back to article What's with AI boffins strapping GoPros to toddlers? We take a closer look

AI researchers looking for better ways to train large language models are turning to the masters of language acquisition – children – to find out how it's done. Large language models – the complex neural networks behind the generative AI boom – are trained on mountains of data. Yet many of these models are little more than an …

  1. m4r35n357 Silver badge

    Yeah that is bound to work

    Desperation is growing.

    1. b0llchit Silver badge
      Happy

      Re: Yeah that is bound to work

      Think of it this way: once we can use 80 years of data to train an AI, covering an entire human lifetime, we get an eighty year old artificial fart telling us that it was better in the past.

      What is not to like?

      1. Anonymous Coward
        Anonymous Coward

        Re: Yeah that is bound to work

        Bíonn ár bpáistí i gcónaí i bhfad níos cliste agus i bhfad níos cruthaithí ná aon ríomhaire.

        1. b0llchit Silver badge
          Go

          Re: Yeah that is bound to work

          We didn't build Deep Thought and we certainly did not build the planet computer earth with life as integral calculation part.

          (I'm waiting for the Vogons to do their work)

        2. heyrick Silver badge
          Happy

          Re: Yeah that is bound to work

          All I know about Irish is that it isn't said anything even remotely like it looks (to a person used to how the letters work in English).

          Upvoted, because Irish girl's names are lovely.

          1. abend0c4 Silver badge

            Re: Yeah that is bound to work

            it isn't said anything even remotely like it looks

            In contrast to choir, queue, cough, epitome, leopard, colonel, debt, island...

            1. captain veg Silver badge

              Re: Yeah that is bound to work

              Not to mention tough, though, through, thorough, trough. And plough.

              -A.

              1. upsidedowncreature

                Re: Yeah that is bound to work

                Proper names too - Cholmondley, Belvoir, Featherstonehaugh.

                1. Sorry that handle is already taken. Silver badge
                  Trollface

                  Re: Yeah that is bound to work

                  Those are made up names stop being silly!

              2. William Towle
                Pint

                Re: Yeah that is bound to work

                > Not to mention tough, though, through, thorough, trough. And plough.

                *hiccough*

                1. captain veg Silver badge

                  Re: Yeah that is bound to work

                  Good one.

                  Where do you stand on the bruff/brow pronunciation of the vintage motorcycle manufacturer(s) Brough (Superior)?

                  -A.

        3. Anonymous Coward
          Anonymous Coward

          Re: Yeah that is bound to work

          An bhfuil cead agam dul go dtí an leithreas?

          1. heyrick Silver badge
            Happy

            Re: Yeah that is bound to work

            Buíochas le Dia as aistriúchán ar líne.

      2. Anonymous Coward
        Anonymous Coward

        Re: Yeah that is bound to work

        "an eighty year old artificial fart telling us that it was better in the past"

        And running for Prez again...

      3. John Miles

        Re: Yeah that is bound to work

        Speaking of the good old days - I can relate to this

    2. Lee D Silver badge

      Re: Yeah that is bound to work

      "Just throw more data at it."

      Everyone knows that the secret sauce to turn binary signals into actual intelligence is more processing / data / time. We've known that since the 60's right?

      One day we will hit that magical universal data barrier beyond which everything magically turns into intelligence, and the world will be changed forever.

      Because, you know, those supercomputers running for decades across billions of machines on every continent analysing the entire world's data going back 40 years... that's what's inside every cat, dog, monkey, ape, elephant, giraffe, dolphin etc. right?

      They have all that ability, and that's why they're smart, and plants don't which is why they're dumb.

      Just a *little* more funding and OBVIOUSLY we'll hit that magic universal constant this time that'll turn our little sand machines into the world's greatest intelligences.

      1. Sorry that handle is already taken. Silver badge

        Re: Yeah that is bound to work

        Sure, and one day amanfromMars will make sense

      2. captain veg Silver badge

        Re: Yeah that is bound to work

        Optimal strategy for finding a needle in a haystack: add more hay.

        -A.

  2. Bebu
    Windows

    Quelle surprise!

    So intelligent systems acquire language rather than language systems acquiring intelligence.

    Of course (re)defining intelligence as linguistic competence is good for the share price. :)

    Explicit criteria defining language and linguistic competence against purely sonic communication of alarms, etc which would allow one to determine whether apes (chimpanzee, gorilla, orangutan) have language as these creatures clearly have intelligence* (as do many others.)

    If gorilla were determined not to have language but to possess intelligence they would be a clear counter-example to the linguistic competence is intelligence claim.

    The is an embarrassment of examples from the human world demonstrating extraordinary linguistic proficiency totally lacking even a skeric of intelligence.

    * Arguably instances of tool use and tool making. The argument that human language developed in response to the need for individuals to do things together has merit to my mind. Intelligence is performative and adaptive - the walk not talk.

    1. Paul Herber Silver badge
      Happy

      Re: Quelle surprise!

      ' these creatures clearly have intelligence* (as do many others.)'

      I knew we'd get back to kitten videos eventually!

    2. LionelB Silver badge

      Re: Quelle surprise!

      > So intelligent systems acquire language ...

      Intelligent social systems (a tautology - no point in language if you've no reason to communicate with anyone/anything).

      > ... rather than language systems acquiring intelligence.

      Then again, we (and possibly some other social species) most certainly use language to develop, train and generally enhance our intelligence. So while I suspect that few (outside of the LLM hype industry at least) would argue that language is the be-all-and-end-all of intelligence, I don't think it is particularly contentious to posit that language may indeed be useful in the development and enhancement of intelligence.

      > Intelligence is performative and adaptive - the walk not talk.

      Absolutely agreed there... well, except insofar as the talk sometimes is (an aspect of) the walk... if you see what I mean.

    3. Blue Pumpkin

      Re: Quelle surprise!

      I don't remember who it was (am obviously approaching old fart age) but there was, some possibly animal behaviourist, who named one of his wards Norm Chimpsky

  3. Eclectic Man Silver badge

    Sensory input

    This is very difficult. Babies have much more sensory input than merely vision and sound. And actually even baby / toddler vision is usually stereoscopic, seeing things in three dimensions, rather than the two-dimensions from a Go-Pro as well as auditory location, having two ears.*

    *A recent genetic treatment of a toddler allowing her to hear for the first time was proven to work when she turns her head to look at the origins of sounds. https://www.theguardian.com/science/article/2024/may/09/uk-toddler-has-hearing-restored-in-world-first-gene-therapy-trial

    1. Anonymous Coward
      Anonymous Coward

      The brain is also 3-dimensional

      Today I saw an article in "Smithsonian" - a piece of brain half the size of a grain of rice removed from an epileptic patient having a lesion removed - 57K neurons, 150 million synapses, 230 mm of blood vessels. The computational base of current LLM's are a very suboptimal minimum as far as emulating intelligence goes.

      1. Herring` Silver badge

        Re: The brain is also 3-dimensional

        And also it's analogue rather than faffing around with tiny floats.

        I have heard that the human brain takes maybe 14 Watts to run. And that's if you're trying something hard like explaining a problem to a manager.

        1. Anonymous Coward
          Anonymous Coward

          Re: The brain is also 3-dimensional

          I'm sure it's only a matter of time before an AI declares its self non-binary.

          1. LionelB Silver badge

            Re: The brain is also 3-dimensional

            Oh, definitely. They'll be qunary.

          2. Plest Silver badge
            Happy

            Re: The brain is also 3-dimensional

            Or given the garbage woke data it's probably trained on, starts demanding some nonsensical pronouns eveyone uses or it'll be off to HR to complain you called it a "he", "she" or "it".

            1. Anonymous Coward
              Anonymous Coward

              Re: The brain is also 3-dimensional

              define 'woke'

              1. Sherrie Ludwig

                Re: The brain is also 3-dimensional

                define 'woke'

                it's the default attempt at a slur by persons who can't spell empathetic or compassionate.

                1. LionelB Silver badge

                  Re: The brain is also 3-dimensional

                  The original definition* is, roughly, "Alert to social injustice" (80s African-American slang).

                  Those who use it as a slur are either ignorant of the definition (duuuh... it's a bad fing because those people disagree with me so it must be a bad fing), and/or are incapable of perceiving social injustice (unless it happens to those in their own image), and/or genuinely believe that being alert to social injustice actually is a bad thing.

                  *Of course there has been a concerted effort by the alt-right to subvert the meaning to "Perceive social injustice where there is none". Which doesn't of course mean that there is none - just that they don't perceive it. Because, in a neat irony, they're not alert to it. The original definition still works just fine.

                  1. Anonymous Coward
                    Anonymous Coward

                    Re: The brain is also 3-dimensional

                    Don't forget people who can't accept that, to the privileged, equity looks like oppression. They kid themselves that just because some laws have changed, we've somehow achieved equality when, in actual fact, society is less than 10% along the road to becoming truly equitable.

                    Objective reality has an inherently left wing bias.

                  2. Anonymous Coward
                    Anonymous Coward

                    Re: The brain is also 3-dimensional

                    Spoken like a true Social Justice Warrior.

                    There's no oppressed minority you wont get upset about on their behalf, and no internet form you wont go to battle on loudly.

                    As long as it doesn't involve doing anything real life, eh.

                    1. LionelB Silver badge
                      FAIL

                      Re: The brain is also 3-dimensional

                      Spoken like a true Culture War Warrior.

                      Sure I get upset about oppressed minorities - and by oppressed majorities too, funnily enough. We call that "empathy". (I happen to hail from a country where the majority were, famously, enthusiastically oppressed by a racist minority, but I don't think that's a prerequisite for empathy.)

                      > ... and no internet form [sic] you wont go to battle on loudly.

                      Oh, the irony.

                      > As long as it doesn't involve doing anything real life, eh.

                      How would you know that?

                2. Benegesserict Cumbersomberbatch Silver badge

                  Re: The brain is also 3-dimensional

                  ...and too cowardly to use the slur they were thinking.

                  1. Yet Another Anonymous coward Silver badge

                    Re: The brain is also 3-dimensional

                    I'm not afraid to call a Belgian a Belgian

        2. LionelB Silver badge

          Re: The brain is also 3-dimensional

          > And also it's analogue rather than faffing around with tiny floats.

          Well, sort of. The majority of inter-neuron communication takes the form of "spikes" of cell membrane potential at synapses. These are effectively discrete events, often occurring in "trains" of consecutive spikes - although, of course they take place in analogue time. It is thought that information is encoded and transmitted via the relative timings of spike events, and/or the rates of spike trains.

          > I have heard that the human brain takes maybe 14 Watts to run.

          Indeed. Apart from the shear numbers of units and connections, natural neural systems are orders of magnitude more energy-efficient than anything we can achieve with current technologies.

          > And that's if you're trying something hard like explaining a problem to a manager.

          Steady on! That's setting the bar a bit high, isn't it?

          1. Eclectic Man Silver badge
            Joke

            Re: The brain is also 3-dimensional

            "> And that's if you're trying something hard like explaining a problem to a manager.

            Steady on! That's setting the bar a bit high, isn't it?"

            Reminds me of a line in Mark Twain's 'The Adventures of Huckleberry Finn'. Huck has just done a girl a good deed (saved her form losing a lot of money), and she says she will pray for him.

            'Pray for me! I reckoned if she knowed me she'd take a job that was more nearer her size.'

            Still, explaining a problem to a manager is usually easier than explaining a solution ...

        3. m4r35n357 Silver badge

          Re: The brain is also 3-dimensional

          Half precision: the wrong answer - faster!

          (thanks to The Simpsons, Max Power episode)

          1. LionelB Silver badge

            Re: The brain is also 3-dimensional

            > ... the wrong answer - faster!

            And what's not to like? Since simulation of any analogue system by floating-point arithmetic is inherently (almost) always wrong, this is a Good Thing.

            Seriously, though, what really matters is how much wrong you can get away with. Not my field, but I'm going to guess that games programmers working on physics engines or low-level graphics are the experts here.

            1. m4r35n357 Silver badge

              Re: The brain is also 3-dimensional

              What matters most is whether someone has the chops to know what precision they really need. It is highly non-trivial, and most programmers are NOT experts here - you really need mathematicians for this!

              BTW half precision is less than 5 decimal places of precision, before you do any actual FP calculations - good luck!!

              I do all my FP work in long double (SW quad precision on aarch64, HW 80-bit on x86-64).

              1. LionelB Silver badge

                Re: The brain is also 3-dimensional

                > What matters most is whether someone has the chops to know what precision they really need.

                Sure. Landing a half-billion pound rover on Mars and moving a dorky cartoon character around a screen, say, represent rather different values of "need".

                > It is highly non-trivial, and most programmers are NOT experts here - you really need mathematicians for this!

                I happen to be a mathematician, but unfortunately the wrong kind. I can confirm, at least, that the maths is hard, fiddly and (at least for me) no fun at all. I do know a couple of physicists who are very good at this, because sometimes they have to be.

                > I do all my FP work in long double (SW quad precision on aarch64, HW 80-bit on x86-64).

                YMMV. In my field (mostly computational neuroscience these days) double is almost always fine (though single-precision is frequently not). The only "chops" required to establish this is that, at least where tractable, simulation results do not diverge significantly from known analytic results. I regularly need to deal with large neuroimaging data and large-scale simulations, so the computational costs of extended precision can be prohibitive.

                1. m4r35n357 Silver badge

                  Re: The brain is also 3-dimensional

                  My work is private "vanity" projects, but my problems always end up exploring the roundoff limits (I even used MPFR when I first started looking at chaos). When you are careful about this, it is easy to see those limits and not lose precision.

                  I don't know that 80-bit (long double) is more expensive than 64-bit (double) is much more expensive (on x86-64), it all goes though the FPU so it is only a word-length difference, but I guess you know your domain better than me ;)

                  Not sure what problem the downvoter (not me!) had with your post . . .

                  1. LionelB Silver badge

                    Re: The brain is also 3-dimensional

                    > I don't know that 80-bit (long double) is more expensive than 64-bit (double) is much more expensive (on x86-64), it all goes though the FPU so it is only a word-length difference ...

                    That's a fair point. The last time I tested this was many years ago (and I don't remember the exact hardware or GCC support at the time), so maybe worth revisiting (the hardware is mostly Xeon boxes). I work mostly in Matlab these days - which doesn't support 80-bit long double - but do still code in C99 a fair amount too.

        4. cookieMonster Silver badge
          Trollface

          Re: The brain is also 3-dimensional

          Explaining something to a manager, 14 watts?? Wow, from past experience I’d put that value wayyyyyyy higher

  4. This post has been deleted by its author

  5. Pete 2 Silver badge

    Get Smart

    > over time, the child will likely become much smarter and more creative than the model.

    An assertion I would love to see tested.

    Although first, that would require a philosophical debate for what "smart" means and an even more difficult debate over what "creative" is all about.

    But maybe that's the point. If all these philosophers (who determine such things) keep disagreeing with each other violently enough and slagging each other off in the popular press, you can keep yourself on the gravy train for life

  6. heyrick Silver badge

    The problem here is obvious

    Mini people, usually, have innate intelligence. For a while the parent(s) tend to the little person as that little person gets bigger, explores their bit of the world, and gains the ability to communicate with others. And then the cycle repeats. Therefore the capacity to learn and some degree of intelligence is built in. Without it, those people would be dumber than a flat earther and wouldn't survive, which in evolutionary terms does rather defeat the purpose of being born.

    The machine, however, is desperately lacking in anything that resembles intelligence. It might be able to amass huge amounts of data and start to pick up patterns in it. There's no magic here, we could do that as well, it might just take us a lifetime because of how we work (but, note, many people's lifetimes got us to this point). Feeding yet more information to a machine to try to pick up more patterns isn't intelligence, it's basically throwing spaghetti at the wall and hoping the patterns might spell out some sort of word.

    1. Eclectic Man Silver badge

      Re: The problem here is obvious - Aside Flat Earthers

      There is an interesting explanation of 'Flat Earthers' in chapter 1 of 'How to talk to a science denier' by Lee McIntyre (ISBN 978-0-262-04610-7). He attends a 'Flat Earth' conference, and finds that most of them are coping with severe personal trauma, and the Flat Earth groups are basically support networks (although without actually addressing the causes of the trauma). The books is well worth reading, IMHO.

  7. Howard Sway Silver badge

    First sample output from baby-trained AI

    "Will someone please take this fucking camera off my head?"

    1. Anonymous Coward
      Anonymous Coward

      Illustrative instance of Toddler trained AGI for Analysis

      User: can you crunch these numbers for me?

      AI: I’m scraping The Wiggles.

      User: [sternly] I’m not asking mister, I’m telling!

      AI: [sobbing] You’re not the boss of me!

    2. spold Silver badge

      Re: First sample output from baby-trained AI

      Given their innate wobbliness, I assume "attaching" said camera involves some Super Glue (although experiments with sheets of self-adhesive velcro are looking promising in secondary applications, such as keeping them in car seats, attaching them to kitchen cabinets, or enabling the dog to take them for a walk).

  8. Anonymous Coward
    Anonymous Coward

    Restated results:

    Given a number of videos where various words are spoken, the computer successfully showed the bits where a particular word was spoken.

    How is this more impressive than speech-to-text feeding into a hash? Key is the word, images are the values.

  9. Winkypop Silver badge
    Terminator

    And so

    The Borg Race emerged…

    1. Plest Silver badge
      Unhappy

      Re: And so

      I'd love to laugh at this but you've hinted at something very, very sinister coming down the road.

      1. mattaw2001
        Terminator

        Re: And so

        Just to add to that shiver bionic eyes have been growing in popularity since 2012. For those who have them, there really is nothing preventing the computer from adding or removing items from their view :)

  10. Bartholomew
    Meh

    Maybe point that camera at the smartest species on Earth.

    If they somehow manage to get this to work, maybe we need to strapping one onto a Dolphin, or a whole pod, and finally translate that last message "double-backwards-somersault through a hoop whilst whistling the 'Star Spangled Banner'" correctly as "So long and thanks for all the fish."

    "Humans think they are smarter than dolphins because we build cars and buildings and start wars etc., and all that dolphins do is swim in the water, eat fish and play around. Dolphins believe that they are smarter for exactly the same reasons." - Douglas Adams

  11. Mungo Spanner

    This was the exact subject of my PhD 25 years ago...

    Modelling language acquisition on the neurological and perceptual development of infants was the core of my thesis - and it worked well on a '486 - but I didn't have a cloud of GPUs to scale it up (and I wasn't strapping Go-Pros to toddlers) -

    https://eprints.soton.ac.uk/464080/

    1. that one in the corner Silver badge

      Re: This was the exact subject of my PhD 25 years ago...

      So, anyone got access to read the full article El Reg is reporting on?

      Did they do their literature search all the way back to the Stone Age of Computing, 25 years ago?

  12. EricB123 Silver badge

    So 2 New Sentences ?

    So the boffins learned two new phrases. "What tbe hell is that Gopro on that child for " and I'm calling the police!"

  13. that one in the corner Silver badge

    Totally missing the feedback loop

    Children learn language as an integral part of their learning about the rest of the world - physics[1], linguistics, psychology[2] (and the other ologies) all going on at the same time. Baby can even combine all of these things and use their words to control the adults into affecting the desirable physical changes to the world ("Bigger, Daddy" and "Upsy").

    And, strangely enough, there are feedback loops involved: Baby does A, the world does B - yay. Baby does C, the world doesn't do B - frustration. Baby does A again - you get the idea[3]. And these interactions are strangely subtle - learning that saying "Upsy" to a shop dummy doesn't work - oh, wait, it makes Mummy laugh, so maybe it *does* work, in another way.

    So, what does this strapping cameras to toddlers strategy capture? Practically nothing!

    As comments noted above, it misses all the other sensory input (including directional cues from binaural and binocular), smell, touch. It misses the outputs from the toddler - especially facial expressions that are used to influence their surroundings (aka adults). Most importantly, it misses all of the internal interactions - including every single element of intention. Even if the video does capture an event, interpreting that, in order to label it, is something that adults are often rubbish at, let alone a self-training ubernaive[4] bit of software (it turns out, Toddler Tina wasn't clumsy, she just found out that dropping onto her bum shook the table and made all the bricks fall over - yay!); so even employing adults to interpret and label the data before feeding it to the LLM-in-training isn't capturing the child's learning process.

    If this experiment did capture the way that the child is learning language, rather than being a shallow mimic that falls apart as soon as it is used beyond its ill-defined limits, then it has managed to not only model that learning process, but did so *on top of* a model of all the missing inputs, all the interactions between child and carer (even those where the carer is not present but the child looked for them), and that simplest of things, a complete theory of mind to determine how much the acts observed were intentional (and thereby strengthened an existing connection) or random flailing of the limbs (if only that bit of rusk in the romper suit would stop itching!). Never mind that the infant itself barely has a working theory of mind yet! Somehow, all that metamodelling seems, shall we say, unlikely.

    We are seeing the limits of LLMs that have (initially) been trained on textual data that was labelled by humans[5] and therefore had a fighting chance of getting its nadans adjusted in the right direction based upon the punishment/reward feedback applied to its outputs during training.

    What are the chances that the LLMs resulting from this are going to fall apart less readily, with so little accurate feedback?

    And what are the chances that these (probably interestingly new and different) limitations are accurately probed and evaluated by the researchers? And the people who leap on this and hypye it?).

    [1] lots of chemistry and biology going on as well, but that is less - intentional - on the toddler's part than building up towers and seeing which ones make the most noise as they fall.

    [2] aka social engineering: smile and frown at the right time and you control them completely! Dance for me, giggle gurgle!

    [3] In fact, I understand that we are so aware of this feedback that it has even been used to create some crude software models! What are they called again, it is on the tip of my tongue.

    [4] please forgive the the missing diaresis

    [5] and usefully labelled, as even if the textual inputs were misunderstood by the labeller, that is still "first-person" interpretation direct from the source data

  14. captain veg Silver badge

    messing around

    Stop faffing around with cameras and get brain implants in there. There's money at stake!

    -A.

    1. that one in the corner Silver badge

      Re: messing around

      > get brain implants in there

      Paging Mr Musk

  15. that one in the corner Silver badge

    Use some compute instead of borging the babies

    The graphics and physical modelling chaps can do great things with recreating video scenes as though they had been shot from different angles. Add in some head location estimation & eye tracking, like what your camera does nowadays, to feed in the viewing angles.

    Cross-disciplinary research, it is all the rage.

    Then maybe you could lots and lots more footage, without the need to train the infants not to pull the cameras off (honestly, given how many hats and pretty bows get flung from strollers, those things must be glued on!).

    (Of course, that would lose the GoPro sponsorship, but they could go into partnership with the nanny cam suppliers).

  16. that one in the corner Silver badge

    Ok, there was no need for a wall of text

    Shouldn't attempt to criticise their methodology.

    All they want to do is fling different stuff into the bucket to train an LLM. From that p.o.v. job done.

    Just treat the rest as valueless hype, meant to pull in the cash and get people to talk about it. So, job done, sigh.

    We like to hear of research that will push the bounds of human knowledge and analyses of data that lead to ground breaking discoveries, but we it is just unfair of us to expect everyone to have the same attitude.

  17. Triggerfish

    Inference

    The human brain can infer things. That's why you could teach a child what a chair is, but when saying where can you sit they would choose things that are not just chairs, tables, low walls etc.

    Is AI capable of this? Because I feel it's a big factor in things like using language, that is not just having a good vocab but also understanding nuance, idioms and so on.

  18. Anonymous Coward
    Anonymous Coward

    I would love to have trained a pet AI

    when I was 5, 6, 7, 8... I'd take it everywhere and it would see everything I see, and I'd tell it what I think of everything and it would listen.

    And then when I turn 13 or 14 I would want the thing memory-wiped or at least nobody else is allowed to make any queries ever.

  19. JRStern Bronze badge

    Back to the 1960s

    Just FYI Chomsky noted back in the 1950s and 1960s the extreme efficiency with which children learn vocabulary and sentence structures.

    But he never got to even first base trying to figure out how they did it.

    All that is clear at this point is that the way LLM's are trained today is pretty much the exact opposite of how children do it.

    In effect children learn the transforms as well as the reference grid.

    Has the GoPro game never been tried before? Dunno. Probably a worthwhile study, but likely a looooong way from helping LLMs.

  20. Locomotion69 Bronze badge
    Thumb Down

    Strapping electronics to a toddler...

    ... is one of the worst ideas ever.

    I mean, these are not medical devices, are they?

    You do not strap electronics like this to a child - not to mention if BT/WiFi/5G is enabled as a "feature".

    1. captain veg Silver badge

      Re: Strapping electronics to a toddler...

      Indeed not. Electrode implantation is far more efficient, not to say cheaper.

      -A.

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

Other stories you might like