hgm wrote:M$50 at an interest rate of 2%/yr, for 8 hours ~ 1milli-year is $1000. So yes, I would say that for a company like Google this is just a joke.
Again you are comparing apples and orangutans. What the hell does interest rate have to do with it???
Google will not rent TPU in cloud services for less than a 1$/h.
5'000TPUs x (9h+12h+34h) = 275000k$ of lost revenue just for self-play games in the paper. So you are at least 2 and a half orders of magnitude wrong.
I wonder if you make such "great" estimations also in your main field of work?
Come on, HGM is surely right in his arguments in this thread. This arXiv paper of DeepMind authors is completely valid scientific paper for getting priority. They can elaborate later, or even not at all, depending on the say of "Nature" peer-reviewers. To get priority, I had some even sketchier valid papers. Long ago I lost priority by wasting time with putting together a LaTeX crap (at which I am very bad) just to see an arXiv sketch by others giving me a blood. I just abandoned that my work.
Second issue is that against a human Lee Sedol they indeed used monster hardware, and nobody complained. Even an adapted Giraffe would beat a top GM on that sort of hardware. Why complain now?
hgm wrote:Of course the hope is that this form of AI could be used in the future towards solving real-life problems, rather than just silly games. But not many real-life problems are equivalent to two-player zero-sum games with complete information, so this is rather a long shot.
I don't think it's a long shot at all. I could see it being an excellent automobile driver, for instance, although probably overkill for that. We'll just have to wait and see what they come up with.
hgm wrote:M$50 at an interest rate of 2%/yr, for 8 hours ~ 1milli-year is $1000. So yes, I would say that for a company like Google this is just a joke.
Again you are comparing apples and orangutans. What the hell does interest rate have to do with it???
Google will not rent TPU in cloud services for less than a 1$/h.
5'000TPUs x (9h+12h+34h) = 275000k$ of lost revenue just for self-play games in the paper. So you are at least 2 and a half orders of magnitude wrong.
I wonder if you make such "great" estimations also in your main field of work?
Come on, HGM is surely right in his arguments in this thread. This arXiv paper of DeepMind authors is completely valid scientific paper for getting priority. They can elaborate later, or even not at all, depending on the say of "Nature" peer-reviewers. To get priority, I had some even sketchier valid papers. Long ago I lost priority by wasting time with putting together a LaTeX crap (at which I am very bad) just to see an arXiv sketch by others giving me a blood. I just abandoned that my work.
Second issue is that against a human Lee Sedol they indeed used monster hardware, and nobody complained. Even an adapted Giraffe would beat a top GM on that sort of hardware. Why complain now?
By the way, I remember my father having hand-written preprints in physics rushed before the secretary was able to print a proper preprint, in late 1960s The competitivity was always high, and heck, nowadays we are typists too with this LaTeX and such crap.
Last edited by Laskos on Fri Dec 15, 2017 12:55 pm, edited 1 time in total.
Laskos wrote:Come on, HGM is surely right in his arguments in this thread. This arXiv paper of DeepMind authors is completely valid scientific paper for getting priority. They can elaborate later, or even not at all, depending on the say "Nature" peer-reviewers. To get priority, I had some even sketchier valid papers. Long ago I lost priority by wasting time with putting together a LaTeX crap (at which I am very bad) just to see an arXiv sketch by others giving me a blood. I just abandoned that my work.
With publishing in "Nature" type of publications it is how you do it, you send a sketch and later work on details.
But you forget 2 thing.
1) They don't need to publish on arXiv in such an essentially crappy form (and no that's certainly not a scientific paper) especially not during London Chess Classic to get priority. They almost certainly submitted the full paper to respected journal or conference before (and I strongly doubt it will be rejected, for god sake they are DeepMind).
2) The resources they used, essentially prohibits literary anyone to overtake them, so that claim about priority in this particular case makes no sense.
Laskos wrote:Come on, HGM is surely right in his arguments in this thread. This arXiv paper of DeepMind authors is completely valid scientific paper for getting priority. They can elaborate later, or even not at all, depending on the say "Nature" peer-reviewers. To get priority, I had some even sketchier valid papers. Long ago I lost priority by wasting time with putting together a LaTeX crap (at which I am very bad) just to see an arXiv sketch by others giving me a blood. I just abandoned that my work.
With publishing in "Nature" type of publications it is how you do it, you send a sketch and later work on details.
But you forget 2 thing.
1) They don't need to publish on arXiv in such an essentially crappy form (and no that's certainly not a scientific paper) especially not during London Chess Classic to get priority. They almost certainly submitted the full paper to respected journal or conference before (and I strongly doubt it will be rejected, for god sake they are DeepMind).
2) The resources they used, essentially prohibits literary anyone to overtake them, so that claim about priority in this particular case makes no sense.
Wasn't a guy from SF devs (forgot his name) saying he is working on a similar thing? Maybe I just misunderstood. With resources, well, even a much weaker pre-A0 would harm a purely A0 "revolution".
Dirt wrote:I don't think it's a long shot at all. I could see it being an excellent automobile driver, for instance, although probably overkill for that. We'll just have to wait and see what they come up with.
AI is already much better in driving than humans, but there are 2 problems there.
1) Unexpected situations. No matter how many hours you've spent training your AI, there will always be many real-life situations where it hasn't been trained (human either), and where it might perform poor compared to humans. When these situations are life and death ones, you have a really big problem.
2) Reliability of the sensors. The AI can be perfect, but can't work if its inputs are bad. Aviation industry still has huge problem with that even after 50+ years of developing auto-pilots. So if sensor fails AI disconnects and if human is not able to instantly takes control, consequences could be grave.
That is why it is so hard to make Level 4 self-driving a reality despite huge current effort.
Laskos wrote:Wasn't a guy from SF devs (forgot his name) saying he is working on a similar thing? Maybe I just misunderstood. With resources, well, even a much weaker pre-A0 would harm a purely A0 "revolution".
Don't know about SF dev, but many ppl are working on reinforcement learning DCNN training and MCTS type of chess engine. Problem is the current performance of those efforts. I don't think there is any that can reach even super GM performance, i.e. even 500Elo below current SF.
With that you don't even get publication in ICGA, or mediocre AI journal, and submitting to "Nature" type of publication would be a laugh.
Laskos wrote:
This arXiv paper of DeepMind authors is completely valid scientific paper for getting priority. They can elaborate later, or even not at all,
Priority before whom?
The only reason to publicise such a "paper" is to see the reaction of chess- world and scientists like you.
I'd bet they won't elaborate anything more later, as you say, even not at all, why should they?
Chessplayers not knowing what had really happened are enthusiastic and scientists like you are enthusiastic too by just having been shown such a cheap (having invested much more before in Go) and biased test, and to convince AI- scientists knowing anything about chess and AI is not a goal for Google at all.
Those scientists working in the same field as DeepMind aren't an audience to convince and impress for Google.
What to earn from competitors?
hgm wrote:M$50 at an interest rate of 2%/yr, for 8 hours ~ 1milli-year is $1000. So yes, I would say that for a company like Google this is just a joke.
Again you are comparing apples and orangutans. What the hell does interest rate have to do with it???
That is how one converts capital investment to cost of use. The M$50 equipment cost was your number.
Google will not rent TPU in cloud services for less than a 1$/h.
5'000TPUs x (9h+12h+34h) = 275000k$ of lost revenue just for self-play games in the paper. So you are at least 2 and a half orders of magnitude wrong.
I wonder if you make such "great" estimations also in your main field of work?
Ummm... The way I make estimations in my field uses more conventional arithmetic, where 5'000TPUs x (9h+12h+34h) x 1$/h = 275k$. Rather than 275000k$. And note that I was just addressing the 9hr Chess part.
