DeepMind's Mustafa Suleyman says general AI is still a long way off
techcrunch.com
techcrunch.com
It's well known in AI that common sense AI systems are incredibly hard to build and expert AI systems are relatively straight forward to model and build. Given that the most 'common sense' type products we see like Siri and Alexa still appear to be rules based systems (that have been enhanced with historical data no doubt), we all know the world is not suddenly going to become a common sense AI paradise unless a new Marvin Minsky makes some breakthroughs.
I find that the term 'AI' is frequently used when a company has taken an existing problem and enhanced the solution by using machine learning at one or more places. Nothing wrong with that fundamentally, but calling that AI is more of a marketing gimmick than anything else.
Google/Facebook do some of the best machine learning enhanced workflows that delight users. They usually refrain from branding it as AI.
What we have right now is a boom in AI engineering, with AI research fairly neglected because it doesn't produce the same kind of hype.
Here's a Twitter thread [1] where NLP researchers describe people who work on machine translation as 'depressed', because they spend their time tweaking neural net parameters to get tiny improvements over the best techniques of last decade, with no opportunity to understand what causes an improvement.
I'm pretty sure Siri/Alexa/GoogleNow don't use rules. Like all modern Q&A systems like IBM Watson, they use a variety of statistical word match/associations and phrase/sentence response templates. Like Watson, the algorithm identifies the most probable word match to fill the void in your query and then inserts it into the most appropriate answer template. Yes, it's a bit hit or miss (like Watson answering "Toronto" for the US city with an airport named after a WWII hero). But if Apple/Google/Amazon adds some simple sequential disambiguation and refinement (clarification Q&A), I suspect this approach should answer most casual questions quite nicely.
IMHO, calling it "AI" means that the problem is not amenable to a closed form solution, so some form of iterative search and refinement is required. In venerable GOFAI parlance, AI problems are inherently 'scruffy'. Once the problem can be solved efficiently, it becomes 'neat' and it's no longer considered to be AI.
IMHO, most of today's progress in AI is scruffy (probabilistic word matches and deep learning), since they're driven not by formal internally-curated information but by smart search against presumed-relevant external referents, requiring secondary refinement -- a process that is largely inefficient and open-ended and error-prone, and will remain so indefinitely, I think, since we're not trying to improve the quality of the referent data, which is the weakest link in the cognitive chain.
Having said that, I still see abuse of the phrase (more in industry than academia) where standard machine learning applied to a workflow is marketed as some big AI in a product. But I understand why they do it - it usually helps in the long drawn out enterprise sales process to be perceived as an AI powered product.
Something like this would be immensely valuable, and would immediately improve healthcare everywhere, since the software could be updated with new or obscure knowledge right away, so you wouldn't have to shop around for doctors until you came across someone who had read of a very rare disease somewhere, the software would already know everything you might have.
Is there something like that out there? It sounds like there must be, but I don't know of any doctors using anything like this.
There are a lot of barriers to presenting it as you imagine, though: some liability, some workflow, and some just the deep lack of appropriate datasets. There is also some adoption issues - clinicians are rarely going to be happy with a black box approach but it can be difficult to explain/expose the details of such analysis in a way that clinicians will accept.
A last thought: this is actually very hard to do well and dangerous to make mistakes with - so your rather cavalier claim that there would be immediate and wide spread improvement of health care should be tempered a bit.
Another thing to consider in any work like this, you really need to thank about what your false positives mean, in a practical sense. If your numbers are big enough (e.g. screening) and your positives result in a procedure (e.g. biopsy), eventually someone is going to die based on a false positive. This may still be the better option, of course, but you really need to demonstrate that, and think about what your sensitivity and specificity tuning needs.
Going back further, medical diagnosis was a classic application of expert systems even in the 1970s. They ended up being something of a failure then due to the need to laboriously extract huge amounts of knowledge from doctors to get them to work, but it does show that it's long been a popular idea in the field.
MYCIN was never actually used in practice but research indicated that it proposed an acceptable therapy in about 69% of cases, which was better than the performance of infectious disease experts who were judged using the same criteria. [...] Some observers raised ethical and legal issues related to the use of computers in medicine — if a program gives the wrong diagnosis or recommends the wrong therapy, who should be held responsible? However, the greatest problem, and the reason that MYCIN was not used in routine practice, was the state of technologies for system integration, especially at the time it was developed. MYCIN was a stand-alone system that required a user to enter all relevant information about a patient by typing in responses to questions MYCIN posed. The program ran on a large time-shared system, available over the early Internet (ARPANet), before personal computers were developed. In the modern era, such a system would be integrated with medical record systems, would extract answers to questions from patient databases, and would be much less dependent on physician entry of information. In the 1970s, a session with MYCIN could easily consume 30 minutes or more — an unrealistic time commitment for a busy clinician.
Given that the systems integration problems should be soluble now, if not already solved, it does seem a little surprising that there isn't a modern version of such a thing, far more sophisticated, in general use. I guess the ethics and liability problems still loom larger than the above quote suggests.
[0] https://en.wikipedia.org/wiki/Mycin
[1] http://people.dbmi.columbia.edu/~ehs7001/Buchanan-Shortliffe...
What's more, its definitely open to the public, except it isn't really promoted at all (it's only open in the interests of openness, you're not really meant to use it.) As such I can't find the URL for the life of me. Perhaps they closed it.
Any doctor or nurse will know about it, perhaps someone can post the link.
Lots, and for decades. If you Google "AI medical diagnosis" you get a zillion hits, including a dedicated Wikipedia page.
Also such system is quite complex in many dimensions, so the natural risk minimization strategy would be to introduce it in stages: a system that warns about drug interactions here, monitors drug dosage there, reminds you to order extra tests, etc .And of course everything needs to be clinically tested to be accepted by docs(even though it isn't stricly necessary, it's just a second opinion, reminders, not the final word) . It makes for a very slow process.
And they they all face very big adoption barriers , in some cases upto 96% of their recommendations are skipped. Some of this is due to usability , but a lot due social/political factors.
Now after seeing this, what company in it's right mind will invest in developing a full diagnostical system ?
Actually , i know of one who did, http://www.isabelhealthcare.com , but at least according to the founding story , it began because the founder nearly lost his daughter to a hospital misdiagnosis, so he wasn't really driven by money.
And as far as i know(haven't followed them in a long time), they only achieved small scale sucsess.
The other company now entering this is IBM , but that's a bit of a unique case: IBM may be in big trouble so they take a huge bet, and the field they choose - cancer , is such that patients have much more power than usual , a patient has time, and can demand a diagnosis by Watson , heck it's even a cloud service, you just need to send blood and pay.
For what it's worth, I completely agree with the sentiment presented here that we're far enough away that we don't know what direction even we need to start down to make the best progress, which in my view makes any sort of prognostication of questionable value. It's also why I think it's not that productive to have people like Stephen Hawking arguing about future AI as a threat that somehow needs to be mitigated. Even if he's right in the general point, we know so little about how that future system would look that it would be like asking Isaac Newton to make cars safer.
Love your point about hawking. I treat that as pure entertainment but since the general population treats hawking with high regard (well deserved) his opinions about things he doesn't know much about also carry weight in shaping public perception.
As far as speculating on how scary an AI driven future can be, a filmmaker in the 60s or a high school student today can do as compelling a job as Hawking.
Even worse, we have some strong theoretical results suggesting that none of the techniques we have developed so far can ever lead to AGI, and that refers very much to machine learning primarily.
I'm talking about the work that was done on inductive inference, and then PAC-learning, in decades past (starting in the 1960's). There are many negative results and very few positive ones.
I wouldn't say I've reviewed the bibliography extensively (it's a huge area spanning a couple of fields), but from the little I've read I'm getting the feeling that, if you want to be able to learn any function in polynomial time, with a polynomial number of training examples, and with arbitrary error, you're basically stuck with the easiest classes of concept to learn. And if you want to learn specific distributions or representations (like DFAs or PCFGs) then you can just forget about it.
Which reminds me of something that Andrew Ng has said, about how recent results in machine learning and deep learning only reached the "low hanging fruit" and further progress is going to be much harder. No references to all this- I'll dig 'em up if required.
Any result which predicts that human intelligence cannot exist.. has to have an error somewhere. Perhaps the error here is in assuming that a learning algorithm able to fit any function in polynomial time is needed in the first place.
People are too ready to leap at results that say, "Well, machines can never do X, and I can do X, therefore human-level AI is impossible."
In general, my counter to that would be "you probably can't do X either". It's just that you can do it well enough. You don't encounter the weird cases where your ability fails. But if that's allowed as an argument, then of course machines are more powerful too. I can absolutely solve the halting problem if "solve" can be interpreted as "almost always come up with the correct answer for the kinds of problems people encounter in the real world".
Also, don't forget what Terry Pratchett said, about how thousand-to-one odds crop up one time out of ten. The problems with AI in the past started when people tried to attempt to solve anything beyond the trivial, toy-problems they could demonstrate were solvable "in the lab". Like SHRDLU, which solved blocks world, but could not really scale. The same goes with machine learning today: some problems are solvable in restricted settings, or in a lab environment, under controlled conditions and with safe assumptions etc, but once you try to do the same thing in the wild...
You can even look at the big success of Deep nets for examples of that: image processing and speech recognition. Whereas it's possible to get really good results in both fields, image processing still relies very much on having good quality images to work with, and speech recognition is really crap if you have a bit of an accent, or a speech impediment etc. (at least that's what I gather- these are not my specialties).
Yes, it's entirely possible that we have missed something really important and when we figure that thing out we'll end up with a completely different Theory of Computation than we have now, and that may allow us to progress in learning concept classes that, today, are not learnable. That however, would be a revolution on par with Einstein's in physics.
Another thing that might save our collective ass, in terms of learnability, is that many of those results are assuming that P != NP. If that turns to be wrong, then those results will be off.
However, all this is very unlikely and in any case a) we can hardly expect such a profound re-interpretation of all that we know to happen within the next couple of generations and b) even if it does it remains the case that at this point in time everything we know suggests that there are some concept classes that we can simply not learn efficiently, at least not with computers.
Edit: what I'm really trying to say is that we may not understand human minds, but we're talking about computers here, so basically I agree with what you say, but with an added flourish :)
Dana Angluin, Inductive inference of formal languages from positive data. [http://www-personal.umich.edu/~yinw/papers/Angluin80.pdf]
Dana Angluin, Learning regular sets from queries and counterexamples. [https://people.eecs.berkeley.edu/~dawnsong/teaching/s10/pape...]
Mark E. Gold, Complexity of automaton identification from given data. [http://sebastian.doc.gold.ac.uk/papers/Language_Learning/gol...]
Mark E. Gold, Language Identification in the Limit.[http://web.mit.edu/6.863/www/spring2010/readings/gold67limit...]
Pitt, L., & Pitt, L. (1989). Inductive Inference, DFAs and Computational Complexity. 2ND INT. WORKSHOP ON ANALOGICAL AND INDUCTIVE INFERENCE (AII, 18--44. Retrieved from http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.187....
Leonard Pitt & Manfred K. Warmuth. The minimum consistent DFA problem cannot be approximated within any polynomial. [http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.86....]
Also see below for a gentle introduction to PAC-Learning, which actually allows for learning several concepts in restricted (but non-trivial) settings:
http://web.cs.iastate.edu/~honavar/pac.pdf
There is a lot more bibliography on both Inductive Inference tractability and PAC-Learnability. You'll have to follow the trail of references :)
See deong's comment below. We may not understand how our minds work, but we do understand how computers work and that's where the learnability results come in: because they pertain to computers, not human minds.
Also, it may be that we don't need to learn every single thing we know. See Chomsky's Universal Grammar, etc.
I've been looking for work in this area without much success since publication is dominated by the opposite movement. I'd be grateful for any references you can dig up.
If we accept that GAI is far off, though, it's very likely that we simply can't have a productive conversation about it at this time. In fact, such conversations could even be counterproductive if we make incorrect assumptions, leading to the illusion of safety.
Most talks about the dangers of AI assume we'll one day finally "break through" and create a metaphorical ghost in the machine, at which point all bets are off. However, this scenario seems awfully unlikely to me. Human intelligence is far beyond that of any other known organism, but there are plenty of intelligent animals on Earth, and the process of solving GAI will almost certainly produce intelligence on par with those animals before it breaks through to human-level intelligence. Dog-level intelligence is a decent milestone; at that point, we'll clearly have created GAI in the sense of creating a being that can exhibit truly "intelligent" behavior (including emotions), but it's clearly not intelligent enough to cause any problems that are specific to GAI. The rebuttal to this argument is that a generalized learning algorithm running on a computer will be far more powerful than a dog's brain, but it's fallacious to assume that the only difference between a dog and a human is the processing powers of their brains. That would imply that a bigger brain (however you choose to measure it) is all that's needed to create more intelligence. As far as we can tell, this isn't true, and architectural differences (which imply "software" differences) seem important for creating higher intelligence.
When we do get to GAI, we don't know if it'll be through software running on a "normal" computer, or if it'll be running on specialized hardware in a specialized shell designed to navigate a real environment, or if it'll be in some other form that we haven't really thought of yet. The dangers and resolutions are very different between those forms, and when we don't seem close to GAI at all, it can be argued that talk about its "dangers" may be a complete waste of time, or perhaps even counterproductive to safety in the long-run.
Why are emotions required for intelligence?
But second, it may very well be true that emotions are necessary for GAI. The GAI probably needs to have some sort of basic instinct or drive; otherwise it'll just exhibit completely random behavior because it has no actual intentions (remember we can't explicitly program intentions for tasks, because then it's not general intelligence). The very fact that the GAI will "want" to act according its instincts/drives implies it needs some concept of satisfaction or at least desire. I think that satisfaction and desire meet most definitions of "emotion," and they can absolutely lead to emotional (as opposed to logical) behavior because any GAI is going to be basing satisfiability and desire off of heuristics, not perfect data, because the instincts/drives need to be general themselves, such that it's impossible to deterministically satisfy them. (Otherwise, the behavior degenerates into something we would recognize as non-intelligent.)
I say "probably" at the start of all of that because we've never made GAI, so it's impossible to know for sure what it takes to get there. However, that model is coherent with evolution and the animal (particularly mammalian) condition. It's unclear exactly what the drives of animals are--easy guesses are survival and reproduction--but we know there are biological responses that heavily guide animal behavior.
For example, we don't eat to live so much as we eat because relief from hunger is enjoyable. However, we've developed an even more abstracted response, and so most of us just kind of like to each in general, even when we're not particularly hungry. When our body can tell we're eating, it releases dopamine, which associates to a mood of well-being. This elevate mood is an emotional response; it's the high-level abstraction we use to choose our behavior.
More complex decisions (involving social situations or whatnot) can just be mapped to ever-higher abstractions that are enforced through mechanisms we developed through evolution. Therefore, it can be argued that literally 100% of high-level decision making is emotional. Even if we ultimately use logic to try to make the best decision, our determination of what we want out of the "best decision" is still emotional. Even more extreme, it may be true that emotion turns out to be the definition of general intelligence, because it's at that point that the agent is making decisions in arbitrary situations based on some sort of internally consistent desire, rather than a prescriptive goal.
So, the idea of a cold drone that can do everything a human can do, but doesn't feel any desire or emotion, may in fact be a contradiction. Achieving behavior that complex may not be possible (in the real world, with the data available in our plane of existence) without some emotional mechanism.
AI improving in those specific tasks are bound to affect people today, it's just the comfy CEO and his pals are not going to be outsourced (or AIsourced or whatever) any time soon, although it might happen eventually.
The point being: You do not need general AI anytime soon, the consequences will still be significant, I would bet.
"Well, what if they took all the AI power that makes that GO thing beat the best player, and converted it into regular intelligence? It would probably be as smart as the average person!"
Is a real question I've been asked several times by friends. Of course, no one wants to be friends with "that guy" at a party or family dinner, so I usually just say "Yeah, that would be neat" and move on. But I really do like talking having serious conversations speculative AI and geeking out over AI. It's hard though, because most people just want to talk about the kinds of AI they've seen in movies, which are unrealistic compared to what's possible and practical.
I tell them that any AI that's designed to only solve one type of problem (or a finite set of problems) is "weak." An AI that can figure out how to solve an arbitrary problem is "strong."
I then tell them that we've never been able to create a strong AI, and we've never been able to convert a weak one to a strong one. They ask why not. I tell them that every weak AI, at its core, has a mathematical way to evaluate its outcome. Every time it does something, it does it by assigning some number to the outcome, and it tries to get that "number" as close to the "good" number as possible. All of our advances in weak AI have been in ways to come up with better numbers, or use computers to come up with good numbers to situations that we understand very well but aren't good at attaching numbers to. Finally, I say that with the way we make weak AIs, to make it into a strong AI, we'd have to come up with a way to assign an accurate number to literally every single decision the computer might be asked to make. But there are infinitely possible decisions, so we'd have to come up with infinite ways to come up with a number. To come up with strong AI, we need some "universal" way for the computer to think about the world, but we have no idea what would look like.
This explanation is hardly adequate for a true understanding of AI, but it works pretty well for helping them understand the gist of the problem without using some misleading metaphor that would just give them another bad understanding.
Quoting from the excellent "Wait But Why?" blog post The AI Revolution: Our Immortality or Extinction (http://waitbutwhy.com/2015/01/artificial-intelligence-revolu...):
> In 2013, Vincent C. Müller and Nick Bostrom conducted a survey that asked hundreds of AI experts at a series of conferences the following question: “For the purposes of this question, assume that human scientific activity continues without major negative disruption. By what year would you see a (10% / 50% / 90%) probability for such [Human-Level Machine Intelligence] to exist?” It asked them to name an optimistic year (one in which they believe there’s a 10% chance we’ll have AGI), a realistic guess (a year they believe there’s a 50% chance of AGI—i.e. after that year they think it’s more likely than not that we’ll have AGI), and a safe guess (the earliest year by which they can say with 90% certainty we’ll have AGI). Gathered together as one data set, here were the results:
> Median optimistic year (10% likelihood): 2022
> Median realistic year (50% likelihood): 2040
> Median pessimistic year (90% likelihood): 2075
> A separate study, conducted recently by author James Barrat at Ben Goertzel’s annual AGI Conference, did away with percentages and simply asked when participants thought AGI [Artificial General Intelligence] would be achieved—by 2030, by 2050, by 2100, after 2100, or never. The results:
> By 2030: 42% of respondents
> By 2050: 25%
> By 2100: 20%
> After 2100: 10%
> Never: 2%
a) What is the computational power required? (stuff figured out back in the 80s turned out to work well for deep learning, but it took another two decades until the power was there at an economical price.)
b) What kind of mistakes do human-like thinking machine make? Additionally, are those mistakes actually a fundamental piece of general intelligence?
Some of the core pieces to general intelligence may be solved soon, but it could take a long time for the economics to become compelling or even possible. Presumably you can collapse time by throwing more resources at an "AI", but humans, with their brains running more or less 24/7 take many years before they can start producing really valuable work. We have assumed you can preload an "AI" with concepts or just copy and paste, but this might magnifying weaknesses just as if every employee at Google was a copy of Sergey Brin.
I except we will start seeing things, very soon, that begins to look very much like general intelligence, but the hardware limits will be an issue.
Remember, the same experts predicted flying cars in the 1960s. We are yet to have self driving cars.
TL;DR: Experts have a poor track record for predicting AI development, they aren't particularly qualitatively different in their predictions from non-experts, and "15-25 years" has been a staple prediction for decades now.
(Statisticians often deride machine learning people for having a poor understanding of statistics).
Hassabis is CEO
Step 2: Rescue the problem from necessitating general AI by developing an algorithm. Any problem which has been split off from general AI is vastly closer to being solved.
How do they know that they can't take their current deep learning-type techniques or some variation/enhancement of them and apply them to more demanding and general circumstances?
For example, they have the agent that goes around trying to score in a 3d first-person-shooter with the pixels as input. Has it been determined that there isn't a way to gradually train that system to recognize words, phrases, or even sentences that would give it hints about how to score better, maybe by associating words with other (spatial/navigation) parts of the model or more general association between aspects of the learning system in different domains.
To me it seems like some type of deep learning (or maybe a variant using spiking neurons since those can reportedly learn with fewer examples?) when grounded by integrated sensory inputs of different times and incrementally trained, may be able to already exhibit general animal/human-like intelligence, or will be able to with one or two minor 'breakthroughs'.
I'm still going with 2029 (as suggested by Kurzweil I think).
The company is working with the National Health Service (NHS) in the UK on a project around the early detection of acute kidney injuries. Critics, however, argue that the collaboration with DeepMind is wider than the company and the NHS previously disclosed. In addition, it’s also working with the Moorfields Eye Hospital on looking at how it can use the hospital’s eye scans to diagnose eye conditions better and faster. The NHS, however, mostly used its own algorithm in its project, with DeepMind focussing on the front-facing app. Suleyman argued that this is due to the fact that it’s still early days for this collaboration, which only started 12 months ago.
So in the health space they're currently doing anomaly detection and building applications for their clients. Sounds very Watson/consulting. Lots of money in healthcare due to the amount of data generated and the complexity of building regulation-compliant.
Judging from their website, they've also built an algorithm for reducing cooling energy usage in Google's data centers.
There are a ton of IoT platforms that are currently competing to connect a bunch of industrial/medical/automotive/aerospace and defense devices and servers, analyze that data with machine learning and statistical techniques, make it easy for users to build their own cloud/edge applications, and eke out every last drop of efficiency they can for their customers.
DeepMind is fortunate to have Google's brand, infrastructure, and reach. I'm just not sure what else separates them from any other "AI" (read machine learning/algorithm-building/efficiency consultancy) out there.
> I'm just not sure what else separates them from any other "AI" (read machine learning/algorithm-building/efficiency consultancy) out there.
probably the amount of pure research they produce: https://deepmind.com/research/publications/
I wouldn't be surprised if google sunsets the company and absorbs the employees in its more profitable products.
The idea that they'd take the people and put them to work making Blogger better or something? Not a chance. AI research is pretty much Google's core function. They'd no more break up that functional area than Apple would take Jonny Ive's team and put them answering support calls.
If they don't figure out compelling products in reasonable time, the researchers will become part of google research (maybe retaining the deepmind name as part of a group within google research) and the engineers will become absorbed or find new jobs. Also, DeepMind is not such a big name outside our little geek community. Try it with someone :)
My point was that ideally we like to believe that research supports products, but generally speaking, most of Google research or Microsoft research papers are rarely applicable to their products. And good researchers like being independent (that's how you woo them - with the promise of independence) so you can't usually dictate what they work on.
I have no inside information here, but my hunch is that that's overly pessimistic. From the outside, we hear about DeepMind finding cats in YouTube videos or learning to play Atari games or Go, and it's tempting to say, "Google doesn't sell Go programs, so what was the point."
However, Google does do incredibly good image labeling in Google Photos. Their autonomous vehicles might eventually be a profit maker, potentially using deep reinforcement learning methods developed to play Atari games. Certainly, Google Translate and Inbox use ideas developed from their machine learning groups.
You're right that there's often a pressure to make internal research groups more profit-driven and more connected to the revenue stream that runs a company. Microsoft Research saw this happen to some degree 10 or so years ago. But Google's core competitive advantage is their ability to collect and analyze data. Data is the revenue stream that runs the company.
Also, I think the "independence" aspect of research is overblown to some extent. Almost no one is truly independent. If nothing else, you're subject to the needs of funding agencies. Typically, researchers just want the ability to work on problems that interest them using ideas they have some control over developing. You can do that inside a company, provided that company is interested in the same types of problems you're interested in. That's been the success of corporate AI/ML labs -- many of the best researchers want to go there because it's a steady stream of interesting work free from the need to do the constant dance for grant money.
Retaining researchers who write papers is highly correlated with providing a de facto academic environment for them to work in. Microsoft research has become almost the most interesting place for researchers to go to in the last 10 years because they have developed the place to feel like a university department. Look at their research papers (public). Any overlap with commercial products looks providential more than intentional.
I'm obviously not talking of people like Jeff Dean, whose intent is to build, not publish papers. The big table paper was not the intent - building Big Table was the intent and when they were done, they ended up writing a paper about it.
If Google uses some of the DeepMind research itself, that's really good. DeepMind can make revenue from Google that way (if they have that arrangement). If they can't, it's hard to justify their existence as an independent financial entity - they're better off just being part of Google.
Probably mostly the CEO https://backchannel.com/the-deep-mind-of-demis-hassabis-1561...
I remember at AAAI 1982, the conference was handing out bumper stickers saying "AI its for real" that my coworkers and I happily put on our cars. Yeah, a little optimistic:-)
That said, I like the way Deep Mind is run, especially the forays into health science.
And we already saw that google doesn't want to touch the "uglier" aspects of ML/AI/robotics with a ten-foot-pole (e.g. see how they reacted to Boston Dynamics's humanoid robot video).
So I'm not sure if we should take this seriously. DeepMind technical views are mostly on its papers. Not press statements.
It will surprise all of us, and will likely be a "earth-shaking", akin to an extra-terrestrial alien contact. Perhaps even more so - because it would be a hard example against the idea of dualism (unless it arises from a quantum computing system - then humans can posit an "out" of course, regardless of whether it is true or not).
All pure speculation, of course. I honestly don't think we'll be able to "design" a general AI - we don't even understand how sentience and consciousness (among other things in the philosophy of mind) even works or exists in human brains. We understand the gears and wheels, but can't describe the factory.
That said in many regards, the sum total of the internet, social networks and automated systems that already exist, to me is fast beginning to resemble an emergent AGI / ASI, and furthermore is already out and about busy doing its business.
That it just happens to use networks, humans and machines as cells and DNA seems kind of irrelevant.
Of course this is a silly POV thought experiment, but seriously, if that were the case, would we even notice?
I'm not sure our own cells and DNA know they are part of a larger entity for example.
Okay, enough weird thoughts.. time for bed.
Real AI will not simplify the complexities of social problems — and to believe so shows a complete lack of understanding of those problems and what real AI will become.
When the CEO of one of the most well-known AI-companies makes a claim you disagree with, you should probably give a better reason for him being wrong than "a complete lack of understanding".
It's not clear to me at all that throwing AI at a social problem somehow solves it, so...
I mean Schmidt, Ballmer or Scully, didn't seem to have much expertise in their field.
Adding one more agent who happens to be artificial is not going to make these social-reasoning problems any simpler.