back to article That this AI can simulate universes in 30ms is not the scary part. It's that its creators don't know why it works so well

Neural networks can build 3D simulations of the universe in milliseconds, compared to days or weeks when using traditional supercomputing methods, according to new research. To study how stuff interacts in space, scientists typically build computational models to simulate the cosmos. One simulation approach – known as N-body …

  1. Blockchain commentard Silver badge

    If neural networks are so good at simulations, can someone do one to try and explain Donald Trump?

    1. Evil Auditor Silver badge

      Not that I'm qualified to answer but within this context that should be perfectly acceptable. Trump, I think it is fair to assume, contains some form of neural network. And from my limited experience, that neural network behaves scarily similar to one artificial neural network that I built about twenty years ago.

      It was a simple one, very trivial: input/output field was something like an six-by-six dot matrix and it was trained to recognise single-digit numbers 0 to 9. It did reasonably well for an experiment. But it had one major flaw: whatever the input was (maybe even for a blank matrix), it would always "recognise" a number. So, from a rather high number of stimuli, the output was limited to a very small number of possible reactions - but it would always and instantly react.

      1. KittenHuffer

        Wonderful comment, well written!

        Very well done. Your piece is successfully 'whooshing' those that would normally downvote such a comment. This can be seen by the fact that your comment is currently 8-1 whereas the parent is 5-3.

        Being able to insult someone or something without them even realising seems to have become a bit of a lost art these days. I congratulate you on reviving this, and also for your choice of target.

        1. DropBear Silver badge

          Re: Wonderful comment, well written!

          Relative rarity of the requisite skills aside, the thing is that type of insult is meant for the benefit of others of comparable stature who witness it, while being completely lost on the target. It only works when those who see it are the majority, by lowering their esteem of the victim if the scorn is deserved. When those who see it are in the minority or completely absent however, wits are supremely ineffective and do none of what an insult is supposed to achieve. So it's hardly surprising it's not a popular pastime these days, and that whole "never offend people with style when you can offend them with substance" thing...

        2. asdf Silver badge

          Re: Wonderful comment, well written!

          >Being able to insult someone or something without them even realising

          Like saying with all due respect to most Yanks (am one).

          1. Robert Forsyth

            Re: With all due respect

            I'm not sure it needs explaining.

            The amount of respect deserved based on their past actions or statements, which might be none due.

            One of the sentences which sounds complementary to pompous people with inflated self worth.

          2. Martin-73 Silver badge

            Re: Wonderful comment, well written!

            I think you're being unfair to your countrycritters by saying 'most'. Most of my American friends are simply silent through shock. Much like most brits

    2. fandom

      Because why talk about boring things like science when you can put political shit in everything.

      1. Anonymous Coward
        Anonymous Coward

        News flash, everything is political. Either you can participate, or be a victim of other people's choices.

        1. Anonymous Coward
          Anonymous Coward

          Real Flash

          If you don't bring it in it's not part of the equation.

          You can talk science without religion and Politics.

          You can talk without bring in Race as well, well maybe not anymore since Barry though.

        2. Anonymous Coward
          Anonymous Coward

          >Either you can participate, or be a victim of other people's choices.

          You're a victim either way - it's a much better use of your time to improve the choices made by others - and that has little to do with politics.

      2. Anonymous Coward
        Anonymous Coward

        RE: you can put political shit in everything

        'snot just for breakfast anymore?

    3. Anonymous Coward
      Anonymous Coward

      That cahnt (I refuse to use his name) is easy to figure out.

      He is very cleverly playing the system and leveraging the stupidity of the people. I don't just mean the voters here, but the press (which publish every 'controversial' thing he says - doesn't matter if it's for or against him, it's publicity and keeps him in the news cycle and his opponents out if it).

      The left are also responsible as they try to engage him with old rules and expectation of debate. These rules don't apply any more. The systems (of politics and media) are falling apart and easily exploitable.

      The blonde cahnt in the UK is using the same tactics. Yesterday almost every major media outlet was reporting bs about painting cardboard busses (including for and against this as an acceptable debate). With this level of stupidity going on in the systems that should be balancing out the corruption and people lapping it up, then no wonder this pair can easily be successful.

      No AI needed to see it.

      1. Benson's Cycle

        His body language on the buses thing, even as he reached for the next bit, clearly said "I can't believe I'm getting away with this guff".

      2. CountCadaver

        Turning "chucking a dead cat on the table" up to 11

        Basically iterating the methodology leveraged by Sir Lynton Crosby : https://en.wikipedia.org/wiki/Lynton_Crosby#Dead_cat_theory

    4. Anonymous Coward
      Anonymous Coward

      Dont need a neural net to aswer that...

      The reason for President Trump is..

      ...Hillary Clinton

      Replace that term with any other name and first statement become null.

      For the same reason the answer to the phrase President Regan (I) is..

      .. Jimmy Carter and Jon Anderson

      And for President Regan (II) is..

      ..Walter Modale

      Now the phrase President Bill Clinton(I) and (II) is a bit more interesting as the answer to both is..

      ..Ross Perot.

      That's politics for you. The people who decide the outcome are usually those who vote against someone rather than vote for them.

      As for the Universe, I'm afraid it was all a ghastly mistake.

    5. Anonymous Coward
      Anonymous Coward

      The Donald?

      Mmm.

      Left wing liberal: Monkey say.

      Donald: Monkey do.

      The confusion arises in the Left wing brain because it is conditioned to listen to words and believe in them. Especially when uttered by celebrities.

      The Donald is programmed to challenge this behaviour by:

      (a) Being a celebrity

      (b) Uttering nonsense.

      Curiously what the Donald does appears to be making America great again.

      This really annoys people who want to see a world government and the end of democracy.

  2. Zog_but_not_the_first Silver badge
    Alien

    Dark Star

    "Let there be light!

    1. Screwed

      Re: Dark Star

      Commander Powell for PM.

    2. Robert D Bank
      Pint

      Re: Dark Star

      Dark Star Revelation, lurvely!

  3. Anonymous South African Coward Silver badge

    Skynet Awakens.

    1. 0laf Silver badge
      Pint

      Beat me to it. Here is your pint ->

  4. Thoguht Silver badge

    Infinite Fun Space here we come!

    1. Roland6 Silver badge

      Where is the fairy cake?

      1. GX5000
        Megaphone

        Still?

        The Cake is a lie....c'mon now.

        1. Anonymous Coward
          Anonymous Coward

          Re: Still?

          Butt here are cookies. You promised cookies.

      2. Ugotta B. Kiddingme Silver badge

        Re: Where is the fairy cake?

        Also no mention of the requisite "You are here" sign.

  5. Semtex451 Silver badge
    Joke

    Could someone please apply this to the weather because I've been promised lightning storms for the past 6 weeks and there has not been a single one.

    1. Paul Crawford Silver badge

      You can look for where they are hiding here:

      http://www.lightningmaps.org/blitzortung/europe/index.php?lang=en

    2. hplasm Silver badge
      Boffin

      "Could someone please apply this to the weather because I've been promised lightning storms for the past 6 weeks and there has not been a single one."

      ...and the , er, experiment on the slab is starting to smell...

      1. KittenHuffer

        I believe that Igor has connected the lightning rod with the polarity reversed .... and it's actively repelling the lightning!

  6. Caspian Prince

    Obligatory HHT2TG quote

    “And to this end they built themselves a stupendous super-computer which was so amazingly intelligent that even before its data banks had been connected up it had started from I think therefore I am and got as far as deducing the existence of rice pudding and income tax before anyone managed to turn it off.”

  7. Authentic Name

    Simulation

    Clearly it is using the same algorithm as the simulation we exist in.

  8. The Oncoming Scorn Silver badge
    Thumb Up

    Probably Involves A Fairy Cake.....

    ASSISTANT ARCTURAN PILOT:

    Well you know what they say, don’t ya? They had to move to a bigger planet because he got so fat he kept sliding off the old one. I mean I’ve heard ya know, I’ve heard they’ve created a whole electronically synthesized universe in one of their offices so they can go and research stories during the day and still go to parties in the evening. Yeah, bloody clever, of course, but it’s got nothing to do with the real galaxy is it? Nothing to do with life.

    Icon for HHT2TG.

    1. Louis Schreurs

      ?.....

      HHG2TG

      ?????

      HHGTTG........

  9. steelpillow Silver badge
    Joke

    The elephant in the room

    So, Tubul, Jerakeen, Berilia and Great T'Phon turn out to be logical consequences of a universe built from cats and dogs.

    But can the net recognise the Great A'Tuin?

  10. Def Silver badge
    Headmaster

    Class, repeat after me:

    ML is not AI.

    1. Azerty

      Re: Class, repeat after me:

      it's not GAI, but it's a narrow form of AI, AI nevertheless.

      1. William1940

        Re: Class, repeat after me:

        Actually, AI is not AI. Rather, it is Machine Based Intelligence. Nothing artificial about it at all. John McCarthy should've known better.

  11. RichardBarrell

    The link to the paper is broken

    The link to the paper in the article goes to https://www.pnas.org/content/early/2019/06/21/18214581162 which is a HTTP 404.

    I think you wanted to link to https://www.pnas.org/content/early/2019/06/21/1821458116 instead.

    1. hplasm Silver badge
      Happy

      Re: The link to the paper is broken

      It's the little perforations!

      1. Anonymous Coward
        Anonymous Coward

        Re: The link to the paper is broken

        "It's the little perforations!"

        Very old joke:

        Designers discussing the mystery of why the prototype aeroplane's wings were breaking off in the wind tunnel tests. Passing cleaner says "Drill a line of holes across the wing where it would break".

        They decided to try it - and it worked! Thanking the cleaner they asked how they knew that answer - "Easy - toilet paper never tears along the perforation".

  12. Anonymous Coward
    Anonymous Coward

    Rain man?

  13. Milton Silver badge

    Skeletons

    It's like teaching image recognition software with lots of pictures of cats and dogs, but then it's able to recognize elephants," said Shirley Ho, first author of the paper and a group leader at the Flatiron Institute. "Nobody knows how it does this, and it's a great mystery to be solved.

    I must be missing something here, because the answer would seem obvious: the neural net is inferring the existence of skeletons. In the analogy provided, you are giving it lots of pictures of cats and dogs, and it "notices" that some bits are rigid and sized according to specific ratios, that others are "bendy" points, also located according to certain ratios, and it therefore infers underlying structure and rules governing movement. In the cats'n'dogs case, if you included data on their patterns of movement in specific environments, introducing the concept of behaviours toward goals (e.g. hunting by stealth in environments with abundant cover; scavenging in open regions where carrion/waste may be deposited), my wild-ass guess is that you could get the neural net to infer the approximate form factor for a "new" animal based on previously unimagined environments. From mice to cows. (It is intriguing to speculate what deep parallels there may be with how evolution actually works.)

    Returning to the universes, it would seem that the neural net is inferring the existence of underlying rules despite not knowing their precise formulae. This is surely exactly what you wanted.

    This leads me to ask: at what point can you inspect the neural net's results and ask "Why?" When does it become able to tell its human interlocutors "I have discovered the inverse square law for gravity"?

    Because the neural net's next answer—some time later—may reveal rules/formulae that we didn't already know ...

    1. The Oncoming Scorn Silver badge
      Thumb Up

      Re: Skeletons

      FORD:

      Hey Marvin!

      MARVIN:

      What do you want?

      FORD:

      Give Zaphod a yell will you?

      MARVIN:

      Ahhh. Mind-taxing time again is it?

      FORD:

      Just get on with it.

      MARVIN:

      I’ve just worked out an answer to the square root of minus one.

      FORD:

      Go and get Zaphod.

      MARVIN:

      It’s never been worked out before. It’s always been thought impossible.

      FORD:

      Go and get -

      MARVIN:

      I’m going. Pausing only to reconstruct the whole infrastructure of integral mathematics in his head, he went about his humble task. Never thinking to ask for reward, recognition, or even a moment’s ease from the terrible pain in all the diodes down his left side. “Fetch Beeblebrox,” they say, and forth he goes.

      [Door hums open]

      ARTHUR:

      Don’t you think we should do something for him?

      FORD:

      Hmm… we could rip out his voice-box for a start.

    2. NetBlackOps

      Re: Skeletons

      The next step is to insert "probes" into the net during training to see the shifts that are occurring in the weighting of the net as you train it. They either aren't aware of the technique (fairly new) or just haven't found their roundtoit.

    3. boltar Silver badge

      Re: Skeletons

      Artificial neural networks are just very good statistical analysers. Learn enough stats and you can infer most things, but that doesn't mean the end result will be accurate and I for one would - for example - prefer the auto pilot in any plane I'm flying in to be calculating its output based on hard coded physical laws, not infering the the best response from a bunch of previous flight stats.

      1. Robert Forsyth

        Re: Skeletons

        But isn't that what the fleshy pilot does?

        1. boltar Silver badge

          Re: Skeletons

          That and a whole lot more. The human brain 6 or 7 multiple orders of magnitude more complex than an ANN and can make far more in depth decisions.

          1. Alien8n Silver badge

            Re: Skeletons

            And yet it can still come up with the idea that tying a rope around your ankle and jumping off a bridge is a "good idea".

            1. Anonymous Coward
              Anonymous Coward

              Re: Skeletons

              "[...] tying a rope around your ankle and jumping off a bridge is a "good idea"."

              The rope is a good precaution if the aim is to test your emulation of a bird. It took folk a while to be able to do controlled flight using empirical experiments.

          2. itzman
            Paris Hilton

            Re: Skeletons

            Do you realise that the human brain is the size that it is not to encompass the beauty of Romantic Poetry, but in order to be able to walk on two legs?

            1. boltar Silver badge

              Re: Skeletons

              Hardly. An ostrich can walk on 2 legs and it has a brain the size of an egg.

            2. The Dark Side Of The Mind (TDSOTM)

              Re: Skeletons

              Funny enough, but I think the size is attributable to the complexity of lies that it has to serve to the frontal lobes. Consider only the saccades...

      2. Anonymous Coward
        Anonymous Coward

        Re: Skeletons

        "[...] not infering the the best response from a bunch of previous flight stats."

        For that you need a Probability Drive.

  14. Arthur the cat Silver badge

    containing 32,728 particles

    Strange number. Typo for 32,768 maybe?

    1. Jay Lenovo Silver badge

      Re: containing 32,728 particles

      A universe not based on a power of 2?

      The missing 40 must have been reserved for chaos theory.

      1. Korev Silver badge

        Re: containing 32,728 particles

        32,728 is overkill; the answer is 42...

  15. Anonymous Coward
    Anonymous Coward

    FFS A.I is not intelligent

    A.I as being done now is not fucking intelligence in any shape or fucking form.

    it's a weighted pattern identifier with a funky twist that makes it even harder to know why it's doing shit.

    unless you know why it makes the choices, you don't know shit, and the model knows shit all too, and tells you shit all.

    Brains are pretty good at making quick rough guesses, but are actually pretty bad at accurate processing, we pretend our brains are infalliable but in actual fact are pretty bad, dealing with people everyday I worry how 90% actually manage to do anything useful as there decisions are so shit.

    Why are we trying to make something intelligent by modelling a flawed system seems pretty stupid.

    1. Hollerithevo Silver badge

      Re: FFS A.I is not intelligent

      So you are saying that artifical intelligence is flaws based on the example of flawed human intelligence, which is largely a weighted pattern identifier, and perhaps one not even as good at processing as I? It seems that AI is exactly the right term: an intelligence modelled on the only other intelligence we know well. What else could be model it on? A $Deity's intelligence? A whale's, wot we know not of?

    2. boltar Silver badge

      Re: FFS A.I is not intelligent

      "Why are we trying to make something intelligent by modelling a flawed system seems pretty stupid."

      Biological neural networks might be flawed, but they're good enough to have survived half a billion years on a changing planet where the rules arn't written down.

    3. TomG

      Re: FFS A.I is not intelligent

      Downvoted you, not for what you said but, for the inept way you said it.

  16. Uncle Ron

    Nobody Knows...

    Nobody knows how it reached over and grabbed my sandwich either. Plus, it made a phone call over to Skynet for some reason.

  17. Anonymous Coward
    Anonymous Coward

    Searching for the AI Whisperer

    Someone to calm the machine, speak sweet nothings into its noise cancelling microphones and smile reassuringly into its camera arrays, soothing its overactive networks, and stroking its emergent ego while reaching for the power cord and killing that thing deader than a stuffed dodo.

    Anon natch.

  18. rpark

    Dark matter

    ...Shirley Ho believes Dark Matter is an entirely separate category from the particle simulation (dogs, cats, elephants) leading to her confusion. The particles originate from the Dark Matter and resolve back into the 'Dark Matter' - DM isn't separate from 'regular matter' it originates it.

  19. d3vy Silver badge

    To be fair to them, I dont know how most of my code works most of the time either,

    1. Anonymous Coward
      Anonymous Coward

      I was always impressed when my code handled an unexpected situation for which it had not been explicitly programmed. Explaining how it did it was often a challenge.

  20. spold Bronze badge

    Hmmmm

    >Neural networks can build 3D simulations of the universe in milliseconds

    On no account give them a ray gun

    1. Ugotta B. Kiddingme Silver badge

      Re: On no account give them a ray gun

      certainly not a phased plasma rifle in the 40-watt range...

  21. Will Godfrey Silver badge
    Alien

    Be afraid

    Be very afraid.

  22. TechnicalBen Silver badge
    Thumb Up

    I could tell them...

    Should I email them? Like, mathematically, there is a good reason this (a neutral net/weighted network) should work pretty well. Who do I email/hit up on twitter to explain it to them? :)

  23. StuntMisanthrope Bronze badge

    Pinball wizard. Pocket O'Credits.

    Like a back actor. Can I have a sim please. Let's fire an ever descending serial particle spectral sequence from a lagrange gun, curved forward in time through the offset IDL and bounced back. Like a complicated flag with a planar barcode scanner at asteroids level. #wordsshapesconceptsandgenerationmaynotbeapplicable #magnetslater

  24. Tom 7 Silver badge

    Cutting corners I'd imagine.

    We tend to use some pretty accurate FP maths even when something less computationally involved would do when we model 3d worlds. I'd hazard a guess that the AI has spotted where it doesnt have to piss about with 64bits and can just leave it out for a few cycles.

    I used to work on electrical cct simulators which couldnt be arsed to calculate lots of shit when it wasnt necessary - saves a shitload of CPU not working out fuck all to 20 decimal places.

    1. Anonymous Coward
      Anonymous Coward

      Re: Cutting corners I'd imagine.

      "[...] which couldnt be arsed to calculate lots of shit when it wasnt necessary {...]

      Effects which were conventionally ignored "because they are insignificant" - have been known to bite people in a new context where they had become very significant.

  25. Charles Calthrop

    they should empoly me

    I don't know why my code works so badly. Together we'd be a like a fortune teller who finds a tenner. A happy medium

  26. Julz Bronze badge
    Joke

    Proof

    That the Universe is a simulation.

  27. tfb Silver badge
    Boffin

    Black boxes

    It's not clear from the article which of two things this is doing:

    - finding, by a possibly-opaque mechanism (ie an NN) a good configuration for a simulation which can then be run;

    - or running a simulation by an opaque mechanism which seems to produce answers which look reasonable.

    The first of these is fine I think: it would be interesting to know how it finds the good configuration but that doesn't actually scientifically matter: what matters is that the simulation, when run, can be explained in terms of the physics we know. This is like solving a differential equation by an ansatz: you just say 'let's try this as a solution', plug it into the equation and show that it is indeed a solution. How you arrived at the ansatz is interesting but you are allowed to arrive at it by guessing or magic or, really, by any mechanism you like ('looking it up' is the normal one).

    The second is not fine. If the actual simulation is using some opaque mechanism to produce plausible-looking results then that really tells you nothing useful, until you understand the opaque mechanism and can translate it into physics (possibly new physics). Unless you can do that it's just a black box doing magic.

    It looks to me from the abstract that what they are doing is indeed the latter. Which is probably why they are not completely happy with it.

    (This is like the difference between predicting the weather and predicting the climate: if someone builds an opaque NN model to predict the weather and it demolishes the current numerical models (and there's a crap load of training data, so this is almost certainly going to happen) then that's fine, because the purpose of a weather forecast is to be accurate, and if it achieves that accuracy by opaque magic, well, who cares? But an opaque NN model to predict climate is useless, because the whole purpose of a climate model is to be able to understand what the underlying mechanisms are and how adjusting those mechanisms might alter the trajectory of the climate, for which an NN is useless as it's this opaque blob of weights.)

  28. Toni the terrible
    Angel

    Multivac

    All this assumes that the inputs it is trained upon, and the results we expect are in fact true/correct of course.

    It also reminds me of an old SF story about the Giant AI computer known as Multivac. Once constructed the users wondered what to do with such a powerful AI, so they decide to ask it some of the most difficult questions that have plagued mankind since antiquity, and it answered them. Eventually, they decided to ask it if there was a 'God' and after years of input and decades of 'thinking' Multivac annouced it had an result, the Users gathered and asked the question " Is there a God?" to which the reply was "There Is Now"

    1. Roxor

      Re: Multivac

      Probably the best of Asimov's Multivac stories is The Last Question, where Multivac and its descendants are repeatedly asked how to reverse entropy.

  29. ZeiXi

    Works Well!

    A teenager, in similar fashion, knows everything in a span of 13 years, more than everyone else above the age of 40. I am particularly amazed how good researchers are at patting each other’s back.

  30. Mike 137 Bronze badge

    N bodies

    I'm not a high powered mathematician, but last time I checked it seemed the N body problem was analytically insoluble. We're probably in the realm of extreme sensitivity to initial conditions, which means the outcome is not likely to be reproducible.

    Apart from which - do we actually know what a universe is, and how do we therefore know we've adequately simulated one? Last time I checked, we had only managed to ascribe mechanisms to pretty a small proportion of the real one.

  31. fraunthall

    Very interesting and almost frightening article

    What intrigues me is that if a neural network can do what is described, that is to learn on its own without programmatic guidance to analyze and simulate things so very different from what it started with, and do it so quickly, then it is approaching the level of biological intelligence. I would love to see it being used to analyze and predict the outcomes of human planned activity based on available information about the plans, the risks associated therewith, the reasons why the activity was wanted in the first place, the predelictions or intentions and hoped for outcomes of the originators of the plan, and frailties of the planned concept, and of the planning and implementation processes involved, etc. What a tool that would be.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019