Tool AIs want to be agent AIs
gwern.net
gwern.net
https://www.amazon.com/Superintelligence-Dangers-Strategies-...
It has helped me make informed, realistic judgments about the path AI research needs to take. It and related works should be in the vocabulary of anybody working towards AI.
"There’s no sense in being precise when you don’t even know what you’re talking about."
Bostrom is one of those medieval cartographers drawing fantastical beasts in the blank spots of continents which he has never visited.
The analogy I like to use for our understanding of AI is alchemy. We threw Sir Isaac Goddamned Newton at chemistry and he couldn't make forward progress, because the tools were not precise enough. Similarly, we just don't understand minds enough yet to formulate sensible questions about AI.
This doesn't bother Bostrom. He builds castles of thought in the air, and then climbs up into them.
We most certainly do understand enough to formulate questions, and even answer some of them. The problem is that the people making the most noise (Bostrom et al) are not trained in neuroscience or computer science, nor do they have practical experience in deploying working systems. They have about as much training and expertise as science fiction writers, and the end result is similar.
There is also the Advances in Cognitive Systems journal and associated conference, which is AGI even if they prefer to avoid that specific acronym:
And there are always a small but growing number of papers related to AGI in each AAAI conference.
2. Setting aside cogsys (CogSci is a whole different beast than AI/ML in computer science), the only impactful journal/conference you've listed is AAAI.
3. Papers are also typically incremental and all of the AGI papers I've seen in AAAI (and there have been very few) are no different, tackling some small theoretical subproblem.
4. I'm not saying the research is useless. It's very valuable. But it's is pure theory right now, and to claim it has insights for us about what AGI would actually look like is very premature.
This is incorrect. Among his degrees, Bostrom has a master's in computational neuroscience. His arguments have also convinced PhD neuroscientists (such as Sam Harris) and computer scientists (such as Stuart Russell) about the potential dangers of AI.
It would be a different story if he published in the meantime, but he did not. Nor did he work on practical projects in industry or anything. He shifted gears to philosophical speculation which he has done since.
There are good arguments to be made against Bostrom's Superintelligence, but malignments and surface analogies aren't appropriate. Please engage the ideas, not the man.
Read what I wrote again please. I think you misinterpreted.
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?
That said it is also not computable with finite time or resources, so it is unclear what relevance it has to practical applications.
> 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.
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.
1. https://www.jair.org/media/3125/live-3125-5397-jair.pdf 2. http://people.idsia.ch/~juergen/oops.html
The largest publicly available research datasets for machine translation are 2-3 million sentences [1]. Google's internal datasets are "two to three decimal orders of magnitudes bigger than the WMT corpora for a given language pair" [2].
That's far more data than a cell phone's translation app would receive over its entire lifetime. Similarly, the amount of driving data collected by Tesla from all its cars will be much larger than the data received by any single car.
This suggests that most learning will happen as a batch process, ahead of time. There may be some minor adjustments for personalization, but it doesn't seem like it's enough for Agent AI to outcompete Tool AI.
At least so far, it seems far more important to be in a position to collect large amounts of data from millions of users, rather than learning directly from experience, which happens slowly and expensively.
This is not about having a human check every individual result. It's about putting a software development team in the loop. Each new release can go through a QA process where it's compared to the previous release.
[1] https://github.com/bicici/ParFDAWMT14 [2] https://research.googleblog.com/2016/09/a-neural-network-for...
> Similarly, the amount of driving data collected by Tesla from all its cars will be much larger than the data received by any single car.
Well, Tesla benefits from the fact that its cars are already agents, acting in the real world. So using it as a counter-example doesn't really work... You would need to imagine some sort of counterfactual Tesla which eschewed any real-world deployment or simulators. Which no one would ever do because the usefulness of trying to create a self-driving car database without agents or reinforcement learning is so obviously zero.
And anyway, most of Tesla's data, however, will be totally useless. You can train a simple lane-following CNN with a few hours of footage (as Geohotz demonstrated), but another million hours of highway driving is near-useless and doesn't get you much closer to self-driving cars (as perhaps Geohot also inadvertently demonstrated). What you need is the weird outliers which drive accident rates. So for example, the Google self-driving car team doesn't depend purely on collected data even from its agents, but simulates very rare scenarios as well. (If I may quote myself: "You need the right data, not more data.")
> This suggests that most learning will happen as a batch process, ahead of time. There may be some minor adjustments for personalization, but it doesn't seem like it's enough for Agent AI to outcompete Tool AI.
Even if we imagine that the dataset covers all possible sequences and doesn't need any kind of finetuning the loss, the agent AIs in this scenario are still going to benefit from actions over (going by section): 'internal to a computation', 'internal to training', 'internal to design', and 'internal to data selection'. In fact, if you want to have any hope of running effectively over extremely large datasets like hundreds of millions of sentences, you are going to effectively require some of those to keep training time tractable and runtime cheap (even Google can't afford a million TPUs for Google Translate). For example, regular hyperparameter optimization using a few dozen examples doesn't look very good when it takes a month on a GPU cluster to train a single model, but using a meta-RL RNN to do some transfer learning and pick out near-optimal architectures & hyperparameters in just a few iterations looks pretty nice.
I haven't seen reports that any self-driving cars learn to drive better in real time. Instead the data is collected and used to improve the software. There might be frequent software releases (perhaps even every day) but this isn't quite the same thing.
An offline learning process seems considerably more predictable. The release engineers can test each release to see how it would behave under various extreme conditions.
It's not clear that there's a compelling reason to switch to true online learning. Self-driving cars will improve over time, but not to the point where new training data gathered at 9am should make a significant difference to how the car drives at 6pm. Probably this year's model will drive better than last year's, but the changes are likely to be gradual.
As you say, there are diminishing returns to gathering more data, which suggests that in many domains, updating the software in real time won't be a competitive advantage. Unless it's a system that's reacting to breaking news, the delta in training data gathered in the last 24 hours is unlikely to make a huge difference.
Most self-driving cars probably use some variant of local search[1]. In this case it probably doesn't matter if the agent learns offline or online.
[1] https://en.m.wikipedia.org/wiki/Local_search_(optimization)
Tool AIs will never "want" anything because the meaning of want will be completely foreign.
I especially liked this passage. It connects many ideas onto one theme:
> CNNs with adaptive computations will be computationally faster for a given accuracy rate than fixed-iteration CNNs, CNNs with attention classify better than CNNs without attention, CNNs with focus over their entire dataset will learn better than CNNs which only get fed random images, CNNs which can ask for specific kinds of images do better than those querying their dataset, CNNs which can trawl through Google Images and locate the most informative one will do better still, CNNs which access rewards from their user about whether the result was useful will deliver more relevant results, CNNs whose hyperparameters are automatically optimized by an RL algorithm will perform better than CNNs with handwritten hyperparameters, CNNs whose architecture as well as standard hyperparameters are designed by RL agents will perform better than handwritten CNNs… and so on. (It’s actions all the way down.)
This is the general trend. Going meta, through RL. Optimize the optimizer, learn gradient descent by gradient descent.
As far as 'learning gradient descent by gradient descent' goes, I think it's still an open question if a tiny RNN actually can improve meaningfully over ADAM etc :) I don't recall the paper showing any wallclock times, which suggested to me that the RNN was way slower even if it trained faster. The individual parameter adjustments are so low in the hierarchy of actions that the value may be minimal compared to higher up like architecture design.
It's tough enough to get value out of A.I. that this trick should not be left on the table. Thus Tool A.I.s need to be Agent A.I.s to maximize their potential.
Yes, we should be aware of the limitations and market instability of tool AI, but I think it's unjustified to suggest that tool AI is essentially impossible ("highly unstable equilibrium") and all we can hope to do is solve value alignment.
Only somewhat. There are not repeated practical demonstrations, nor is there any really good theoretical reasons, to think that value alignment mechanisms would be seriously and systematically detrimental to performance & cost in the same way that we have for reinforcement learning/active learning/sequential trials/etc. You can reuse my little trivial proof to argue that UFAIs >= FAIs, but of course they could just be == on intelligence. Then FAIs can be a relatively stable equilibrium because they are not instantly outclassed by any fast-growing UFAI. Contrast that with Tool AIs where everyone who cooperates is at enormous disadvantage to one defector, and defectors have large and growing rewards.
> But that doesn't make us conclude that value alignment is impossible, just hard to achieve soon enough in the real world.
I'm not sure anyone seriously thinks that value alignment will be easy to achieve, much less at sufficient scale.
(It's weird to try to reason that 'value alignment looks hard for the same reason that Tool AI looks hard; but we can do value alignment, thus, we can do Tool AI as well'. Maybe we can't do either? We desperately want value alignment to work, but the universe doesn't owe us anything.)
Some vaguely similar arguments:
1. The race by organizations to develop AI is slowed >10% by worrying about friendliness. In a big efficient global marketplace, the winning organizations will almost certainly be ones who ignored friendliness.
2. A seed AI which is amoral can take immoral actions not available to a moral AI. Even a tiny advantage amplifies over time because of the exponential nature of recursive self-improvement. So the first super intelligent AI will be immoral.
3. Nations who obtain nuclear weapons have immense power. Even if most nations decide not to obtain them, or only obtain them for peaceful purposes, at least a few will get them and then use them to take over the Earth. Therefore, the Earth will be taken over completely by a nuclear-armed nation.
All of these arguments have some merit, but none are strong/inevitable. They all require quantification and comparison against countervailing forces (e.g., the effect of laws, the moral motivations of nearby human programmers, the counterthreat provided by many allied nations/AIs).
> (It's weird to try to reason that 'value alignment looks hard for the same reason that Tool AI looks hard; but we can do value alignment, thus, we can do Tool AI as well'. Maybe we can't do either? We desperately want value alignment to work, but the universe doesn't owe us anything.)
I'm not making any sorts of claims like that.
What is still unclear -- to me at least -- is the technical challenges that lie ahead of this "neural networks all the way down" approach. I get the impression we'll need quite a few breakthroughs before usable Agent AIs are a thing. Insights on the order of importance as, say, backpropagation and using GPUs.
I think a 10x-100x increase in efficiency is going to come soon, based on more research into efficient hardware and efficient models. New algorithms, new hardware and especially a lot of money and talent invested in research are going to power the next step.
So, there certainly is a lot of room left for performance improvements!
1. https://arxiv.org/abs/1611.05763 2. https://arxiv.org/abs/1611.02779
Also it seemed like the author assumes great technological advances in AIs, but not in biology. If we're gonna dream shit up why not dream that brains in the future will be 10,000 times as dense and computers won't be able to keep up except as tools.
"Suppose, says Searle, that this computer performs its task so convincingly that it comfortably passes the Turing test: it convinces a human Chinese speaker that the program is itself a live Chinese speaker. To all of the questions that the person asks, it makes appropriate responses, such that any Chinese speaker would be convinced that they are talking to another Chinese-speaking human being."
> If we're gonna dream shit up why not dream that brains in the future will be 10,000 times as dense
Because... that is not a thing which is happening. And deep learning and AI progress are things that are happening. (Quite aside from the many issues with your proposal, like a brain 10kx as powerful due to 10kx density would probably break thermodynamic limits on computation and of course cook itself to death within seconds.)
Like i mentioned, i don't know all the AI terminology but isn't there an unresolved argument that ai architecture in the short run can't mimic biological decision making, and so the decisions will always be different/ tasks for which tool AI will be better to help the biological decision making processes?
That's one perspective on anatta, but not a particularly useful one. You may find this interesting:
http://www.accesstoinsight.org/lib/authors/thanissaro/selves...
If you've ever taken an introductory course on Buddhism, you've
probably heard this question: "If there is no self, who does the
kamma, who receives the results of kamma?" This understanding turns
the teaching on not-self into a teaching on no self, and then takes
no self as the framework and the teaching on kamma as something that
doesn't fit in the framework. But in the way the Buddha taught these
topics, the teaching on kamma is the framework and the teaching of
not-self fits into that framework as a type of action. In other
words, assuming that there really are skillful and unskillful
actions, what kind of action is the perception of self? What kind of
action is the perception of not-self?