The Chisel-Construct Cycle · AI Series · 05

Turing Replaced “Can Machines Think?” with Another Question

Han Qin · About 28 minutes · Published

1. I Shall Replace the Question with Another

The first article in a quarterly in October 1950 was “Computing Machinery and Intelligence,” by Alan Turing. It occupied twenty-eight pages.

The quarterly published psychology and philosophy. Its editor was the philosopher Gilbert Ryle, and the other pieces in the issue were by philosophers. Here was a mathematician submitting a paper to a philosophy journal.

He was then a senior lecturer in the mathematics department at the University of Manchester. There had been meetings about mind and machine in Manchester the previous autumn and winter. People taking part included the mathematician Newman, Polanyi, who had trained as a chemist, Jefferson, a neurosurgeon, and a zoologist.

The article has seven sections and no abstract, illustrations, or bibliography at the end. The references are in the text: Gödel, Church, Kleene, Rosser, Turing's own result of 1937, Jefferson's 1949 lecture, Lovelace's notes on the Analytical Engine, and Hartree's 1949 book. Babbage comes up in the account of the history. A mathematician writing in a philosophy journal is drawing half on logicians, half on a neurosurgeon and a nineteenth-century countess.

The first sentence raises the question: Can machines think?

The usual way to answer would be to define machine and think. But if one is to find the meanings in ordinary usage, both the meaning of the question and its answer will have to rest on a statistical survey, something like an opinion poll. He calls that absurd.

His objection is not to defining the words. It is to making statistics about usage the ground of the definition. If you settle the meaning by counting heads, then whether machines can think depends on how many people are willing to say so right now.

So he writes that he will replace this question with another.

He does not say the original is a false question, that it ought not to be asked, or that the new one is the same as the old. He says the new question is closely related to the old and is put in relatively unambiguous words.

Closely related does not mean identical. Relatively unambiguous does not mean unambiguous. He added both qualifications. In the seventy years and more since, both have seldom been quoted together.

The sentence people now know so well, about the question being too meaningless to deserve discussion, is not here. It is at the start of Section 6, by which time the game has been described and the machine formally defined. The opening paragraph is much more restrained. Move the line from Section 6 into the first paragraph and the tone of the whole article changes. A person saying he will try asking a different way becomes a person saying the original question is not worth discussing.

2. The Game as He Set It Out

The game in Section 1 starts with three people.

A is a man. B is a woman. C is the interrogator and can be either sex.

In this setup the distinction to be made is between a man and a woman, not between a person and a machine. All three players are people. There is no machine in this version. The machine enters in the next paragraph, taking one person's place and taking over that person's task.

C stays in a room apart from the other two and knows the unseen players only by labels, which we may as well call X and Y. At the end, C has to decide which label belongs to the man and which to the woman.

C can ask questions of both. Turing's first example asks X to say how long his or her hair is.

A, the man, has been given the part of the troublemaker. He has to get C to make the wrong identification. In Turing's sample answer, A describes his hair in a woman's style. The point of this first game is impersonation. One person tries to pass as another sort of person, a second tries to tell the truth, and a third has to tell them apart.

B, the woman, has the opposite task: help C. Her best strategy, Turing says, is probably to tell the truth. She can say outright that she is the woman. He immediately points out that A can say the same thing.

The answers must be written, preferably typed, so the voices do not give the players away. The ideal arrangement is a teleprinter connecting two rooms. Two rooms, one teleprinter line. No faces, no voices, just the words as they come through.

Then comes the move.

The machine takes the place of A.

It takes the part of the man who was supposed to mislead the interrogator. It does not take the place of B or C, and it does not join as a fourth player.

The version familiar from later accounts has a person on one side and a machine on the other, with the judge deciding which is the person. In the opening version, the machine takes over a part someone is playing. It has to do what the man was doing: get C to make a mistake.

The question after the substitution is a comparison. Will C make as many mistakes in the machine version as in the one where a man and a woman were hidden from sight? The baseline is not the truth. It is the performance in another game of impersonation. The machine is being compared to a man who is trying to pass as a woman.

These questions, Turing says, replace the original question of whether machines can think.

Section 1 asks about the error rates in the two forms of the game. It cannot simply be reduced to whether a judge can tell a person from a computer. But the original paper itself does not fix the first scene for good. By Section 5, Turing says a man can take B's place. The later formula drops the opening comparison between man and woman, but it does not invent an entirely different question out of nowhere.

Section 1 has no time limit or percentage. It does not stipulate the number or subjects of questions or what qualifies a person to be the interrogator. Nor does it say that a machine can think once it reaches a particular mark.

What the section gives is an arrangement, not an exam.

The name Turing test is not in the paper. Turing uses imitation game, criterion, experiment, and test. Each carries a different weight, and he uses all four. A criterion marks a line. An experiment tries something out; a test asks you to show what you can do; a game is something you play. He does not say which of the four is definitive. The proper name came from other people later.

3. Two Predictions

At the start of Section 6, he distinguishes what he believes from what has been proved and marks what follows as conjecture.

Then comes the number people have quoted for more than seventy years.

The gist is that in about fifty years it should be possible to make computers with a storage capacity in the region of a billion and program them so that an average interrogator has no more than a seventy percent chance of making the right identification after five minutes of questioning.

Three numbers: fifty years, five minutes, seventy percent. They all appear here, not in the rules in Section 1. People who have since run the game as an exam have mostly used those numbers. In the paper they are parameters of a conjecture, not provisions of a rulebook.

Immediately afterward comes another prediction.

By the end of the century, he says, ordinary usage and educated opinion will have changed to the point where saying a machine is thinking will not usually draw an objection.

The two sentences sit side by side, but they are about different things. The first is an empirical prediction about performance in a game, with conditions of storage, time, and probability. The second is a prediction about language and educated opinion. To check the first, use a machine and a stopwatch. To check the second, listen to how people talk.

One man makes two predictions on one page. One is about a machine, the other about a word. The line the first four essays have followed, a boundary that moves with what people know and say about it, is here made into a prediction by one of the people involved. He did not say the machine would cross a line. He said the way people speak would change.

In the first essay a sentence picked up a word as it traveled. In the second a limited technical statement lost its qualification. In the fourth someone wrote in a manuscript in a drawer that the boundary lay on the side of the observer. Here one of the people involved makes it a prediction about the future, and the object of the prediction is the use of words itself.

But the second sentence does not cancel the first. In Section 7 he says that the eventual, satisfactory support for the view in Section 6 would be to do the experiment at the end of the century. The prediction about language does not swallow the one about the experiment. Both remain.

It is also at the start of Section 6 that he writes the best-known sentence: in its literal form, the original question is too meaningless to be worth discussing.

Nearby he writes another, quoted much less often. The original form cannot be entirely abandoned, because people will disagree about whether the substitution is appropriate.

That is the move in this essay.

A question out of reach, one that cannot be measured, is replaced with a proxy that is within reach and can be measured. The original has not been answered, disproved, or declared nonexistent. It has been moved to the side. The person who moved it says the move can be questioned, which is why the original form cannot be thrown away.

This is where colonization takes place. Nobody has seized anyone else's land. Beside a piece of land that cannot be measured, someone marks out another that can, and most of the dispute moves to that second piece. The first is still there. Fewer people look its way.

This move differs from the earlier ones. Before, the construction did something, and the boundary receded. This time the construction did not do anything; the thing being measured was exchanged for one that could be measured. In 1950 not a single machine could play the game. What was replaced was the question, not the machine's powers. Behind each retreat in the preceding essays there had been something that had been done. This time there wasn't. It was the ruler that moved.

More than half of Section 6 deals directly with objections to the idea that a machine can think. Turing does not do away with the question. He argues with it for twelve pages. Put the two things together: someone says a question is not worth discussing, then spends twelve pages answering nine views on it. What he found wanting was not the importance of the question, but its literal form as something one could answer directly.

4. Nine Objections

In Section 6 he sets out nine objections and answers them one at a time.

The first is the theological objection. Thought belongs to the immortal soul God gave to human beings. Animals and machines do not have such souls. They cannot think.

Turing does not argue for materialism. He replies within the other side's terms. There is a problem with limiting God's powers that way. Making a thinking machine can be compared to having children; people are instruments of divine creation, not its rivals. Turing calls parts of this speculative theology. He is no more dogmatic about it than the objector. The objector invokes theology and so does he. His answers differ among the nine, but each time he stands on the other side's ground. This is a way of arguing: do not change fields, answer where the other side drew the lines. Yet the move of the whole article is to change fields. Both things happen in the same paper. He asks people to change the question, then meets each objection on the old question's ground.

The second is the heads-in-the-sand objection. The consequences of machines thinking would be too dreadful, so let us hope and believe they cannot.

He takes it as an emotional position, especially understandable in people who have staked much of human superiority on intelligence. It is not substantial enough to warrant a logical reply, he says. There is sympathy, no concession. Of the nine it is the one he expressly says does not deserve a logical rebuttal.

The third is the mathematical objection.

He refers to results from 1931 to 1937, including his own paper of 1936. They show that for a machine that proceeds step by step under fixed rules, however large it is built, there are questions it cannot answer. This is not an engineering shortcoming. It can be proved.

He makes a definite concession here. Any particular machine of the kind has a limit. There will be a question it cannot answer. This is the point of the third essay: the boundary is not drawn by a person. It follows from the rules of the construction.

His reply turns on the comparison. Nobody has proved that human intellect has no corresponding limitation. Human beings also make mistakes.

Of the nine, this is one he says cannot be passed over lightly. What he grants is that a particular machine has mathematical limits. He does not grant that human beings therefore have unmatched powers. Between the two lies a premise nobody has proved.

The fourth objection is the argument from consciousness.

Here he quotes the June 1949 lecture. The passage begins with a machine having to write a sonnet or compose a concerto because of thoughts and feelings actually experienced. It goes on to the condition that the machine must know it has written the work, and to the list: pleasure at success, grief, a response to flattery, making mistakes, sex, anger, and dejection.

He leaves out none of the three conditions discussed in the fourth essay.

His first reply takes up the problem of other minds. Insist on having direct access to the machine's subjective experience, then apply the same standard consistently to another person, and you slide toward the position that you can only know you are conscious yourself.

The second part of the reply is an imagined oral examination. He gives the machine questions about poetry, then asks whether verbal performance of this kind would be taken as evidence if another person supplied it.

He does not deny that consciousness is hard. He says plainly that there is a mystery there. He says only that it need not be solved before the machine question can be answered. He does not settle consciousness. He moves it farther down the road, as he does with the original question in the paper as a whole.

The fifth is the argument from various disabilities.

Under this objection is a list. People say a machine will never be kind, resourceful, beautiful, or friendly; have initiative or a sense of humor; tell right from wrong; make mistakes; fall in love; enjoy strawberries and cream; make someone fall in love with it; learn from experience; use words properly; take its own thoughts as a subject; match a person's variety of behavior; or do anything really new. Sixteen things people around 1950 said machines would never be able to do.

He says claims of this sort generally lack evidence and are often drawn from the handful of extremely limited machines a person happens to know. This is a point about induction. The machines you have seen cannot do it; that does not mean machines cannot do it. And in 1950 there were scarcely more than a dozen machines anyone could see anywhere in the world, most of them in laboratories.

He takes several of these, the pleasure of strawberries and cream, falling in love, and feelings of that kind, for the objection from consciousness in disguise, not separate limits established by proof.

On mistakes he distinguishes errors of functioning from errors of conclusion. A machine can be arranged to give deliberately misleading answers; it can also reach a wrong conclusion in the course of a process that is functioning as designed.

As for taking itself as a subject, a machine can deal with its own operations or its own program, he says. As for variety of behavior, he points to the enormous range offered by sufficient storage. He does not grant that finite machinery necessarily behaves in monotonous, repeating ways.

The seventh is the argument from continuity in the nervous system. The nervous system is not a discrete-state machine. Small continuous changes may matter. A digital computer is therefore physically organized differently from it.

He accepts the physical contrast but points out that the game looks at the answers. A digital machine can approximate a continuous device, such as a differential analyzer, to a point where the interrogator cannot tell from the answers what kind of mechanism is behind them.

He grants that the nervous system need not really be a discrete-state machine. He does not grant that the difference itself provides a way to tell the two apart in this game.

It is the method of the whole paper applied to a particular objection. A difference you cannot observe from the other side of the door is not a difference in the game. That is the strength of the method. It is also its price.

The eighth is the argument from the informality of behavior. No finite set of rules prescribing an action for each situation can exhaust how people behave. A machine governed by rules therefore cannot reproduce that behavior.

He expressly grants a narrower claim. There really is no simple code of conduct telling a person what to do in every situation.

Then he distinguishes codes of conduct from laws of behavior. Not finding a complete, explicit law of behavior does not prove that none exists. The one is a rule a person acts on. The other is a law by which a person is described. There are no complete rules of the first sort. There may still be laws of the second.

The third, seventh, and eighth objections draw his largest concessions. All three concessions concern the same sort of thing: a concrete formal or physical limit. In none of them does he concede something a person has that a machine must lack.

What he grants are particular limits one can point to. What he does not grant is the step from those limits to conclusions about people.

5. The Sixth Objection

The sixth is called Lady Lovelace's Objection.

Turing says the fullest account of Babbage's Analytical Engine is in Lady Lovelace's notes and dates them to 1842. They were published in 1843.

He takes two propositions from them. The Analytical Engine has no pretensions to originate anything. It can do what people know how to order it to perform.

These are the two sentences from the last note in the first essay. One hundred and seven years later they are printed again in a philosophy quarterly, at the head of an objection to be answered.

The first sentence he quotes leaves out an emphatic word in the original. Lovelace's paragraph also continues after the two sentences. What she says about the machine's possible indirect effects on science is left outside, as is her point about why operations need not be confined to numbers.

He goes on to quote Hartree's 1949 book. Hartree had considered whether a machine with a learning mechanism something like a conditioned reflex might complicate the question of thinking for itself. He had also judged that machines already built or then being planned did not have the relevant property.

Turing agrees with Hartree's judgment about the available evidence. He does not accept the stronger conclusion that such a property is impossible in principle.

As the first essay noted, Hartree left room when he quoted the sentence. Turing reprints what he says in leaving it open and agrees with him. The three people keep on the same side of that line. Lovelace is talking about the unbuilt machine of her time. Hartree is talking about the machines built or planned in 1949. Turing is talking about the same machines. All three are speaking of the present, not forever.

Turing then recasts the objection in two stronger forms.

One: a machine can never do anything really new.

The other: a machine can never take us by surprise.

He answers the second in a sentence: “Machines take me by surprise with great frequency.”

The sentence shifts the place of the argument. It was about whether a machine had a particular capacity. Now it is about whether a person has worked out all that will follow from the rules that person sets.

He explains that people often do not see the consequences of the rules and data they supply. The person makes the rules. The person has not worked out all that follows.

Near the end of Section 7 he adds a footnote to a formulation he has put down himself: “the machine can only do what we know how to order it to do.” The footnote expressly refers back to Lovelace and records one thing. Her sentence does not contain the word only.

This is the first time the record in this series closes a loop. The first essay's sentence was written in 1843. A hundred and seven years later someone quotes it, then points out later in the same paper that the sentence picked up an extra word along the way.

And in that same paper he gives the date of the notes, off by a year.

While correcting someone else's rendering of a document, he misremembers a detail of the document himself. Both things are there in the same twenty-eight pages, nine pages apart.

6. A Telepathy-Proof Room

The ninth is the argument from extrasensory perception.

Later accounts often leave it out. In the paper it takes up two pages, comes last of the nine, and receives one of Turing's strongest endorsements as an objection.

He names four phenomena: telepathy, clairvoyance, precognition, and psychokinesis.

First he says the subject presents a difficulty for his argument and that he would rather these things were not real. Then he says something at odds with that wish. He took the statistical evidence, at least for telepathy, to be overwhelming at the time. He does not dismiss the matter as nonsense. The wish and the judgment appear separately, the wish first.

He calls the objection strong for a specific reason. If such phenomena are possible, the assumption that the interrogator and the players are cut off from one another as intended does not hold.

The objection does not attack the idea that a machine can think. It attacks the room. Does the door really separate the two sides? The whole method rests on the separation, and the objection says there may be a leak. The worst defect of a ruler is not where its marks are, but whether it measures the thing you think it measures. That is what the section says. And the person saying it made the ruler.

On that assumption he sets out a numerical example. The interrogator holds up cards and asks which suit they belong to. A person with telepathy might get a hundred and thirty out of four hundred right; a machine guessing at random might get a hundred and four on one run. The gap would let the interrogator identify the person by the scores. Then the game would pick out a different ability, not the ability to think.

He then complicates the example. If the random number generator of the computer can be affected by the interrogator's psychokinesis, even the baseline of random guessing can be upset. Clairvoyance could give away the witness's identity without the ordinary questions at all.

His solution is to isolate the people taking part, for instance with a telepathy-proof room.

A telepathy-proof room. That phrase appears in the body of a philosophy quarterly in 1950.

This is where the passage sits. The paper's method is to trade an immeasurable question for a game that can be measured. Here it says that an assumption of the measurable game might fail. Turing does not take the problem out. He adds a room to the game.

When a device for measuring comes to a place where it cannot measure reliably, there are three choices. Define the difficulty out of existence, say nothing about it, or change the device. He chose the third.

This is the paper explaining itself. He knows what his method rests on. He knows the ground may not hold.

When the method is retold, this section is usually not there.

7. A Child Machine

Section 7 is about learning machines.

He divides the making of an adult mind into three sources: the initial state at birth, education, and experiences outside education. Only in the second can another person intervene. He proposes taking over the second, leaving the first to the programmer and the third to chance and the surroundings.

Then he turns the usual programming problem around. Rather than try to simulate the adult mind directly, make something like a child's mind and educate it.

His comparison is a brand-new notebook. The child is imagined to have little built-in machinery and many blank pages. He immediately notes that a real child's brain is unlikely to be this simple. A comparison is a comparison. He is the first to write down its limits.

He divides this into two experimental tasks that belong together: make the child program, and design its education. He expects to try repeatedly, not to get it right in one go. It is the same road as in his internal report of 1948. Do not design the finished product. Make something that can change and teach it. Two years have gone by, the report is still unpublished, and its idea appears in print here in different words.

He compares successive changes to the child machine with biological evolution. Its structure corresponds to hereditary material, changes correspond to mutations, and the experimenter's judgment corresponds to natural selection.

In that comparison the experimenter is standing on the side where natural selection would stand. Deciding which change counts as good has not been handed over. In evolution nobody is in charge. In this plan somebody is. There is always someone doing the teaching in the education he imagines.

He does not require education to reproduce every channel open to a child. With adequate two-way communication, he says, one could in principle get very far with education. He takes Helen Keller as an example: even severe sensory limitations need not prevent the development of intellect. What he wants is not for the machine to grow eyes and ears, but for there to be a channel wide enough to go both ways.

As for reward and punishment, in the teaching scheme he imagines, an event or pattern just before punishment is made less likely to recur. One just before a reward is reinforced.

The model need not assume that the machine actually feels pleasure or pain. Reward and punishment can take the form of signals that change what it does later.

This is the second substitution in the paper, on a much smaller scale. Two names for things that are felt are replaced with two signals that feel nothing, and the program of education goes through. It is the same move as using a note to stand in for a state of mind in 1936.

He says he has tried methods of this kind. His way of doing it was too unorthodox for the trial to be wholly successful. It is one of the few first-person accounts of an experiment in the paper, and the account is of something that did not work.

Rewards and punishments are not the whole thing. He adds channels that carry no feeling and can convey orders, propositions, and symbolic language. You cannot teach it by reward and punishment alone. You have to be able to talk to it too.

On randomness, he thinks some element of it is a good thing in a learning machine. Otherwise a search would systematically go through many poor possibilities one by one before reaching a good one. He compares the random element to mutation in evolution. Randomness proposes alternatives; selection or evaluation keeps what is useful.

The last paragraph says he expects machines eventually to compete with people in all purely intellectual fields. He offers two places to begin. Start with an abstract pursuit, such as chess, or give the machine the best sense organs available and teach it to speak and write English as one would a child. He does not choose between them.

People went down both roads. One leads to chess and formal games, the other to language. Over the following decades the weight of the field shifted back and forth between them more than once.

The last sentence says, in effect, that we can see only a little way ahead, but that there is already plenty to do in the distance we can see.

8. Afterward

The first public response was a short piece by Leonard Pinsky in the same journal in July 1951. Its title was “Do Machines Think About Machines Thinking?”

It is satire, not a theory set up to rival Turing's. Pinsky suggests getting one of the highly complicated electronic or digital computers Turing describes, taking out a subscription to the same journal for it, and observing its response to the dispute over whether machines can think. That would be an experiment in the other direction. He reads Turing as proposing a decisive experiment somebody will actually run, then uses parody to see what it would prove.

The first public answer was a joke. In 1951, among readers of a philosophy quarterly, his proposal was first of all a thing you could make a joke about.

In April 1952, a different philosophy journal published “Can Machines Think?” by the philosopher Wolfe Mays. He too was in Manchester, in the same city as Turing.

The responses in the first ten years that one can put a name to are these: a joke, an article from the same city, a few scattered exchanges. What was later called a beginning made about this much of a splash in its first decade.

On May 15, 1951, Turing gave a talk on the Third Programme called “Can Digital Computers Think?” This time he entertained the idea of a digital computer as a brain more directly than in 1950. In another talk that year he used a much longer scale. If machine intelligence were one day to surpass people's, it would happen within the next thousand years, not millions of years from now.

In January 1952, four people took part in a discussion on the same program entitled “Can Automatic Calculating Machines Be Said to Think?” They were Newman, Turing, Jefferson, and Braithwaite.

Two moments stand out. In one, Turing says he does not want to define thinking. To make the comparison he proposes, he says, it is not necessary to have a definition.

In the other, the arrangement has changed. In 1950 there is one interrogator. In the 1952 broadcast there is a jury. The task is now to get a substantial portion of them to say the machine is human.

In two years one person has changed the key element of his method from one person to a group.

The judgment of one interrogator and the judgment of a sizable share of a group are not the same thing. One is a single act of telling apart. The other is a vote. Two years earlier he had objected precisely to handing the meaning of a word over to a statistical survey. The vote in one case is on how to use a word. In the other it is on who is human in the game before them. Still, the two things have a very similar shape.

The first appearance in print of the name Turing test has not been located here. Turing did not call it that. A person's name usually gets attached to a method after the person is no longer there. This is the fourth instance in the series: the machine's name, the title given to Lovelace, the two names joined in the thesis, and this one.

The first competition organized expressly around a test of this kind was the Loebner Prize in November 1991, funded by Hugh Loebner. In that first contest the topics were restricted. Human judges used text terminals to talk to people and programs and decided who was human and who was a machine. What took place forty-one years later was the familiar version, not the 1950 arrangement as first described.

Put the accounts that have circulated next to the paper and they look like this.

The story is that Turing proposed a criterion for deciding whether machines have intelligence. In the paper he does use the word criterion. There is no fixed passing score, and he does not say that passing is the same as having intelligence.

The story is that Turing predicted a machine would pass the test in 2000. What comes closest in the paper is the conjecture in Section 6, with its conditions of storage, duration, and probability. Passing, getting a mark, reaching a given year: this is not how it appeared in print.

The story is that Turing thought machines could think. In the paper he considers the literal question not worth discussing, writes down a conjecture about performance in the game, and predicts a change in ordinary usage.

The story is that the test is a five-minute conversation. Five minutes is not in the rules of Section 1. It is in the prediction in Section 6.

The story is that Turing avoided consciousness. The paper gives it a subsection called “The Argument from Consciousness,” running over several pages. He quotes Jefferson at length, discusses solipsism and other minds, imagines an oral examination, and expressly says there is a mystery about consciousness.

The stories have a shape in common. The paper has conditions, qualifications, and disagreements its author marks out himself. The versions that have come down are clean, executable, and come with a score. For a method to be used on a large scale, it first has to be simplified in this way. What gets simplified away is exactly what the author used to mark off how far the method goes. Once the qualifications are gone, what remains is easier to use and easier to take for a boundary.

Finally, a few dates. On March 31, 1952, Turing was tried in Cheshire on a charge of gross indecency. The sentence was not imprisonment but twelve months' probation on condition that he undergo medical treatment. On June 7, 1954, he died at home in Wilmslow, Cheshire. He was found the following day. The official finding was suicide.

Four years earlier, the same person wrote in a philosophy quarterly that he would replace this question with another.

The question put aside is still where he left it.

Nobody has answered it. Nobody has proved it should not be asked. Since that year most of the argument has gone to the question that took its place: to its rules and its length, to whether it counts, to who has passed it. That piece of ground was marked off by one person in 1950. In Section 6 the person who drew the line had already put it in writing: readers would disagree over whether the substitution was appropriate, so the original had to be kept.

That sentence is rarely quoted alongside the other. Quote the first alone and you get a man sweeping the old question out the door. Read both and you get a man moving the old question out of the way and leaving a note beside it to say it is still there.

The machines are still being built. Each generation says again where the boundary lies.