Artificial Intelligence: The Revolution Hasn’t Happened Yet (2018)
medium.com
medium.com
That's the wrong problem with AI. The trouble with AI is that it still sucks at manipulation in unstructured situations and at "common sense". Common sense can usefully be defined as getting through the next 30 seconds of life without a major screwup. At, at least, the competence level of the average squirrel. This is why robots are so limited.
If we could build a decent squirrel brain, something "higher level" could give it tasks to do. That would be enough to handle many basic jobs in unstructured spaces, such as store stocking, janitorial, and such. It's not the "high level reasoning" that's the problem. It's the low-level stuff.
A squirrel has around 10 million neurons. Even if neurons are complicated [1], somebody ought to be able to build something with 10 million of them. Current hardware is easily up to the task.
The AI field is fundamentally missing something. I don't know what it is. I took a few shots at this problem back in the 1990s and got nowhere. Others have beaten their head against the wall on this. The Rethink Robotics failure is a notable example.
The real surprise to me is how much progress has been made on vision without manipulation improving much. I'd expected that real-world object recognition would lead to much better manipulation, but it didn't. Even Amazon warehouse bin-picking isn't fully automated yet. Nor is phone manufacturing. Google had a big collection of robots trying to machine-learn basic manual tasks, and they failed at that.
That's the real problem.
[1] https://www.sciencedirect.com/science/article/pii/S089662732...
We likely have a very warped view of what intelligence is, because the most prominent examples of it have been aggressively honed over an extremely long period of time to be good at tasks crucial to their survival, such as effectively navigating a 3D environment. We consider art to be a difficult and complex task, and making a sandwich to be a simple one, but that's because our particular brand of intelligence is optimized toward the latter.
Not just that, but it's grown on a body which has been optimized for survival during a few billion years; and that body is built on cells that have evolved to survive hostile environments, and those cells are built with self-replicating molecules, evolving from complex chemical reactions in several changing environments, that competed with and displaced other less-successful self-replicating molecules that disappeared.
Each of those layers provides a degree of adaptability and self-healing that is extremely hard to replicate. And if we managed to reverse-engineer and replicate one of those layers, it would still be missing all the layers below.
Our best hope to create fully independent agents will come from re-adapting and controlling biological entities, not from tools built from the ground up with current engineering techniques.
I can only speculate but it certainly is for a reason that certain parts of our body are vegetativelly controlled whereas others are under the active control of our consciousness.
If you step on a nail, the first reaction comes from vegetative stimulus, later your consciousness processes that information. A squirrels neuronal network is also separated in that way. That may be a reason.
And second, AFAIK AI still doesn't 'think' in concepts, it has no notion of the 'world'.
And third: The capability of reproduction and acting accordingly may be another thing.
There's an old (and sometimes forgotten) idea in AI that perhaps things we think are simple, like vision and control (robotics), are actually incredibly complicated and took millions of years to evolve.
Whereas things we think are complicated, like playing Go or picking stocks or computer programming, are actually quite simple to learn.
This would be counter-intuitive but---as you observed, and taking my argument recursively---common sense might be much more difficult to get right than obscene pathological thinking.
Anyway, I've always thought a good startup would be to automate away Silicon Valley using AI. It's so punk rock that a lot of disillusioned smart techies would join under this banner. A collaborator of mine has already used AI to do high-level bug finding in blockchain code.
I've worked on a number of projects where it looked simple to automate from the outset and impossible in retrospect.
This is an example paper that, for instance, mathematically blurred the distinction between offshoring and automation:
https://talkbusiness.net/2017/07/ball-state-study-automation...
There was another paper (that i cant find right now) that basically surveyed people about how creative they thought their job was and just assumed that creativity was inversely proportional to automatability.
Ironically I think a widespread belief in this myth helped, among other things, lead to the trucker shortage. Who wants to join a profession with a high barrier to entry that they believe will be automated soon?
And it's understanding what the right thing is to build that's the critical challenge in programming.
Those "Simple" tasks have so much variation in them that it takes a billion+ years of evolution + the genetic pretraining to be able to perform.
I don't think so. If you want to model a single synapse in full to capture all effects that might lead to "learning", you have a system of ordinary differential equations. Solving that is very hard, and solving that for 10 million neurons is impossible.
On current hardware can only implement but a poor caricature of a real neuron.
1) Our brains, and moreso those of animals, come with a really good pretraining at birth. This is collective genetic knowledge of millions of generations distilled into your brain.
2) Our brains have a lot of sensors and actuators to interact with the world. We only learn by reading as adults when our brains can already do the synesthesia of translating words into thought. But even as adults, most of us learn better if we do something, write something, engage in dialog, instead of passively listening, reading, or watching.
Passive data can never replicate the rich environment our brains grow up in.
We don't need a complete physiological model for it to be useful. We don't need a perfectly accurate silicon-based mirror of a mammalian brain to outsmart ours on every task we do (and many we don't even realize we could). The challenge will be to coexist and cooperate with these completely alien intelligences that share almost nothing with ours.
This is a neat result. This research started with the differential equation model of a neuron and tried to train various neural nets to get the same result to within 99%. They succeeded. Worst case took an 8-layer net with 256 elements per layer. See Fig. 4. So, 10 billion elements for a squirrel. Not that big by current standards.
It's not clear that a model which tracks the biological neuron that accurately is needed. They discuss simpler models that are almost as good.
Low-end mammal brains should be buildable right now. It's not a hardware limitation.
[1] https://www.sciencedirect.com/science/article/pii/S089662732...
One thing that has held back progress is the way putting knowledge directly into the system has become taboo. So much so that they often fail to even guide the training towards really core aspects of the world model. Or even deliberately going about it with the assumption that everything from start to finish must be determined from the barest input data such as pixels. Then being surprised when it learns random inaccurate and overfit models that miss the underlying hierarchical structures.
I don't believe this is correct. It's too low.
It's more like 400 million.
> somebody ought to be able to build something with 10 million of them
Build them, sure, but they need to be connected in the right way.
One of them is physical structure. You can get 10M somethings, sure, but how do you wire them together is probably more important than how many there are. And there’s many possible combinations.
The other missing part is that we have not figured out the high level software. A squirrel brain is a “desktop PC running windows”-level of utility. A bunch of neurons interlinked is some fashion is equivalent to a blank CPU. We know how the individual transistors work, but the BIOS, and OS are still unknown.
It’s quite possible that problem 1 and problem 2 are related, because evolution doesn’t care about making things easy to understand for us with clear delimitations.
I'd go as far as saying that ML is now at a point where it's basically a mirror image of GOFAI with the exact same issues. The old stumbling block was that symbolic solutions worked well until you ran into an edge case, everyone recognized that having to program every edge case in makes no sense.
The modern ML problem is that reasoning based on data works fine, unless you run into an edge case, then the solution is to provide a training example to fix that edge case. Unlike with GOFAI apparently though people haven't noticed yet that this is the same old issue with one more level of indirection. When you get attacked in the forest by a guy in a clown costume with an axe you don't need to add that as a training input first before you make a run for it.
There's no agency, liveliness, autonomy or learning in a dynamic real-time way to any of the systems we have, they're for the most part just static, 'flat', machines. Honestly rather than thinking of the current systems as intelligent agents they're more like databases who happen to have natural language as a way to query them.
If GOFAI includes semantic reasoning in real world concepts modeled eg. with theasauri and concept maps - I think AI research was on the right track but went astray as there was not enough resounding business success to warrant further funding.
Sure, because it's already a training input. We'd run because we recognize the axe, the signs of aggression, the horror movie trope of an evil clown, and so forth. We have to teach "stranger danger" to children.
"There's no agency, liveliness, autonomy or learning in a dynamic real-time way to any of the systems we have, they're for the most part just static, 'flat', machines."
Well, that's at least in part because we design them that way. It's more convenient to separate out the "learning" and "doing" parts so we have control over how the network is trained.
not in any meaningful sense, no. I can tell you, "if something's fishy about the situation, just leave". You can do this not because of some particular training inputs or examples I give you, but because you have common sense and a sort of personality and intuition for how to behave in the absence of data. If you told that sentence to a state of the art ML model you'd probably get "what fish?" as an answer.
>Well, that's at least in part because we design them that way
It's mostly because we have no idea how to design them anyway else. I think if anyone knew how to build complex agents with rich internal states that have the intent and communication abilities of humans we'd do that. It's not even really conceivable right now how you could have an ML type system that also can just directly adopt high level concepts dynamically just by communicating them.
"Fishy" is doing a lot of work in this sentence. How much training went into refining an instinct for what's "fishy"? Do you not agree that everyone has a different view on what's fishy?
> I think if anyone knew how to build complex agents with rich internal states that have the intent and communication abilities of humans we'd do that.
I'm not so sure. There doesn't seem to be much commercial value in having an agent with intent and its own goals, and most AI advancements are for commercial entities these days.
Only if I had a baseline to compare the situation to. If you took out all familiar elements, I'd have no way of telling whether a situation was normal or suspicious. My understanding of the word "fishy" is born from 300 thousand hours of training data.
"It's mostly because we have no idea how to design them anyway else. I think if anyone knew how to build complex agents with rich internal states that have the intent and communication abilities of humans we'd do that."
That's a different question. We can build machines that learn autonomously; they just don't have the capability of biological minds.
Sorry if i went a bit off topic, but just needed to get my thoughts since yesterday out my head.
But like almost all scientific theories tackling The Hard Problem, it’s built on the assumption that matter gives rise to consciousness.
As time goes by and my own understanding deepens, I’m becoming more and more convinced that this assumption is wrong. Instead we should start considering that consciousness is fundamental, and matter is a product of universal conscious experience.
Idealism is still compatible with the material world, but it seems futile to search for “the experiencer” within the experience itself.
[1] https://en.m.wikipedia.org/wiki/Orchestrated_objective_reduc...
You're right, that's why the concept of "the experiencer" is ultimately an illusion. It's the same sort of illusion as "tables" and "a day job". None of these concepts fundamentally exist in physics, they are labels we apply to loosely defined categories of observations.
Ultimately, Descartes was wrong, "I think therefore I am" is false because it's circular; it assumes the existence of "I" to conclude that "I" exists. The fallacy-free version is "this is a thought therefore thoughts exist", and as you can see, no "I" can be inferred.
If you want to understand what sort of answer neuroscience is starting to provide to the hard problem, I recommend this paper:
A conceptual framework for consciousness, https://pnas.org/doi/10.1073/pnas.2116933119
I'm laughing here because when I posted their ideas to HN oh so long ago, I got downvoted to oblivion because "there's no way organic matter can act as a quantum device". For a place that considers itself a "safe" place to explore ideas, it can be quite dangerous to share too much too early, sometimes.
Time will tell, but my instincts are that we're getting close. We needed computers dreaming first, and we have that now with generative networks!
The other point to make is that we already build systems that can exceed their programming and they are called children.
Even tcp/ip has devolved into a "failed social experiement" with petabytes of low quality/low aptitude vocabulary.
AI is just ambiguous phrasing to color gibberish.
What I mean by that is that AIs, the way they are currently built, need to learn very slowly on short term inputs or they overfit. Whereas humans can learn something just by explanation short term and don't have overfitting problems.
I suspect this is solved by sleep, and I haven't seen AI with a similar mechanism.
How far away do you think we are, exactly?
Hardware is still accelerating exponentially in density, albeit a bit slower. What you're not considering is that algorithmic improvements in machine learning are outpacing hardware improvements.
For instance, NVidia recently revealed how to switch from 32-bit floats to 16-bit floats with no perceptible loss in effectiveness, and they're working on 8-bit floats next. That's a full doubling in number of parameters in your model in only a single step. Other improvements are refinements to language models themselves to reduce overfitting and boost effectiveness with fewer parameters.
Arguably a machine learning model will achieve parity with human neuron density, in terms of number parameters, within the next decade. What that actually means is unclear.
And why is “the reach of normal commuting of the ordinary user” a relevant bar — Google search (as an example) requires computation beyond the reach of normal computing of the ordinary user yet has still had a tremendous impact.
PaLM is not an attempt at AGI, a parameter is not equivalent to a neural connection, an activation function is not equivalent to a neuron (of which you have many different types), biological connection patterns are much richer, and biological stimuli are not like slideshows of a single type of data, so...
That said, there are plenty of multimodal networks (ie not slideshows), and we know very little about the relevance to intelligence of the “richness” of neural connections, activations, etc. — but it’s inarguable that we’ve made great strides in scale alone.
I do believe that AGI is possible and that it does not have to resemble a biological brain though.
Our brains are made of tiny little animals because that's just how life evolved on this planet. It's not a given that this is the best or even a good way to approach the problem.
Brains are very well optimized for computation/energy (while also being self replicating and self repairing), the tasks researchers care about just aren't the ones evolution cares about.
Plus, there are a couple different lines of research for new materials as well, so any one of those can yield something interesting.
https://www.lesswrong.com/posts/K4urTDkBbtNuLivJx/why-i-thin...
Twitter taught Microsoft’s AI chatbot to be a racist asshole in less than a day.
https://www.theverge.com/2016/3/24/11297050/tay-microsoft-ch...