AlphaGo solved one of the hardest challenges which we believed to be very far out in the future before it came along. I'd say the field is on the right -- and very exciting -- track; people just need to learn to approach it in a clear-headed manner.
Good on IEEE though for welcoming opposing viewpoints.
Also the objective was different - they tried to train a network that would play well against any opponent. It should be easier to train a network that exploits some very narrow weakness in specific AlphaGo version.
For example, my phone consistently beats me at chess. But I found a sequence of moves in one opening where is always makes the same mistake, and I after that I can win easily.
1. how many handicaps can AlphaGo give to human players, while still come out winning? the idea of handicap is basically the stronger player lets the weaker player a few "freebie moves" before the game proper starts. For reference, a typical professional player can give 2 to 3 handicaps to the strongest amateurs.
2. Can AlphaGo defeat a group of professional who are also given unlimited time to think and discuss?
If I recall correctly this was part of the commentary during the matches so I can't find a good source for it right now.
It's a response to people who think because of AlphaGo that we are on the cusp of achieving true / strong / general AI.
> An AI that solves a lot of problems better than humans, and transfers that learning across many problems with relative ease, is getting close to strong AI.
We're not close. See Yann LeCun's statement on AI after AlphaGo [1] and the HN response [2]
[1] https://www.facebook.com/yann.lecun/posts/10153426023477143
You probably haven't studied much or any AI. In my opinion, those who haven't researched a subject don't understand the details and are likely to make more inaccurate predictions than experts.
In the past 50 years, AI has seen hundreds of small successes in narrow tasks that used to require humans. But none so far has shown the potential to scale up, generalize, and serve tasks other than the narrow one for which it was designed. Like IBM Watson, AlphaGo too is likely to be consigned to the AI scrapheap in the sky.
BUT... the deep net technique used by AlphaGo shows more promise to solve the remaining unsolved AI tasks than any AI method before it. Yes, we still don't know DL's limits, like whether it can integrate one-shot learning, or build and reuse a diverse knowledgebase, or transfer specific methods to solve new more general problems. But as of right now, it's shown greater promise to solve novel weak AI tasks than any past technique I've seen. The author overlooks that potential deliberately and provocatively, and IMO, pointlessly.
Can DL scale up into strong AI too? I think the important thing here isn't that the answer isn't obviously yes (as the author posits), but that the answer isn't obviously no. And in the 50+ year quest for strong AI, that's a first, at least for me.
It would be self-serving if I lied in an interview and said I was qualified for a job making robots if I'd never had any experience doing so.
This is just a short title, and ITT people didn't read the first paragraph.
So it uses a strong claim — i.e. AlphaGo is not AI — then subsequently changes it to a smaller claim — AlphaGo is not AGI.
Reading the entire article does not change the fact that this can be seen as a technique to attract readers. Claim rare X; change X to be a subset of Y; then claim it was really new X all along. Moving the goal post.
With respect, you are being pedantic.
I think that's what the author was thinking when he wrote the title, and I'm giving him the benefit of the doubt.
I'm being flexible, not picky. Disagree with me? Fine, I really don't care.
I could guess that he didn't use a qualifier because there are many different ones and none universally accepted. Or, maybe he wanted to reach a non-tech audience who wouldn't know the difference between strong AI and AI.
There are plenty of good reasons, and reading the article clarifies the author's meaning.
Perhaps we need a new term to refer to "true" AI, preferably one that matches people's general notions of what an AI should be. "Synthetic Thought" maybe?
> Several research labs are now trying to go further into acquiring grammar, gestures, and more complex cultural conventions using this approach, in particular the AI Lab that I founded at Aldebaran, the French robotics company—now part of the SoftBank Group.
This is very interesting. In the human realm, there is the celebrated case history example of Helen Keller, who progressively developed her blindness and deafness over a period of years. She would arguably "pass" whatever tests that were administered for evaluating a strong AI.
However, walking backwards from that, towards case histories where patients were born deaf and blind, they also could pass those kinds of tests. But add more sensory disabilities, and it becomes murkier. Has anyone performed a detailed survey of medical histories? Is there a point beyond which lack of a specific set of senses leads to sapience incapacitation 100% of documented instances in humans? Could that be a promising lead to build a virtual environment that simulates a minimum threshold of senses?
Also… it seems weird to consider having “human” inputs embodied but having direct access to information stored elsewhere as not — surely there’s disk platters, SSDs and other devices existing in the real world, no?
My cursory search says hearing and sight are not necessary. Leaving us with smell, taste, and touch. Smell and taste are closely interlinked, so there is a possibility that a simulated smell-taste and touch are all that is required to "bootstrap" the first AGI, so to speak.
Your point recalls to mind the scene in Battlestar Galactica (the reboot) where Cavil fulminates "I don't want to be human". [1] I suspect once we boostrap the first AGI, you are absolutely correct, we will rapidly add more senses to the AGI's repertoire. But we only have one data point for a template to build AGI off of, ourselves, so I also suspect the first AGI will somewhat resemble us. If so, and if the embodiment principle is one of the right routes to take, then it makes sense to me to simulate what we empirically determine as the minimum sense set, and then expand from there. Walk before we run, so to speak.
It wouldn't hurt for someone investigating the embodiment principle to build for as many senses as they can envision, though. Many eyes, and all that, for attacking the problem space. I'm just expressing an uninformed gut feeling here, as I am not in the AI research space.
Playing go or chess would have been classified as activities clearly requiring intelligence in the recent past. We're just moving the goalpost whenever AI advances. At some point, it'll probably divide mankind into fractions which either want or don't want to grant certain rights to AI (possibly obsoleted by the great AI uprising of 2025-20-11 6:45:23am to 2025-20-11 6:45:25am)
The 'real world' is not nearly as nicely ordered and specified as a turns based game is. So a computer will be ideally positioned to take advantage of its ability to:
- remember perfectly, forever
- execute algorithms at extreme speed
- in parallel
So you end up with a race that is comparable to a sparrow against a jet airplane. Technology can and does beat biology regularly in all sorts of domains.
Slowly but surely computers are making their way into the real world: their programmers - and by extension the systems - learn how to adapt to the messiness of interacting with life as we know it.
Just like you'd first get kids to play in a playground (the games) and then let them loose in the real world (self driving cars and other much harder tasks).
In the end, all of these amount to a slow but steady march towards human equivalence on a large range of tasks, and excelling humans on plenty of them.
The hard AI question is one where we ask ourselves whether or not there will one of these days be a system that is better than humans on essentially all tasks without exception, including learning, creativity and so on. Once you reach that level all bets are off, until then the slow-but-steady march continues.
The difference is evolution vs revolution, but we've already seen enough progress in limited AI domains that we could conclude we're already seeing a minimal revolution. Whether it will result in a major one is still an open question, it could easily be that progress will at some point plateau.
Like Scott Aaronson said, "You can look at any of these examples -- Deep Blue, the Robbins conjecture, Google -- and say, that's not really AI. That's just massive search, helped along by clever programming. Now, this kind of talk drives AI researchers up a wall. They say: if you told someone in the sixties that in 30 years we'd be able to beat the world grandmaster at chess, and asked if that would count as AI, they'd say, of course it's AI! But now that we know how to do it, now it's no longer AI. Now it's just search."
Not by the normal definition of intuition. It reasons about millions of moves.
"During the match against Fan Hui, AlphaGo evaluated thousands of times fewer positions than Deep Blue did in its chess match against Kasparov; compensating by selecting those positions more intelligently, using the policy network, and evaluating them more precisely, using the value network—an approach that is perhaps closer to how humans play. Furthermore, while Deep Blue relied on a handcrafted evaluation function, the neural networks of AlphaGo are trained directly from gameplay purely through general-purpose supervised and reinforcement learning methods"
The other part that increases its strength is the positional evaluation through the value network. It's ALSO intuition based on which points are going to belong to which player.
Those two combined give it a high dan level, maybe around 8d.
Only the last component (Monte-Carlo tree search) is not quite as similar because it uses random game roll-outs to ALSO give an evaluation.
AFAIK, value network evaluation is an overestimate of territories, while the MCTS is an underestimate so using both it's stronger than professional level.
The levels I put is the version that played Lee Sedol. The version that went 50-0 on the Internet is clearly stronger and might already be professional level without MCTS.
He makes even a stronger curious claim: That "This realization is often called the “embodiment problem” and most researchers in AI now agree that intelligence and embodiment are tightly coupled issues."
That makes it sound like it is a consensus between researchers, that strong AI requires a robotic body. Which I doubt.
I'm intrigued by the potential of deep nets to 'simulate' the bases for symbolic grounding using something like thought vectors (TVs), especially if they can be revised and augmented through personal experience. Bootstrapping one's knowledgebase using someone else's rich set of TVs may go a long way for an AI to become grounded vicariously.
Helen Keller had smell, taste, touch, and proprioception all functioning. In what way did she have no senses but a teletype?
It occurred to me later that one can argue that Helen Keller, like the rest of us, was born with a brain shaped by hundreds of millions of years of interacting with the environment. I do not doubt that interaction is probably an essential part of us getting to AI, but I doubt that physical embodiment is the only way to do it.
Alpha Go is an AI project with historical relevance, that had a well specified and measurable goal, which is beating humans in a game that requires to develop an intuition for moves rather than just reading moves in advance.
Having said that I don't understand what's the point of trying to delegitimize the significance of Alpha Go. Implementing what we call "intuition" is not a minor thing.
In contrast, I encountered the Pepper robot in a mall in SF. It was almost impossible to interact with it, so if there is going to be a critique, why not starting from one of his own projects.
Intelligence implies a lot more than pattern recognition. Current state is just a better way to treat large quantity of data. The smart is to feed the computer the right way with good data, iterate through different ways the computer can guess a model from the data and finally interpret the final results. Computers doesn't know how to do that. Yet.
That's exactly how AlphaGo was trained.
The author and his editors deliberately conflate the two terms in the headline, and clarify after about a paragraph. Sure, AlphaGo is not strong AI. And Abraham Lincoln is dead. And there are lots of other things everyone knows that don't deserve to be in the news.
Jean-Christophe Baillie wrote this piece. He has done some work in AI and computer vision, and he had the opportunity here to write a piece that reflected that expertise. Instead he made a rather facile point that is similar to saying "Graffiti isn't art." That argument won't lead anywhere interesting.
He rehashes the Chomsky-ian argument about meaning. That AI's won't understand meaning until they can connect representations to the world. Then he makes the point that this requires embodiment, that they have a physical body in space through which they connect to the world. I don't believe either of these are necessary requirements for AI.
An AI that solves a lot of problems better than humans, and transfers that learning across many problems with relative ease, is getting close to strong AI. It doesn't really matter whether the machine constructs meaning to itself or not. It doesn't have to be humanlike to be strong. Secondly, AIs can train perfectly well in virtual environments. They don't have to be embodied as robots to be considered strong. We can model the world with enough complexity to give AIs problems that, if solved, might define them as strong.