>But neither the NN nor the researcher will be able to give you a rule like: "F=(G m_1 m_2)/(r^2)" to explain the underlying reasons for the object-trajectory dynamics.
It is not impossible. There is a combination of a special neural network architecture and strong sparsity-inducing regularization that makes it possible to learn equations from dynamics dataset: https://openreview.net/pdf?id=BkgRp0FYe
>Everyone talks about extending life and eventually becoming immortal, yet no one asks what happens when we are immortal.
I see the opposite picture when these discussions pop up: everybody tries to find a downside in increased lifespan, be it personal or social. It looks like a pretty obvious defense of status-quo to me.
What is really-really sad and what really bothers me is that this attitude is so prevalent in the modern West, when the West is the only society on planet Earth that has knowledge and resources to make some version of life extension happen. Other parts of the world are de-facto firmly in the survival mode.
The societal problems this would exacerbate are already very hard to fix. It's not that the status quo is good, but immortality is solving a non-problem that may make the real problems intractable.
I think a lot of it is rationalization. People figure they are going to die and so the brain comes up with reasons to rationalize why it should be so - extended life must suck and so on. I find it a little sad too but I think attitudes will change if it becomes real.
This is very important. It seems that the automation of labor is underexplored due to political and economical inertia (I know that industrial robots do somewhat improve, but they still are very costly, very proprietary and require experts to use them. Could this be done another way? Companies like Rethink Robotics show that maybe the answer is "yes").
There needs to be more fresh thinking in this space.
There was an ambitious NASA project 35 years ago: a self-replicating lunar factory http://www.nss.org/settlement/moon/library/1982-SelfReplicat... . The engineers tried to design a manufacturing system aiming for almost full parts closure.
The project was too ambitious (e.g. it looked optimistically at AI's capabilities), and didn't went past design study stage then;
But maybe now, 35 years later, the technology is good enough for something similar to be viable?
Actually there are at least two decades-old branches of computer science/mathematics that have formulated precise definitions of AI, and proved many theoretical results that gave way to lots of practical applications. These branches of CS are called "Reinforcement Learning" and "Universal AI".
While Gwern has already mentioned Reinforcement Learning, UAI is a less known (but even more rigorous and well received) mathematical theory of general AI that arose from Marcus Hutter work [1].
My point here is how can one say that there is no definition of AI when there are several precise mathematical definitions available with many theorems proven about them?
You are confusing narrow AI for AGI. None of those things have proved anything practical about what an actually achievable AGI would look like, rather than some theoretical construct that is provably incomputable.
Because AIXI_tl has failure modes (it doesn't model itself as being embedded in its environment so it can't ensure its own survival) demonstrating that any approach which is just a weaker version of it will have those same problems.
> That said it is also not computable with finite time or resources, so it is unclear what relevance it has to practical applications.
You can define it as space or time-bound and then it's finite but still intractable.
I agree with the first sentence, but I'd like to note that there are practical (though weak) approximations of AIXI that preserve some of its properties, and while not turing-complete, prove to be more performant when compared to other RL approaches on Vetta benchmark. See [1].
Also there is a turing-complete implementation of OOPS, a search procedure related to AIXI that can solve toy problems, programmed by none other than Jurgen Schmidthuber 10 years ago [2]
Even more important: there is a breadth of RL theory built around MDPs and POMDPs. There are asymptotical, convergence, bounded regret, on-policy/off-policy results, etc. Modern practical Deep RL agents (the ones DeepMind is researching) are developed on the same RL theory and inherit many of these results.
From my POV it looks unfavorable to researchers that produced these results over decades of work when the comment's grandfather (and grand-grandfather) write that there is no definition and theory about AI, and that AI is like alchemy.
Meta- reinforcement learning could prove to be such breakthrough, see [1],[2].
Also next generation ASIC accelerators (Google's TPU, Nervana) can give 10x increase in NN performance over a GPU manufactured on the same process, with another 10x possible with some form of binarized weights, e.g. BNN, XNOR-net.
There are also interesting techniques to update the model's parameters in a sparse manner.
So, there certainly is a lot of room left for performance improvements!
Finally there is a trend of using recurrent neural network as a top component of the Q-network. Perhaps we will see even more sophisticated RNNs like DNC and Recurrent Entity Networks applied here.
Also we'll see meta-reinforcement learning applied to a curriculum of environments.
Is he a CEO though? Wikipedia and other press articles say that Hassabis is the CEO: https://en.wikipedia.org/wiki/DeepMind and Suleyman is Chief Product Officer, the head of applied AI at DeepMind.
> You have a lot of solid points, but note that Linux is currently being used by SpaceX in an even more safety-critical aerospace setting
I disagree with the 'more' - how many lives are at risk with a SpaceX failure vs. a self-driving car failure? This is even without multiplying by number of users.
DeepMind looks like a hilariously wrong project to criticize because it is a true moonshot, something very different from the majority of other SV projects.
If hiring hundreds of PhDs to create a general purpose learning agent, all while publishing all the intermediate results in freely available papers isn't a moonshot with socially beneficial outcome, then I don't know what is.
Also note that DeepMind went even further than that, there is DeepMind health division aiming at using this technology to help doctors and patients directly.
If I were the author I'd choose some social media unicorn or an ad network as an example of inherent misallocation of human talent.
I think part of the issue is that the results of DeepMind, while enormously impactful, do not have the sort of tangible quantity that the Manhattan Project or the Apollo Program did. It's easy for critics to dismiss them when the most relatable thing to the layperson is "This thing plays Go". It's a common lament in the scientific community, especially in the theoretical space. I don't see an easy answer for it, other than "Wait for it."
To me, the more apt comparison isn't say, DeepMind to the Manhattan Project, it's DeepMind to the early physics experiments, or Chicago Pile 1. You can imagine in the future, a letter being written much like Einstein's to FDR, pointing to an early and modest project saying "because of this, we can now create <X>".
(I'm ignoring the wartime necessities and issues surrounding that letter in this example as they're outside of the discussion.)
> If hiring hundreds of PhDs to create a general purpose learning agent, all while publishing all the intermediate results in freely available papers isn't a moonshot with socially beneficial outcome, then I don't know what is.
I'm sorry, I don't buy it. What exactly are the social benefits of general purpose learning agent?
It's an impressive technical and scientific challange, agreed, but most applications that I can immediately think of are harmful to society (reduction of white-collar jobs, increased potential for surveillance and profiling, etc) - so how would such an agent be used to actually improve society?
A general purpose reinforcement learning (RL) agent is a machine that can be taught to perform any task from a very wide range of tasks via sparse rewards given by a human or software trainer.
The agent can, like any software, be snapshotted, saved, loaded and copied, creating as many identical agents as needed (given hardware, of course). Agents can and will be trained to perform various tasks, and their snapshots will be sold or made available for download over the Internet.
By saying that your main concerns are technological unemployment of white-collar demographic and increased state surveillance you make it clear that your views reflect that of an upper-middle class western person. On the global scale affluent westerners are a minority.
So, How would such an agent be used to actually improve society?
Consider universally valued, life-critical services: healthcare and education. Only the western people have access to high-quality medicine and education due to a whole lot of reasons (global economical inequality, a very long and hard path to become a doctor or a professor, a very long time needed to establish the necessary social institutions, lack of social stability outside the west, ...).
If we had a general RL agent we could train several variants of it to perform high-quality work in the fields of Diagnosis, Radiology, Paediatry etc. We could also train artificial education agents for many subjects.
The training needs to only be done once. Given sufficiently powerful mass-produced hardware (smartphone SoCs with Nervana-like NN accelerators?) these agents could be given almost for free to billions of people that wouldn't be able to afford such services in any point of their lives otherwise.
How could one be against giving essential high quality services to every human with a smartphone?
And if even that is not enough to justify the utility of RL agents, then consider how much progress in molecular biology and medicine could be done if thousands of agents trained to do life science research worked around the clock to push the state of art further. How many people with debilitating diseases could be cured by such an effort?
And then consider how we could make our currently-crumbling cities and infrastructure permanently well-attended by RL agents inside simple robots. The world certainly could use more smart attention everywhere. I guess the quality of life in such a world would be remarkably different.
An example was already given. Medical professionals screw up all the time. Having a highly intelligent entity capable of providing advice and even oversight could greatly speed diagnosis and reduce mistakes in treatment.
Asking how a general purpose learning agent could be useful seems kind of like asking how an intelligent person could be useful. The ways are countless.
The social benefits are that our knowledge develops. You can argue that that is not a good thing because of the social upheaval it causes, but everyone who has argued that since we were cavement have been proved wrong.
For example, humanity's current best attempt at sustainable society is an unstable mix of democracy and capitalism, and in many countries that isn't working out too well - particularly for blue collar workers, but increasingly for your white collar workers too.
It isn't inconceivable that deep mind could design a better political system for the US, for instance, that resulted in broad consensus instead of virtual civil war. Or design a fairer tax system that meant more people could have fulfilling and enjoyable lives.
Whether deep mind's masters would apply it to those questions is moot, but the parent's point is that the huge resources being poured into Facebook makes it a much better example of the squandering of the efforts of our brightest and best.
Note that Facebook also has a formidable AI research group called FAIR and they are pursuing goals close to DeepMind's, while openly publishing their results and tools.
There is a lot of social media unicorns that don't contribute much to research which are not Facebook.
Who knows, maybe there is no real need for a dozen of global social media companies that provide roughly similar features to the same users?
> but most applications that I can immediately think of are harmful to society (reduction of white-collar jobs, increased potential for surveillance and profiling, etc)
Those are not applications, those are side-effects to _some_ applications.
Also note that DeepMind went even further than that, there is DeepMind health division aiming at using this technology to help doctors and patients directly.
Thing is, there is plenty of existing AI research that could hypothetically revolutionize medicine. Even in the eighties there were expert systems that outperformed doctors in certain areas. The issue with it all isn't that the algorithms are not fancy enough or accurate enough yet, it's the practical application.
I believe the biggest challenge of today's AI research is making AI and ML accessible. Not just for consumption, but for actual training and open ended use.
I don't see Deep Mind doing much in *that" regard.
It is not impossible. There is a combination of a special neural network architecture and strong sparsity-inducing regularization that makes it possible to learn equations from dynamics dataset: https://openreview.net/pdf?id=BkgRp0FYe