The second chapter (technical performance) is a nice skim. It's mostly--but not entirely--about the last few years of progress in ML, divided by topic.
Whenever I explain computer programs that write computer programs– compiler theory– and programming language theory to someone, they immediately think "That's AI." Yet compiler research and programming language research don't get any AI funding and is not considered AI.
it's true, to some extent AI research is "whatever AI researchers do". whereas if work tackling similar problems is done by PL theory, operations research, or whoever, it's no longer AI.
of course this goes both ways; during the AI winter everyone doing AI research was scrambling to rebrand themselves as doing OR, optimization, computer vision, etc.
and there is still a difference in the terminology and conceptual tools. look at the problem of ensuring that a system will output values that satisfy some constraints.
classically there was a lot of work on constraint solving by people who called themselves AI researchers. meanwhile operations research people worked on integer programming approaches to the same problem. and PL theory people work with abstract interpretations.
For most of the history of AI research the vast majority of AI applications consisted of hand-crafted programs. For example automated theorem proves, planners, SAT-solvers, game-playing algorithms, expert systems, search algorithms etc, are all hand-crafted, rather than learned from data.
What you say, that it's AI if it's learned from data, that applies to machine learning, but machine learning is only one branch of AI. Of course it's the branch that most people know today, but go maybe a few years back and have a look at e.g. the classes on AI taught at places like Stanford or MIT etc, and you'll find that they're all about probabilities and logic, and machine learning does not feature very prominently. You can see the same thing in the staple AI textbook, "AI- A Modern Approach", which is pretty much all hand-crafted approaches.
Perhaps a key distinction is whether the algorithm is mostly learned from data or whether the algorithm is mostly hand-engineered.
In a broad view, I don't think there's any meaningful metric for the "mostly" you're talking about.
Sure, "mostly" seems to make sense in the context of laboriously trained deep networks. But if the training process is improved, drawing a line between training and "looking and understanding" become hard/purposeless. At the limit, suppose some crazy genius created a small, "hand crafted" program, from GOFAI or whatever principles and this program "knew how to learn". If you fed it Wikipedia or whatever, it understood that and by it's content, it would then be "mostly trained on data" or would it?
Well that's how neural networks er work. They are hand-crafted systems with a human-devised training algorithm, backprop. Their output is a model trained on data. But the algorithms that train the model, themselves, are coded manually.
Same goes for basically all machine learning algorithms. They are hand-crafted systems that train models from data.
Indeed, it is impossible to learn anything "from data".
Data is just measurement of observable variables. A prior model of the meaning of those measurements is required first to parse them into a coherent "observational model"; and then many prior models is required to parse into a representational model.
Few, if any extant "AI" systems are able to take the latter step, not least, as it requires more than measurements of the target system which are always ambiguous. (In particular, it requires a coordinated body).
It's strange that it came down to it. AI is a field inspired by our notion of human intelligence. Yet a human can learn to recognize road signs with a fraction of the data state of the art "AI" algorithms need these days.
To play chess at the same level as a human master a state of the art neural network based engine needs data from tens of millions of self played games - several orders of magnitude more than a human master encounters (and then the computer has huge advantage in calculation speed and accuracy as well). I don't think deep learning is the end of it. It would be really great to see people venturing into different approaches. We can do much better than pattern recognition on huge amount of data.
> several orders of magnitude more than a human master encounters
Not if you account for millions of years of evolution that got your nervous system to its current state and access to books and training materials which add up to multiple decades/centuries of precompiled and synthesised expertise from other people who learned chess in similar ways to you.
This is a minor point though. I agree that deep learning isn't it, or that if it is, that would be quite underwhelming.
If you want to compare today's state of the art chess playing model to a human that's the relevant comparison, but if you want to compare the current field of AI to humans the line between the system and the training becomes fuzzy.
If you look at the past 10 years of AI we have massively reduced the amount of data required to achieve a given score for computer vision models by making changes to the system of the neural networks. In fact it's where most of the improvements have come from rather than increasing the amount of data. I don't see a reason to believe it won't keep working this way. I'd say we're being damn efficient and it doesn't feel fair to say "look at how much data they need, the approach is fundamentally wrong" when we've been at it for such a short amount of time compared to the millions of years it took humans to evolve.
IMO, cybernetics is more conceptually grounded than AI — at least in so far as it is possible to objectively define what constitutes a cybernetic system. The term "Artificial Intelligence" was literally invented for the purposes of attracting grant money. It is still good for that.
pdf: https://aiindex.stanford.edu/wp-content/uploads/2021/03/2021...