Elon Musk submitted a comment to edge.org a day or so ago, on this article. It was later removed.

The pace of progress in artificial intelligence (I'm not referring to narrow AI) is incredibly fast. Unless you have direct exposure to groups like Deepmind, you have no idea how fast-it is growing at a pace close to exponential. The risk of something seriously dangerous happening is in the five year timeframe. 10 years at most. This is not a case of crying wolf about something I don't understand.

I am not alone in thinking we should be worried. The leading AI companies have taken great steps to ensure safety. The recognize the danger, but believe that they can shape and control the digital superintelligences and prevent bad ones from escaping into the Internet. That remains to be seen...

Now Elon has been making noises about AI safety lately in general, including for example mentioning Bostrom's Superintelligence on twitter. But this is the first time that I know of that he's come up with his own predictions of the timeframes involved, and I think his are rather quite soon compared to most. 

The risk of something seriously dangerous happening is in the five year timeframe. 10 years at most.

We can compare this to MIRI's post in May this year, When Will AI Be Created, which illustrates that it seems reasonable to think of AI as being further away, but also that there is a lot of uncertainty on the issue.

Of course, "something seriously dangerous" might not refer to full blown superintelligent uFAI - there's plenty of space for disasters of magnitude in between the range of the 2010 flash crash and clippy turning the universe into paperclips to occur.

In any case, it's true that Musk has more "direct exposure" to those on the frontier of AGI research than your average person, and it's also true that he has an audience, so I think there is some interest to be found in his comments here.


New Comment
72 comments, sorted by Click to highlight new comments since:
Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

I wonder what would have been Musk's reaction had he witnessed Eurisko winning the United States Traveller TCS national championship in 1981 and 1982. Or if he had witnessed Schmidhuber's universal search algorithm solving Towers of Hanoi on a desktop computer in 2005.

I distinctly recall reading SIAI documents from ~2000 claiming they had until between 2005 and 2010...
Also, in a 2002 interview, Eliezer said that "a few years back" before the interview his actual guess at when the singularity would occur was between 2008 and 2015, but he would say that it would occur between 2005 and 2020 in order to give a conservative estimate.
Eliezer's "The Plan to Singularity" and "Staring into the Singularity" (last updated in 2000 and 2001, respectively) contain numerous references to passive singularity prediction dates and interventionist singularity target dates.

This is not Musk's field of expertise. I do not give his words special weight.

The fact that he can sit in on some cutting edge tech demos, or even chat with CEOs, still doesn't make him an expert.

I have a technical background in AI; there's still massive hurdles to overcome; not 5-10 year hurdles. Nothing from Deepmind will "escape onto the internet" any time soon. It is very much grounded in the "Narrow AI" technologies like machine learning.

I feel pretty confident calling him a Cassandra.


I feel pretty confident calling him a Cassandra.

I agree with the rest of your comment, but calling him a "Cassandra" means "He's right, but no-one will believe him," and I hope that isn't what you meant!

An applicable morality tale here would be the boy that cried wolf, if he hadn't retracted his post. I don't remember if he had a name. (Elon Musk: Inverse Cassandra.)

Stöffler might have the best name among those who failed to update properly.
Well, his comment was deleted, possibly by him, so we should take that into account - maybe he thought he was being a bit overly Cassandra-like too. The other thing to remember is that Musk's comments reach a slightly different audience to the usual with regards to AI risk. So it's at least somewhat relevant to see the perspective of the one communicating to these people.
I think it would actually be helpful if researchers made more experiments with AGI agents showing what could go wrong and how to deal with such error conditions. I don't think that the "social sciences" approach to that works.
This misses the basic problem: Most of the ways things can go seriously wrong are things that would occur after the AGI is already an AGI and are things where once they've happened one cannot recover. More concretely, what experiment in your view should they be doing?
Are you talking about what he's talking about - "risk of something seriously dangerous happening" - or are you talking about AGI? Because I can easily imagine how a narrow AI technology could do a lot of damage, particularly if humans intend it to.
Well, in terms of out-of-control software produced by an AI company, I feel the two risks, 'something dangerous' and AGI, are pretty closely linked. Could more limited AI tech make a more damaging computer virus or cause an unexpected confidential data leak? Sure, but that's not the issue at hand. The most advanced AI today takes input and creates output. It is strictly Oracle AI with nothing present in its architecture that could circumvent that. I don't see that changing anytime soon.
You're free to disregard those, but I'm not sure Elon Musk is doing that. The more damaging computer virus or data leak are only two of the possible worries. If a narrow AI simply helps black market chemists find more novel psychoactives than regulation can ever hope to handle, or if bots eliminate just 10% of jobs (say in transportation and retail to name just the most obvious) leading to massive societal unrest, or if they get better at solving captchas than humans are (which would lead to a massive crisis in anonymous communication and everything that depends on it)... all of these would make Musks prediction true in my book.
But these are just technological issues comparable to other mundane ones; just like how 3D printing could make it easy to create weapons, or how the rise of the automobile has created an enormous new cause of death and injury. There's not reason to think it would be outside the scope of ordinary policy-making methods to handle them. Also, Solving Captchas is already pretty damn easy. A combination of algorithmic methods and crowdsourcing makes it quite cheap, especially for sites using older/easier captcha versions. Captcha is not a security plan; it's a speedbump that's getting easier to pass all the time (but still, no crisis will result from this).
You seem to be much confused :-)
Cleared up my grammar - was that the symptom of the perceived confusion, or do you doubt that much depends on anonymous communication?
How would breaking captchas break anonymous communications?
Some powerful agents (say secret services or the government of... let's say China) would benefit greatly from disrupting anonymous electronic communication as a whole, because that'd force electronic communication to occur in a non-anonymous fashion. People could still encrypt, but it'd at least be known who talked to who, and that's the kind of information that's apparently valued worth billions of dollars and a couple of civil rights. Correct? But how could you do that? Thoroughly anonymized peer-to-peer networks built to defy surveillance (such as Freenet), appear to successfully make de-anonymizing communication very, very, very hard. If you kill or severely impede less than perfect anonymization services such as Tor, anonymity-liking people can just migrate to services such as Freenet, and your plan to disrupt anonymous electronic communication has backfired. Correct? But what you can do is attack not the anonymity, but the communication inside that anonymity. All you need to do is flood the anonymous medium with disruptive pseudo-communication. Spam is the obvious example, but (especially if there are web-of-trust-like structures between the anonymization and the actual communication) you can't make your bots too easy to identify - but as far as it is still possible, you can simply throw in more and more bots. How do you identify bots as such? You do Turing tests of course. How do you identify lots and lots of bots as such? You do completely automated Turing tests, or Captchas. Not necessarily the ones we have, which are apparently somewhat solvable with the current state of machine learning, but better ones. Captchas have already improved, because they had to. Surely there can be better ones, or sites can start to require perfect performance on ten different Captchas at once for acceptance as a non-bot, or charge (even anonymously, using something like bitcoin) for the privilege of getting to take the Captcha. But once you get to the level where narrow AIs
Your argument is basically that anonymous networks can be spammed into uselessness. That looks to be theoretically possible but practically difficult, but that's not the main problem with your argument. The biggest hole, from my point of view, is that you think that captchas are a good (or even only) anti-spam measure. They are not. And, of course, email is a pseudonymous P2P network which used to have a large spam problem and which, by now, has largely solved it. Here is good write-up of how spam wars work in real life.
Spam wars in real life use mechanisms that don't work in fully anonymous networks like Freenet. You can't filter by IP in a network without IPs. Captchas are obviously not a good (or even only) anti-spam measure. But inside anonymous networks, they're one of the few things that work. Webs of Trust, which I explicitly mentioned, are another - they just don't scale well.
DeepMind is very definitely AGI in the sense of the domain of problems its learners can learn and its agents can solve. If DeepMind is easily controlled and not very dangerous, that's not evidence for AGI being further away than we thought before we looked at DeepMind, it's evidence for AGI being more easily controlled than we thought before we looked at DeepMind. Real AGI was never going to look like magic genies, so we should never fault real-life AI work for failing at genie.

Fascinating. Maybe he's been talking to Shane Legg of DeepMind, who also has much sooner timelines than I do.

Do you mind revealing what Shane's timelines are, and the probability that he thinks that he'll play a role in AGI?

How do you know the comment was actually from Musk? My guess is that some crank on the internet sent Edge an email claiming to be from Musk, and they published it without doing an identity check. Then the real Musk found out and asked for it to be taken down. The sentence "10 years at most" seems especially stupidly overconfident and inarticulate (and thus unlike something Musk would write).

Just a quick update of a sort, according to this article, the comment was genuine, and that he will write something longer on the topic.

You know, I suppose it's possible. He does have his own bio on the site though. And the "reality club" thing where famous people comment that edge does is part of their content - I would expect them to be open about it if they had made that kind of mistake and apologise for it, but perhaps that's hoping for too much. How likely do you think it is that it wasn't the real Musk?
Hmmm... This does seem like the most plausible explanation for why the comment was removed - I don't see why Musk would retract his own statement otherwise.

The exposure of the general public to the concept of AI risk probably increased exponentially a few days ago, when Stephen Colbert mentioned Musk's warnings and satirized them. (Unrelatedly but also of potential interest to some LWers, Terry Tao was the guest of the evening).

Warning: segment contains Colbert's version of the basilisk.
"We're sorry but this video is not available in your country." We'll I guess I'm safe. Living in a shitty country has some advantages.
"We're sorry but this video is not available in your country." We'll I guess I'm safe :-).

So what is actually going on at Deepmind right now? Should I be updating on this - is there new data in his estimate (i.e. something going on at deepmind that is more worrying than what we know from other sources)?

Neural Turing Machines are quite interesting, as is their continued work on deep reinforcement learning.

The prediction was wrong, happily!

Musk now says he'll "be surprised if we don't have AGI" by 2029. But this seems significantly less compelling given that his last attempt to time AGI (or "seriously dangerous" AGI-related things) failed.

The "10 years at most" part of the prediction is still open, to be fair.

Kudos to you or whoever saved that comment into an image before it was deleted.

Did you see it on the site, though, or did you only see the image? Because I could easily photoshop such an image and claim it is a legit comment that just happened to be deleted...


Why would Elon Musk have direct exposure to Deepmind?


Ok, he is an investor. I had missed that.

The mainstream press has now picked up on Musk's recent statement. See e.g. this Daily Mail article: 'Elon Musk claims robots could kill us all in FIVE YEARS in his latest internet post…'

This article apparently explains the deletion - it wasn't meant to be a comment for the website. I hope the article is accurate and Musk soon writes something longer explaining his viewpoint.

I suspect that the marginal value of a dollar to Elon Musk is close to zero, which makes it difficult to test his sincerity in his beliefs by offering a bet.

I would structure it like this: I give him $100 right now, and if there's no AGI in 10 years, he gives me a squillion dollars, or some similarly large amount that reflects his confidence in his prediction. This way, he cannot claim that a fooming AI that renders dollars worthless will deny him the benefit of a win, because he gets to enjoy my $100 right now.

Elon is unlikely to accept this wager; would anyone like to accept it in his place?

Note that Stuart Russell has now submitted a comment. It begins with this quote from Leo Szilard:

We switched everything off and went home. That night, there was very little doubt in my mind that the world was headed for grief.


it is growing at a pace close to exponential.

I wonder how he (or anybody else) measures growth of knowledge. Are there any sensible metrics beside amount of paper created? I understand that published pages is a measure as is number of patents but I don't think these are useful proxies for knowledge.

What other measures might be used?

  • Complexity measures of the created knowledge: Depth of the gratph of citations between papers (assuming each citation adds something; might be weithed by the number of outgoing refs)

  • Complexity of the created artifacts (p

... (read more)

We believe we can achieve trans-sapient performance by 2018, he is not that off the mark. But dangers as such, those are highly over-blown, exaggerated, pseudo-scientific fears, as always.

What does "trans-sapient performance" mean?
Well, achieving better than human performance on a sufficiently wide benchmark. Preparing that benchmark is almost as hard as writing the code, it seems. Of course, any such estimates must be taken with a grain of salt, but I think that conceptually solid AGI projects have a significant chance by that time (including OpenCog), although previously I have argued that neuromorphic approaches are likely to succeed by 2030, latest.
You understand that you just replaced some words with others without clarifying anything, right? "Sufficiently wide" doesn't mean anything.
By "we" do you mean Gök Us Sibernetik Ar & Ge in Turkey? How many people work there?
Confidential stuff, it could be an army of 1000 hamsters. :) To be honest, I don't think teams more crowded than 5-6 are good for this kind of work. But please note that we are doing absolutely nothing that is dangerous in the slightest. It is a tool, not even an agent. Although I will be working on an AGI agent code as soon as we finish the next version of the "kernel", to demonstrate how well our code can be applied to robotics problems. Demo or die.
What did you think of Bostrom's recent book?
I didn't read it, but I heard that Elon Musk is badly influenced by it. I know of his papers prior to the book, and I've taken a look at the content, I know the material being discussed. I think he is vastly exaggerating the risks from AI technology. AI technology will be as pervasive as the internet, it is a very spook/military like mindset to believe that it will only be owned by a few powerful entities, who will wield it to dominate the world, or the developers will be so extremely ignorant that they will have AI agents escaping their labs and start killing people. Those are merely bad science fiction scenarios, like they have on Hollywood movies, it's not even good science fiction, because he is talking about very improbable events. An engineer who can build an AI smarter than himself probably isn't that stupid or reckless. Terminator/Matrix scenarios won't happen; they will remain in the movies. Moreover, as a startup person, I think he doesn't understand the computer industry well, and fails to see the realistic (not comic book) applications of AI technology. AGI researchers must certainly do a better job at revealing the future applications. That will both help them find better funding and attracting public attention, and of course, obtaining public approval. Thus, let me state it. AI really is the next big thing (after wearable/VR/3dprinting, stuff that's already taking off, I would predict). It's right now like a few years before the Mosaic browser showed up. I think that in AI there will be something for everybody, just like the internet. And Bostrom's fears are completely irrational and unfounded, it seems to me. People should cheer up if they think they can have the first true AI in just 5 years.
+1 for entertainment value. EDIT: I am not agreeing with examachine's comment, I just think it's hilariously bad.
What would you be willing to make a bet that nothing remotely resembling that happens before 2020? 2025?