the "software engineer careers are doomed" crap is nonsense. Lacking comprehension, AI is an idiot savant and a damn good actor, and nothing more because there is literally nothing inside.

So what specific task will AI definitely not be able to do in 5 years?

The majority of what people want, because people will expect to be able to explain what they want, and that is speaking to a uncomprehending wall. Lacking comprehension is incredibly fundamental.

For example: you question I am answering, that act of answering your question requires comprehension of your question. Lacking comprehension, only a lookup of canned answers to expected questions is possible.

Re: The AI Battle

#122

Post by ben_w »

LeanderK wrote: post #56
stuckinhell wrote: post #13

AI that can do creative work is scary for its implications on humanity. Are we at best biological AI's ? I've played around with several of them. These artwork AI's are way more scary than GPT-3 et al because they seem to do something so creative with such good visible results. I don't think the genie goes back in the bottle. I think humanity has an existential crisis on its hands. Whats the point in doing things if…

> Are we at best biological AI's ? imho yes. But don't have an existential crisis just yet. It's a very narrow task where the AI systems do well in a way that we don't expected from computers. But it's well chosen, there are still only few tasks where significant progress was made. They really can't make "logical deductions" at all and we have no idea how. Everything you see is learned via a massive amount of labeled…

> They really can't make "logical deductions" at all and we have no idea how.

What exactly do you mean by this? Because this sounds like the exact opposite of the problem AI has — logic is the easy part, and has been working in machines since they were clockwork and punched cards and is the foundation for 100% of the functionality of modern computers, but natural language comprehension is only just starting to be possible now, and only at a fairly rudimentary level.

Re: The AI Battle

#123

Post by TaupeRanger »

thorum wrote: post #110
TaupeRanger wrote: post #102

Earlier quoted context omitted.

I don't follow any of that. If I have an idea about a new type of toaster...how does DALL-E help me by giving me a bunch of computer generated toaster images? If I want to make a painting of a solar system floating in front of a nebula, I could ask DALL-E to do it, and maybe I'll get some ideas on composition/style, but is that really such a "gatekept" process? And does it even matter? Is my art better or more meanin…

Toasters are probably not a great example, but concept art is a big deal in most product development / creative industries. Tech, cars, Hollywood, fashion, anime, etc. The process usually involves finding and hiring an artist and working with them as they create lots and lots of designs until you find one you like, and then more sketches until the design is refined. So you either need to have a significant amount of…

That's not how DALL-E works. It doesn't understand specifics enough to give you something useable in any of those fields. A car company is not going to use a DALL-E image as a car. It might have a professional designer, who themselves use DALL-E to look for something interesting, but that is not "disrupting" the car concept drawing market like the author claims. DALL-E isn't capable of that. Same with tech, Hollywood, etc. For fashion, I suppose you could just create whatever weird thing DALL-E throws up, but again this isn't any kind of disruption, it's just a novelty thing.

Re: The AI Battle

#124

Post by LeanderK »

ben_w wrote: post #122
LeanderK wrote: post #56

Earlier quoted context omitted.

> Are we at best biological AI's ? imho yes. But don't have an existential crisis just yet. It's a very narrow task where the AI systems do well in a way that we don't expected from computers. But it's well chosen, there are still only few tasks where significant progress was made. They really can't make "logical deductions" at all and we have no idea how. Everything you see is learned via a massive amount of labeled…

> They really can't make "logical deductions" at all and we have no idea how. What exactly do you mean by this? Because this sounds like the exact opposite of the problem AI has — logic is the easy part, and has been working in machines since they were clockwork and punched cards and is the foundation for 100% of the functionality of modern computers, but natural language comprehension is only just starting to be pos…

Maybe i could have phrased it differently, i'll try my best to explain it. Keep in mind that this is open research, it can change quickly with breakthroughs. What you mean is strictly "following" logic, by executing code or combining axioms like in prolog. What I want to get at is maybe better described as "reasoning". Learning by thinking about stuff and combining knowledge, not by example. Our current models can't do this at all, this was all the rage of old-school, logic based AI (but this also didn't work at all, hence the AI-winter). Just think about the difference between learning to play tennis (repetition and exercise, learning from errors without much reasoning) and my IKEA furniture example, for which you are expected to assemble it on your first try without guidance or repetition. It turns out that we can solve, through repetition and exercise, a lot of problems that were previously thought have a lot to do with reasoning, like dalle-2 or gpt-3, this involves huge amounts of data and long training times. Is it all solvable by repetition and exercise? I doesn't look like it. The learning process is so fundamentally different that we have no idea how to build systems that learn by explanation and have the ability to "think hard about a problem". Some researchers are convinced it can be done, but we currently can't do this at all and there's not really an indication that it is possible using our current approaches.

My personal opinion is that we now have a hammer and everything looks like a nail. I don't think everything is a nail, but surprisingly many problems are, if you phrase the problem correctly. In practice this means that if we can gather enough training data then a lot of problems suddenly become solvable, but this is not possible for all problems. If we can not gather enough training data, then we have have a problem we just can not solve and there's no indication that it is solvable with current tools. It would have to "reason" and "think hard" about the problem, we can't do that. All those fancy things work by ever increasing datasets. This is currently a hard limit and I can perfectly imagine that we have just solved one of the ingredients for better AI. And just like rolling a dice, if you have rolled two 6s in a row the probability for another 6 is still 1/6. If we need another breakthrough this can take years or decades and just because we've made one in 2012 this doesn't mean the next will happen in 2022.

Re: The AI Battle

#125

Post by challenger-derp »

slowmovintarget wrote: post #67
tialaramex wrote: post #29

Earlier quoted context omitted.

> those systems are not a path to AGI If that's the actual meat of your claim, it's weird to just put it unsupported at the end. If you think it's an aside you've got a big problem, because without it the rest is pretty weak. We saw this with chess. Anybody who was clinging on to the idea that well, the machines can't really play chess, because technically there do seem to be a few humans who are better, was screwed…

On "not the path to AGI": Gary Marcus on the Mindscape podcast is worth a listen. https://www.youtube.com/watch?v=ANRnuT9nLEE&list=PLrxfgDEc2N... Transcript: https://www.preposterousuniverse.com/podcast/2022/02/14/184-... "...And then there’s natural language understanding and reasoning, and I would say we have not really made progress at all. GPT-3, which we may wanna talk about, gives the illusion of having natural…

For anyone who's reading the above comment, the additional context that slowmovintarget hasn't provided is that Gary Marcus supports a school of thought in AI that opposes the currently popular school of thought that favors deep neural networks. A frequently contested point is what each school thinks is the path to AGI. Marcus is a well-known figure in AI.

Re: The AI Battle

#126

Post by challenger-derp »

CuriouslyC wrote: post #34
hoppla wrote: post #22

Are AIs capable of creating something really unique, or is it just a washing machine spinning on old ideas?

It can apply patterns learned in different contexts to new contexts, which has the potential to create things that are unique in a sense.

^ quite a good way of concretizing the concept of interpolating between training samples.

Re: The AI Battle

#127

Post by bckr »

CuriouslyC wrote: post #54
Retric wrote: post #49

Earlier quoted context omitted.

Modern word processors haven’t suddenly and dramatically increased the number of great books available. They save a lot of time and effort relative to a typewriter, but such drudgery isn’t the bottleneck on creativity that you’re suggesting.

I know a lot of writers who are very good at creating an outline and describing what's going on but poor at actually sitting down and getting words on paper for any extended period of time. These same people can read and critique/edit/etc endlessly. I think it's a fairly common problem because block is the number one topic in most forums for writers. Having a tool that takes an outline and generates a rough draft of…

This is a great use case and in fact I would pay for this service. Sitting and barfing up text can be fun when I'm inspired, but I'm frequently not inspired but would still like to make progress on my stories.

Re: The AI Battle

#128

Post by CuriouslyC »

awillen wrote: post #44
CuriouslyC wrote: post #32

Earlier quoted context omitted.

I don't think these sorts of tools will ever be used to generate a novel or something on that scale. We're building really sexy autocomplete tools that creators will use to fill in the blanks much like the great masters of the renaissance used apprentices to do much of the work in their masterpieces. People will outline what they want, then ask the AI to fill in the blanks, and iteratively refine the result. As long…

Ever is an exceedingly long time. I would be shocked if 100 years from now, we didn't have AI-authored bestselling novels. Even then, they might not be literary masterpieces, but certainly AI will be able to write formulaic stuff that sells really well. If you don't believe that, just consider where technology was 100 years ago and what the response would have been if you'd described DALL-E in its current incarnation…

I think it'll always be "human decides what book is about, cues AI, then gives feedback to AI to refine output," The cues will just need to be less specific and well crafted, and the amount of feedback required will go down. Maybe eventually AI will be able to one-shot amazing novels, but they'll still need taste makers to read the output and promote it, which isn't really much faster than a taste maker asking for what they want directly then reading/requesting changes.

Re: The AI Battle

#129

Post by marvin »

donkarma wrote: post #117

all of these models (GPT-3, DALL-E) massively infringe on copyright and I expect them to be demolished in court

So does the brain of all artists that have learned and been inspired by copyrighted works.

Re: The AI Battle

#130

Post by YeGoblynQueenne »

carapace wrote: post #111
YeGoblynQueenne wrote: post #73

Earlier quoted context omitted.

You were too pessimistic! Automatic programming, a.k.a. program synthesis, has been a thing since the early days of AI and computer science, for example the idea of deductive program synthesis, where a program is generated wholesale from a complete specification in a formal language that is not a computer language goes at least as back as Alonzo Church himself, in 1957: https://en.wikipedia.org/wiki/Program_synthesis…

That's right, this is your field isn't it? Cheers! Gosh, the robot car work was incredible, I can't believe I've never heard of it before! - - - - Copilot isn't cutting edge but it's significance is just that it's a mass market tool. It will be interesting to see how well it does, and whether users find it worthwhile overall after a couple of years. Will it be improved to the point where it starts to compete with its…

Ah, my field is Inductive Logic Programming - a sub-sub-sub-sub field of program synthesis. But I need to know the basics!

I think Copilot can be used very effectively, as long as its capabilities and their limitations are communicated clearly. For instance, I think it can make a great boilerplate generator, as long as users stick to short code snippets.

Well, I don't know about replacing programmers. I think that's Sci-Fi, for the time being, and for a while longer still. What I'm more interested in is creating tools to help programmers do their job. Copilot does that already, btw, I'm not dissing it. I'm just pointing out it doesn't represent a sudden shift in capabilities, to be clear.

>> (One of the reasons I like Schmidhuber is that his goal, since early on, is to "Create an automatic scientist, and then retire.")

I didn't know Schmidhuber had said that. My thesis advisor, Stephen Muggleton, was part of an interdisciplinary team who created a robot scientist that can develop its own theories and then choose, and run, the experiments to prove them:

https://en.wikipedia.org/wiki/Robot_Scientist

Another one of those things that are not well-known, I guess. I wasn't involved with that, btw, but I think recent advances could make for a much more powerful system. I am considering something similar as a research project, post-doc.

Post reply on HN