[D66] Computers don’t give a damn

[R.O.]

Contemporary philosophy 
<https://www.the-tls.co.uk/categories/philosophy/contemporary-philosophy/>|Book 
Review


  Computers don’t give a damn


    The improbability of genuine thinking machines

By tim crane 
<https://www.the-tls.co.uk/articles/promise-of-artificial-intelligence-brian-cantwell-smith-book-review/#>
“Archive Dreaming” installation at SALT Galata, Istanbul, by Refik 
Anadol, 2017. It uses AI to sort and combine 1.7 million historical 
documents
“Archive Dreaming” installation at SALT Galata, Istanbul, by Refik 
Anadol, 2017. It uses AI to sort and combine 1.7 million historical 
documents© Getty Images
May 15, 2020
Read this issue <https://www.the-tls.co.uk/issues/may-15-2020/>


      In this review

THE PROMISE OF ARTIFICIAL INTELLIGENCE
Reckoning and judgment
184pp. MIT Press. £20 (US $24.95).
Brian Cantwell Smith
Buy 
<https://shop.the-tls.co.uk/the-promise-of-artificial-intelligence-9780262043045.html?utm_source=tls&utm_medium=mainsite&utm_campaign=review>

In 1965, Herbert Simon, one of the founders of the new science of 
Artificial Intelligence (AI), wrote that “machines will be capable, 
within twenty years, of doing any work that a man can do”. He was wrong 
of course – but maybe his mistake was only a matter of timing.

If Simon were to see what computing machines were capable of, fifty-five 
years after he made this remark, surely even he would be amazed. A 
single smartphone contains more computing power than all the world’s 
computers in 1965 put together. And many of the philosophical arguments 
against the possibility of AI from the 1960s and 70s fell flat on their 
face as the technology advanced. The philosopher Hubert Dreyfus also 
claimed in 1965 that “no chess program can play even amateur chess” – 
true at the time, but proved false soon after. When in 1997 the IBM 
programme Deep Blue beat the chess champion Garry Kasparov, this 
conclusively destroyed the idea that world-class chess was something 
computers can’t do; Kasparov has commented recently that “today you can 
buy a chess engine for your laptop that will beat Deep Blue quite 
easily”. And the familiar claim that computers could never really use 
their stored knowledge as well as human beings was shaken in 2011 by 
IBM’s Watson programme, which won the top $1 million prize on the 
American game show /Jeopardy/, beating the best human competitors.

AI sceptics used to claim that computers would never be able to 
recognize human faces or human speech, translate speech into text, or 
convert handwriting to printed text. But today’s phones can do all these 
things. Dreyfus had made gentle fun of the grand claims of AI, quoting a 
fanciful newspaper report from 1968 about “a new idea in gifts … a 
genuine (if small) computer, that costs around $20. Battery operated, it 
looks like a portable typewriter. But it can be programmed like any big 
computer to translate foreign languages, diagnose illnesses, even 
provide a weather forecast”. What seemed then like wild science fiction 
is now our everyday reality.

So Simon’s claim may have been proved false, but maybe he was only a few 
decades out. The achievements of actual AI – that is, the kind of 
technology that makes your smartphone work – are incredible. These 
achievements have been made possible partly by developments in hardware 
(in particular the increased speed and miniaturization of 
microprocessors) and partly because of the access to vast amounts of 
data on the internet – both factors that neither Simon nor Dreyfus could 
have predicted. But it means that enthusiastic predictions for AI are 
still popular. Many believe that AI can produce not just the “smart” 
devices that already dominate our lives, but genuine thinking machines. 
No one says that such machines already exist, of course, but many 
philosophers and scientists claim that they are on the horizon.

To get there requires creating what researchers call “Artificial General 
Intelligence” (AGI). As opposed to a special-purpose capacity – like 
Deep Blue’s capacity to play chess – AGI is the general capacity to 
apply intelligence to an unlimited range of problems in the real world: 
something like the kind of intelligence we have. The philosopher David 
Chalmers has confidently claimed that “artificial general intelligence 
is possible … There are a lot of mountains we need to climb before we 
get to human-level AGI. That said, I think it’s going to be possible 
eventually, say in the 40-to-100-year time frame”. The philosophers John 
Basl and Eric Schwitzgebel are even more optimistic, claiming it is 
“likely that we will soon have AI approximately as cognitively 
sophisticated as mice or dogs”.

The intellectual enthusiasm for the possibility of AGI is matched by the 
vast sums invested in trying to make it a reality. In July 2019, 
Microsoft announced that it would invest $1 billion in Sam Altman’s 
OpenAI, a for-profit company which aims to use AI for the “benefit of 
mankind as a whole”. The British company Deep Mind, led by the computer 
scientist/neuroscientist Demis Hassabis, was bought by Google for $500 
million in 2014. Deep Mind’s best-known achievement to date is the 
machine AlphaGo, which in 2016 beat Lee Sedol, world champion of the 
ancient game of Go. Go is vastly more complex than chess – it is 
sometimes said to be the most complex game ever created – and standard 
AI chess-playing methods had never been successfully applied to it. The 
computing methods used by AlphaGo are often touted as one of keys to 
“solving intelligence”, as Deep Mind’s own publicity puts it.

Brian Cantwell Smith’s new book is a provocative expression of 
scepticism about these recent claims on behalf of AI, from a 
distinguished practitioner in the field. His overall argument is based 
on a distinction between what he calls “reckoning” and “judgment”. 
Reckoning is understood here in its original etymological sense: as 
calculation, like addition and subtraction. Judgment, by contrast, is 
something more. It is described by Smith as “an overarching, systemic 
capacity or commitment, involving the whole commitment of the whole 
system to the whole world”. Our thinking involves not just some kind of 
simple on-off representation of things around us, but an entire 
emotional and value-laden involvement with the world itself. Computers 
have none of this. As the philosopher John Haugeland (a major influence 
on Smith) used to say, “computers don’t give a damn”. Giving a damn is a 
precondition of “judgment” in Smith’s sense, and anything that amounted 
to a real AGI would need to exercise judgment, and not simply calculate.

/The Promise of Artificial Intelligence /gives a brief and intelligible 
survey of two main stages in the history of AI. The first stage, 
starting in the 1960s, was what Haugeland christened “Good Old-Fashioned 
AI” (GOFAI) which solved computing problems by using explicit 
representations of general principles and applying them to particular 
situations. (Think of doing a mathematical proof or presenting an 
argument in logic.) “Second wave” AI, which started to emerge in the 
1980s, began from the opposite end, so to speak: deriving general 
conclusions from huge amounts of simple data as input. This kind of 
approach, variously called machine learning or deep learning, has had 
considerable success at things that GOFAI was very bad at, like pattern 
recognition, or updating knowledge on the basis of input.

First wave AI, it was often said, misconceived the nature of thinking: 
very little thinking resembles calculating or proving theorems. But 
Smith goes further, and argues that “the deeper problem is that it 
misconceived the world”. GOFAI assumed that “the world comes chopped up 
into discrete objects”, and because of this it analysed reasoning into 
its components by using formal logic (the basic ideas of which underlie 
modern computing). Smith argues that first wave theorizing assumed that 
the world must be structured in the way that logic structures language: 
objects correspond to names, properties correspond to predicates or 
general terms. Things fit together in the world as symbols fit together 
in a logical language. Smith claims that this is one main reason why the 
GOFAI project failed: it failed to take account of the “fabulously rich 
and messy world we inhabit”, which does not come in a “pre-given” form, 
divided up into objects.

Second wave AI, according to Smith, does not make this mistake. It does 
not assume a “pre-given” ontology or structure for the world, and for 
that reason, he argues, it has made progress in the areas where GOFAI 
failed: in particular, with tasks like face recognition, text processing 
and (most famously) the game of Go. The distinctive feature of deep 
learning machines is their ability to detect patterns in large 
(sometimes huge) amounts of data. The machines “learn” by being given an 
indication by the programmer of which patterns are the important ones, 
and after a while they can produce results (for example, moves in a 
game) that surprise even the programmers. This is in contrast to first 
wave AI programmes which attempted to anticipate in advance how input 
from the real world should be responded to in every conceivable 
situation. Those early AI machines that worked only did so in very 
constrained made-up environments, sometimes called “microworlds”.

Nonetheless, Smith thinks that we should not be too optimistic about the 
ability of second wave AI to create AGI. Machine learning may not start 
with general rules which make ontological assumptions, but it does start 
with data that is already processed by humans (eg things that we 
classify as faces, or as road traffic at an intersection and so on). 
Much machine learning, as Smith says, is “dedicated to sorting inputs 
into categories of manifest human origin and utility”. So even if they 
are more sensitive to the messy world, second wave AI machines are still 
tied up with the programmers’ own classifications of reality – indeed, 
it is hard to see how they could be otherwise designed.

Smith is surely right that AI’s recent successes give us little or no 
reason to believe in the real possibility of genuine thinking machines. 
His distinction between reckoning and judgment is an important attempt 
to identify what it is that is missing in AI models. In many ways 
(despite his protest to the contrary) it echoes the criticisms of 
Dreyfus and others, that AI will not succeed in creating genuine 
thinking unless it can in some way capture “common sense”. And just as 
common sense (part of Smith’s “judgment”) cannot be captured in terms of 
the “rules and representations” of GOFAI, nor can it be captured by 
massively parallel computing drawing patterns from data.

To make this point about judgment, Smith does not actually need the more 
ambitious ontological claims, that the world does not have natural 
divisions or boundaries, that all classification is simply a result of 
human interest, and so on. Maybe these claims are true, maybe not – for 
many centuries philosophy has wrestled with them, and they are worth 
debating. But we should not need to debate them in order to identify the 
fundamental implausibility of the idea that AGI is on the horizon.

This implausibility derives from something intrinsic to the success of 
AI itself. For despite the sophistication of machine learning, the fact 
remains that like chess, Go is still a game. It has rules and a clear 
outcome which is the target for players. Deep learning machines are 
still being used to achieve a well-defined goal – winning the game – the 
meaning of which can be articulated in advance of turning on the 
machine. The same is true of speech and face recognition software. There 
is a clear goal or target – recognizing the words and faces – and 
successes and failures in meeting this goal are the input which helps 
train the machine. (As Smith says, “recognition” here means: correctly 
mapping an image onto a label: nothing more than that.)

But what might be the goal of “general intelligence”? How can we 
characterize in abstract terms the problems that general intelligence is 
trying to solve? I think it’s fair to say that no one – in AI, or 
philosophy, or psychology – has any idea how to answer this question. 
Arguably, this is not because it is an exceptionally difficult empirical 
question, but rather that it is not obviously a sensible one. I suppose 
someone might say, in the spirit of Herbert Simon (whose famous AI 
programme was called the “General Problem Solver”), that general 
intelligence is the general ability to solve cognitive problems. This 
might seem fine until we ask ourselves how we should identify, in 
general terms, the cognitive problems that we use our intelligence to 
solve. How can we say, in general terms, what these problems are?

Consider for example, the challenges faced in trying to create a genuine 
conversation with a computer. Voice assistants like Siri and Alexa do 
amazingly well in “recognizing” speech and synthesizing speech in 
response. But you very quickly get to the bottom of their resources and 
reach a “canned” response (“here are some webpages relating to your 
inquiry”). One reason for this, surely, is that conversation is not an 
activity that has one easily expressible goal, and so the task for the 
Siri/Alexa programme cannot be specified in advance. If the goal of 
conversation were to find information about a subject matter, then 
directing you to a website with relevant information could be one 
reliable way of achieving that goal. But of course this is not the sole 
thing to which we direct our intelligence when talking with others.

What, then, is the overall goal of conversation? There isn’t one. We 
talk to pass the time, to express our emotions, feelings and desires, to 
find out more about others, to have fun, to be polite, to educate 
others, to make money … and so on. But if there is no single goal of 
conversation, then it is even less likely that there is one goal of 
“general intelligence”. So no wonder AI researchers struggle to even 
define the “task domain” for AGI.

As Smith’s book shows, the claims for the possibility of AGI ignore the 
huge differences between the relatively well-defined areas where AI has 
succeeded, and the barely defined domain of “general intelligence”. This 
is, on its own, enough of a reason for scepticism about extrapolating 
beyond the spectacular successes of actual AI to the real possibility of 
AGI. Smith’s arguments about the ontological assumptions of AI, whatever 
their merits, are not necessary to make this point.

Yet I suspect that many still have this lingering sense that AGI must be 
possible, and that the difference between real human thinking and what 
computers do is just a matter of complexity. What lies behind this 
conviction? One widespread idea is that since the human brain is just a 
complex biological (and therefore material) machine, it must be possible 
in principle to artificially reproduce what the human brain does 
(thinking, perceiving, feeling, imagining, being conscious, etc) by 
building something that functions in exactly the same way as the brain. 
And whatever we thereby build will be an artificial version of our 
mental processes: an AGI.

This argument is based on two ideas: first, that thinking and other 
mental processes go on in the brain; second, the brain is a machine or 
mechanism. So if we can uncover the principles that make this mechanism 
work, and we have adequate technology – the argument goes – then we 
should be able to build a machine that implements these principles, 
without leaving anything out. One of the pioneers of deep learning, 
Yoshua Benigo, puts it this way: “I don’t know how much time it’s going 
to take, but the human brain is a machine. It’s a very complex one and 
we don’t fully understand it, but there’s no reason to believe that we 
won’t be able to figure out those principles”.

Of course building an artificial copy of a real brain is nowhere close 
to today’s scientific reality. But if we believe that we are at bottom 
material beings – if we take away all our matter then there is nothing 
left of us – then such replication seems possible in principle, even if 
it is never actually realized. Suppose, then, a brilliant scientist of 
the future could replicate in an artificial construction everything a 
person (and their brain) does. Obviously this replica would also be able 
to think, since thinking is one of the things a person does. It is 
undeniable that making an artificial replica of a person and all their 
features would be making an “artificial intelligence” in an an obvious 
sense: simply because intelligence is one of the features of people.

The question is, what does this have to do with AI? If the way to create 
a real artificial thinker is to find out first how the brain works, then 
you would expect AI researchers to try and figure out how the brain or 
the mind actually works – that is, to become neuroscientists or 
psychologists. But this is not how AI researchers operate. Just as the 
invention of the aeroplane did not require building something that flies 
in exactly the way a bird flies, so the inventors of AI did not feel 
bound to copy the actual workings of human brains. And despite the fact 
that deep learning computers use what are called “neural networks”, the 
similarity to the brain here is at a very abstract level. Indeed, many 
pioneers of deep learning occupy themselves with very abstract questions 
about general intelligence – Benigo says his goal is “to understand the 
general principles of intelligence” – rather than with the messy 
business of the actual working of the human mind or brain.

This lack of focus on the way human minds (or brains) actually function 
goes back to the beginnings of AI, and it was clearly one of AI’s 
strengths. By ignoring the complexity of actual human thinking, and the 
messy “wetware” of the actual human brain, AI could get machines to 
solve difficult problems without having to bother with how we would 
solve them. Watson, the IBM website tells us, “is not bound by volume or 
memory. Watson can read millions of unstructured documents in seconds”. 
That’s not something we can do. The learning involved in training 
AlphaGo involved millions of practice games of Go – as the cognitive 
scientist and deep learning sceptic Gary Marcus has pointed out, this is 
“far more than a human would require to become world class at Go”. This 
means that whatever it explains about their success, it is not the 
similarity of these machines to human thinkers. So IBM should not claim 
that Watson is “built to mirror the same learning process that we have, 
a common cognitive framework that humans use to inform their decisions”.

Actual AI does not rely substantially on any detailed psychological or 
neuroscientific research. So the fact that the brain is a material 
mechanism which could, in principle, be replicated artificially gives no 
support to the idea that AI as it actually is could build an AGI. In 
fact, given the way AI has actually proceeded, in splendid isolation 
from neuroscience, it is likely that any attempt to replicate the brain 
would require ideas very different from those traditionally used by AI. 
To say this is not to disparage AI and its achievements; it is just to 
emphasize the obvious fact that it is not, and has never been, a theory 
of human thinking.

Philosophical and scientific discussions of AI have tended towards one 
of two extremes: either that genuine artificial thinking machines are 
just on the horizon, or they are absolutely impossible in principle. 
Neither approach is quite right. On the one hand, as Smith, Marcus and 
others have explained, we should be sceptical that recent advances in AI 
give any support to the real possibility of AGI. But on the other, it is 
hard to deny the abstract philosophical claim that if you could 
replicate a human brain in such a way that would produce something that 
did everything the brain did, then that thing would be a “thinking 
machine” in one clear sense of that phrase. However, this mere 
possibility does not mean that today’s AI is anywhere near creating 
genuine thinking machines. On the contrary: when you examine the 
possibility more closely, it shows why AI is unlikely ever to do this.

/*Tim Crane* is Professor of Philosophy at the Central European 
University and Philosophy editor of the /TLS/. The third edition of his 
book /The Mechanical Mind/was published in 2015/

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.tuxtown.net/pipermail/d66/attachments/20210126/5144e72c/attachment-0001.html>