Here’s an interesting detail that I feel should have been hinted at in the headline, and would have made me far more likely to click without being clickbait, and while leaving some interesting meat in the body that would have made the read worthwhile in my opinion:
Amateur exploited weakness in systems that have otherwise dominated grandmasters.
[…]
The tactics that put a human back on top on the Go board were suggested by a computer program that had probed the AI systems looking for weaknesses. The suggested plan was then ruthlessly delivered by Pelrine.
“It was surprisingly easy for us to exploit this system,” said Adam Gleave […]
The winning strategy revealed by the software “is not completely trivial but it’s not super-difficult” for a human to learn and could be used by an intermediate-level player to beat the machines, said Pelrine.
Perhaps we could edit the lobste.rs title to something like, “Amateur demolishes top Go AI with simple tactic discovered by computer program”. Maybe there’s a more elegant way to say it.
Perhaps we could edit the lobste.rs title to something like, “Amateur demolishes top Go AI with simple tactic discovered by computer program”. Maybe there’s a more elegant way to say it.
“AI discovered way to beat another AI. In order to remain hidden, it trained it’s human minion to do it”
No. We’re talking about the Go strategy, which was ‘devised’ by the machine, in both cases (that’s literally the point of AI).
Yes, AI was devised by humans, but I don’t think any person who read about the original victory was confused about whether a non-human might have devised AI, before or after they read any articles or headlines.
The noteworthy thing isn’t that humans beat AI or that an AI devised the method, it’s that AlphaGo had a weakness so glaring even a human could exploit it. It’s unlikely that we’d find such a weakness in, say, a “traditional” chess AI which uses opening books and alpha-beta pruning.
The noteworthy thing isn’t that humans beat AI or that an AI devised the method, it’s that AlphaGo had a weakness so glaring even a human could exploit it.
I’m not sure what the meaningful distinction is. That’s what I was trying to get at. It’s interesting that an amateur human was able to reliably beat a top Go AI (14 out of 15 times) with a relatively simple tactic.
These facts are the very same ones I was highlighting in my initial comment, and I was highlighting them because it shows how glaring the weakness was.
But then some commenters made some points responding to my side point about how these facts probably should have been called out in the headline and I was responding to that.
I agree with you, this is interesting, because, to go even further, it shows that, with some creativity and time, humans could probably still find some low-hanging fruit to advance certain games where we previously thought AI had picked all of those.
I think the paper’s title buries the lede a bit - “AI beats other AI” isn’t newsworthy, “there’s a human-viable strategy to beat AI (that happened to be found by another AI)” is an interesting discovery about the limits of superhuman-seeming AI
I’m not suggesting the paper authors should retitle their paper, we’re talking about this in the context of a lobste.rs headline.
I’m fine with having limited information on AI. I’m not an AI researcher nor do I know what an “adversarial policy” is in the context of AI. This site is full of (technical) laypeople like myself who don’t care for every AI paper, but definitely care about significant advancements in AI understanding, such as a human beating a top AI at Go, and doubly so if the human is an amateur.
This is a technical community which aggregates links around certain technical topics that they find interesting and think others will, too. We don’t just post every paper that is technical here, even though you could do that if you wanted to.
If you’re an AI researcher, or some other person who needs “un”limited information for your purposes, fine, you can click on the uninteresting links as well as the interesting links, what I’m suggesting has no impact on you. I’m sure there are even other sites that aggregate papers that aren’t about how interesting the findings are.
You can be here as well, I’m not saying to go away. But for this site, there is definitely a purpose in making a title compelling. And I’m not saying make every title compelling; if there’s nothing interesting (for a particular definition of interesting) on the other side of a link, no need to trick people into thinking there is.
There are many reasons a link might be worth a click to different audiences without being ‘interesting’. However, if you’re posting on lobste.rs, your audience is generally a lot of laypeople, and for them, if there’s something interesting about what the link points to, I wouldn’t hide it.
And none of what I’m saying goes against making titles informative, in fact, I’m advocating for not burying the lede, a lede which is not just interesting, but actually hella informational; the fact that the human who beat the AI is an amateur, even if the strategy was derived by a computer.
I always thought Gary Kasparov deserved a rematch with Deep Blue.
IIRC he claimed he could have won a rematch. It is possible he learned enough about this opponent in his first encounter to defeat it.
Here’s an interesting detail that I feel should have been hinted at in the headline, and would have made me far more likely to click without being clickbait, and while leaving some interesting meat in the body that would have made the read worthwhile in my opinion:
Perhaps we could edit the lobste.rs title to something like, “Amateur demolishes top Go AI with simple tactic discovered by computer program”. Maybe there’s a more elegant way to say it.
“AI discovered way to beat another AI. In order to remain hidden, it trained it’s human minion to do it”
Would’ve you preferred the original AlphaGO victory described as “A computer beats human at Go with a method devised by another human”?
No. We’re talking about the Go strategy, which was ‘devised’ by the machine, in both cases (that’s literally the point of AI).
Yes, AI was devised by humans, but I don’t think any person who read about the original victory was confused about whether a non-human might have devised AI, before or after they read any articles or headlines.
The noteworthy thing isn’t that humans beat AI or that an AI devised the method, it’s that AlphaGo had a weakness so glaring even a human could exploit it. It’s unlikely that we’d find such a weakness in, say, a “traditional” chess AI which uses opening books and alpha-beta pruning.
I’m not sure what the meaningful distinction is. That’s what I was trying to get at. It’s interesting that an amateur human was able to reliably beat a top Go AI (14 out of 15 times) with a relatively simple tactic.
These facts are the very same ones I was highlighting in my initial comment, and I was highlighting them because it shows how glaring the weakness was.
But then some commenters made some points responding to my side point about how these facts probably should have been called out in the headline and I was responding to that.
I agree with you, this is interesting, because, to go even further, it shows that, with some creativity and time, humans could probably still find some low-hanging fruit to advance certain games where we previously thought AI had picked all of those.
It’s the policy of lobste.rs to roll with the original title:
In this particular case, a much nicer solution would have been to post the paper, rather than the article. Paper’s title is on point:
I think the paper’s title buries the lede a bit - “AI beats other AI” isn’t newsworthy, “there’s a human-viable strategy to beat AI (that happened to be found by another AI)” is an interesting discovery about the limits of superhuman-seeming AI
That paper title would almost certainly have not gotten me to click on it. It’s even worse than the current article title, even.
I can only speak for myself, of course.
The paper’s goal is to inform, not to compel your click. If you require your information sources to be compelling, you will have limited information.
I’m not suggesting the paper authors should retitle their paper, we’re talking about this in the context of a lobste.rs headline.
I’m fine with having limited information on AI. I’m not an AI researcher nor do I know what an “adversarial policy” is in the context of AI. This site is full of (technical) laypeople like myself who don’t care for every AI paper, but definitely care about significant advancements in AI understanding, such as a human beating a top AI at Go, and doubly so if the human is an amateur.
This is a technical community which aggregates links around certain technical topics that they find interesting and think others will, too. We don’t just post every paper that is technical here, even though you could do that if you wanted to.
If you’re an AI researcher, or some other person who needs “un”limited information for your purposes, fine, you can click on the uninteresting links as well as the interesting links, what I’m suggesting has no impact on you. I’m sure there are even other sites that aggregate papers that aren’t about how interesting the findings are.
You can be here as well, I’m not saying to go away. But for this site, there is definitely a purpose in making a title compelling. And I’m not saying make every title compelling; if there’s nothing interesting (for a particular definition of interesting) on the other side of a link, no need to trick people into thinking there is.
There are many reasons a link might be worth a click to different audiences without being ‘interesting’. However, if you’re posting on lobste.rs, your audience is generally a lot of laypeople, and for them, if there’s something interesting about what the link points to, I wouldn’t hide it.
And none of what I’m saying goes against making titles informative, in fact, I’m advocating for not burying the lede, a lede which is not just interesting, but actually hella informational; the fact that the human who beat the AI is an amateur, even if the strategy was derived by a computer.
Link to paper: https://goattack.far.ai/
“humans invent a machine to help them defeat a machine they invented”
I always thought Gary Kasparov deserved a rematch with Deep Blue. IIRC he claimed he could have won a rematch. It is possible he learned enough about this opponent in his first encounter to defeat it.