John Carmack, programmer extraordinaire, and developer of seminal titles like "Doom" and "Quake" has said "Hasta La Vista" to his colleagues at Oculus to to set out for a new challenge. In a Facebook post (https://www.facebook.com/100006735798590/posts/2547632585471243/) he declares that he is going to work on artificial general intelligence.
What are the chances he can pull it off, and what could go wrong?
John Carmack Steps Down at Oculus to Pursue AI Passion Project `Before I get too Old':
Legendary coder John Carmack is leaving Facebook's Oculus after six years to focus on a personal project — no less than the creation of Artificial General Intelligence, or "Strong AI." He'll remain attached to the company in a "Consulting CTO" position, but will be spending all his time working on, perhaps, the AI that finally surpasses and destroys humanity.
AGI or strong AI is the concept of an AI that learns much the way humans do, and as such is not as limited as the extremely narrow machine learning algorithms we refer to as AI today. AGI is the science fiction version of AI — HAL 9000, Replicants and, of course, the Terminator. There are some good ones out there, too — Data and R2-D2, for instance.
[...] Carmack announced the move on Facebook, where he explained that the uncertainty about such a fascinating and exciting topic is exactly what attracted him to it:
When I think back over everything I have done across games, aerospace, and VR, I have always felt that I had at least a vague "line of sight" to the solutions, even if they were unconventional or unproven. I have sometimes wondered how I would fare with a problem where the solution really isn't in sight. I decided that I should give it a try before I get too old.
Skynet? Singularity? With great power comes great responsibility. Can he do it? Should he?
(Score: 2) by Rich on Friday November 15 2019, @06:38PM (2 children)
Well, with the brain it's not the classic punched-card-deck-to-line-printer flow that was imagined when computers were new. Cf. the quote of Asimov's AI definition elsewhere, or the ST:TOS "The ultimate computer" script, or an original 1951 vintage brochure about the Ferranti Nimrod computer that was recently given to me. There is no begin, and no end. No task about objects, except for permanently regulating the neurotransmitters - and going to extreme lengths of doing so.
A decent car analogy would be that of the ECM happily idling. If the revs drop, it opens the idle throttle a bit, and so on. This could be easily done through a single-abstraction-level neural network. Of course there are several factors to regulate, temperatures, pressures, and so on, that have to be balanced out. To improve smoothness, the scientists add multi-layer abstractions, so the ECM can optimize its behaviour. Now imagine that they add another factor: tank level, and it is added with a really strong factor if it nears empty. Would, with stacked abstractions (and how many of them) the car eventually figure out that it has to drive to a petrol garage? (Or just set the indicators and honk its horn a bit once it sees one...).
(Score: 2) by acid andy on Monday November 18 2019, @10:00AM (1 child)
For it to figure this out on its own, it of course needs to develop (or be given) the ability to model future scenarios and assess the potential for them to be rewarding (i.e. This is called a petrol garage. I gain fuel at the petrol garage. I need to gain fuel. If I drive to the petrol garage, I will be at the petrol garage. Therefore, I will gain fuel). Some sort of capability for language processing might help here, because a human would probably be taught some of those facts in words and be able to reason about them by talking to themselves, but it wouldn't have to be English--the machine could be taught using logical statements.
I think most neural networks at the moment are trained based on the correctness of their immediate output, whereas in the above, the AI needs to be able to anticipate the delayed gratification of a potential future reward. I wonder if there have been any approaches yet to build that into machine learning.
Welcome to Edgeways. Words should apply in advance as spaces are highly limite—
(Score: 2) by Rich on Monday November 18 2019, @02:25PM
Precisely.
I have the suspicion that there is not a clear cut algorithmic process (like for the linguistic processing you mention), but a general feedback topology, yet undiscovered, which enables neural networks to operate on this level. There are hardcoded starting conditions, but I assume the ability to abstract gets trained. The ability to assess future potentials might be an extension of the idling mind (experiment: try to think of nothing), which can "lock" on something, which in the best case is the solution for a complex problem, and in the worst case is an earworm of a really crappy song.