Can Machines Think? A Guided Tour of Turing's Test, Its Critics and Its Successors

Fridman opens the first paper of his AI paper-reading club with Alan Turing's 1950 "Computing Machinery and Intelligence", asking whether machines can think. Although it is on the surface a philosophy paper, he calls it one of the most impactful first steps towards engineering intelligent systems, and possibly the most impactful paper in the history of AI. In his view it planted the seed and the dream that machine intelligence is possible, inspiring researchers from computer science and beyond.

He reads and praises the paper's opening lines. Turing says the question should begin with definitions of "machine" and "think", but warns that settling them by how people normally use the words would reduce the answer to something like a Gallup poll, which is absurd. Instead he replaces the question with a related and less ambiguous one: the imitation game. A human interrogator passes written notes to two hidden entities, one human and one machine, and must decide which is which.

Fridman sees this as turning a profound but ambiguous question into a benchmark, and notes it echoes related questions such as whether machines can be conscious, fall in love, make art or enjoy chocolate cake. Turing predicted that by 2000 a machine with 100 megabytes of storage would fool 30% of humans in a five-minute conversation, and that people would stop finding the phrase "thinking machine" contradictory. He also expected machine learning to be critical to success. Fridman separates the paper's two claims: that the game is a good test of intelligence, and that machines can pass it.

Fridman asks why people still find "thinking machine" contradictory when Turing predicted otherwise. Chess was once seen as the height of intelligence, and we credit Garry Kasparov with great intelligence, yet we withhold even an inkling of that from Deep Blue after it beat him. Objections such as brute force or understanding how AlphaGo learns follow, but he suspects something psychological, almost a fear of an AI that passes the test.

The Loebner Prize, running since 1991, offered $25,000 for a text-only system that passes and $100,000 for one using other modalities. Current rules involve a 25-minute conversation and fooling 50% of judges, and the winners over the past decade, Mitsuku and Rose, are mostly scripted, rule-based chatbots. Fridman finds it sad that the prize is, as far as he knows, no longer funded, and surprising that DeepMind, Facebook AI and other major groups have not taken on the challenge. A sample Mitsuku transcript shows small gaps, such as repetition and failing to follow tangents.

He notes that nobody has researched where exactly the test fails; he would like conversations marked for the more and less human-like moments instead of one verdict at the end. The best-known claimed pass was Eugene Goostman in 2014, which fooled 33% of judges by playing a 13-year-old Ukrainian boy with quirks, humour and a habit of steering the conversation back to comfortable ground. Fridman says the lesson is that the rules can let tricks stand in for deep conversation. A transcript of an expert judge, Scott Aaronson, shows that a skilled interrogator refuses the misdirection.

Despite the waning popularity of the Loebner Prize, Fridman says, Google published Meena, an end-to-end deep learning chatbot with 2.6 billion parameters aimed at generating text that fits the conversational context. It also proposes a two-part metric. Sensibleness asks whether a response makes sense in context, and humans score 97% on it.

Sensibleness alone can be gamed by being boring, with generic replies like "I don't know" that fit any context, so specificity asks for something particular to the conversation. Fridman says the beauty, music, humour and wit of conversation come from playing with the specifics. In the paper's combined score, humans reach 86%, Meena 79% and Mitsuku 56%.

He cautions that these numbers should be taken with a grain of salt because the system is closed source and the paper feels somewhat like PR, and says we are far from concluding that a system has human-level conversation. Even so, he finds the idea interesting and thinks that learning-based, end-to-end open-domain conversation is clearly the long-term future, just as Turing predicted 70 years earlier.

Fridman summarises the nine objections Turing himself addressed. The religious objection ties thinking to a soul given by God, and Turing replies that an all-powerful God could give souls to artificial creatures. The "heads in the sand" objection is that human-level AI is scary and far away, so we should not think about it. Fridman notes this still surfaces today in discussion of existential risk, and says feelings about whether something will happen should not stop us from thinking about it.

The Gödel objection, associated with Roger Penrose, says computation is limited and cannot be perfectly rational, and Turing answers that humans are not perfectly rational either, so intelligence need not mean infallibility. On consciousness, Turing separates being conscious from appearing conscious. Since we only know other humans appear conscious, there is no reason machines cannot appear so, and the test sidesteps the deeper question. The "machines will never do X" objection, Fridman says, is a vapid opinion based on today's world.

The Lovelace objection, that machines can only do what we program them to do, is the one Fridman finds most interesting. Turing rephrases it as machines being unable to surprise us, and Fridman points out that a sufficiently complex program, even one whose parts you wrote yourself, behaves in ways you cannot predict. Turing also replies that a large enough digital computer can approximate an analogue brain, and that humans may be a complex collection of deterministic rules. The mind-reading objection, which Fridman reads as possibly Turing trolling, is answered with a telepathy-proof room, and he cautions that we know little about how the mind works.

John Searle's 1980 Chinese Room is the most famous objection, and Fridman sees it as a combination of the consciousness, Lovelace and determinism objections. A person who does not speak Chinese sits in a room following rules to answer notes in Chinese, and so is not truly understanding. Searle's key claim is that syntax alone is neither constitutive of nor sufficient for semantics.

Fridman notes the same criticism is aimed at today's language models, such as OpenAI's GPT-2, which some say only mimic the statistics of language. As an engineer he finds the argument unsatisfying because it amounts to saying humans are special and machines cannot understand. It does not say where computation hits a wall.

The open question for him is whether mimicking thinking, consciousness or love is the same as the real thing, and he goes back and forth depending on the day. He sides with Turing's spirit that engineers can currently only build the appearance of thinking, consciousness and love. He believes that in doing so we will begin to understand the fundamentals, and admits that saying the appearance of consciousness is consciousness is a bit poetic.

The Total Turing Test of 1989 adds perception, computer vision and robotic manipulation. Fridman suggests that a narrow channel such as natural-language conversation might actually be harder to pass than one with extra modalities, and praises the original test for its simplicity. The Lovelace test of 2001 asks a machine to do something that its creator cannot explain, and Lovelace 2.0 in 2014 narrows surprise into producing an impressive piece of creative work.

He says surprise is at the core of how we conceive of intelligence: when we finally build a human-level system we will surely be surprised, and humour is among the hardest cases he has in mind. Creativity-based variants, though, are more subjective than the original test.

The Truly Total Turing Test of 1998 proposes judging a body of work produced by agents across their evolution rather than an isolated agent. Fridman does not find that convincing but likes the idea of judging systems over months or years, such as AlphaZero not on one game but on the millions it played while learning. Intelligence, he muses, may be the journey and not the destination, though he cannot make it concrete.

The Winograd Schema Challenge uses sentence pairs such as a trophy that does not fit in a suitcase because it is too small or too large, where common-sense reasoning resolves the ambiguity. Fridman likes that it is clear, simple, needs deep reasoning and has a right answer without human judges. The weakness is that every question is handcrafted, so it cannot scale to millions, though variants appear in current natural-language benchmarks.

The Alexa Prize, he says, captures the Turing test's spirit through voice conversations with real people on Alexa devices; the bar is a conversation of 20 minutes or longer, in two-thirds of interactions. Duration is a strong metric because people speak with their actions, staying in a conversation when it has meaning. He notes that no team has come close, and finds it surprising that Amazon limited it to students and that major labs have not focused on the problem.

The Hutter Prize rests on the idea that compression correlates with intelligence: compress one gigabyte of Wikipedia as small as possible. The current best is a compression factor of 8.58, about 117 megabytes, with 5,000 euros for each 1% improvement. Fridman finds it fascinating but notes it sets no bar beyond which we would be impressed, unlike the Turing test.

François Chollet's Abstraction and Reasoning Challenge, ARC, was proposed a few months before the video and is running as a Kaggle competition. Fridman has not fully digested it but describes it as set in grid worlds of coloured cells, staying close to the psychometric IQ tests used on humans. By going down to basic reasoning elements, it makes explicit the priors a test-taker brings, so the test comes as close as possible to measuring reasoning.

He gives examples from Chollet's paper "On the Measure of Intelligence". Object persistence means large objects behind noise still exist, spatial contiguity means touching cells of the same colour form an object, and colour contiguity means differently coloured neighbours are separate objects. Tasks show three input-output pairs and ask the test-taker to produce the fourth output.

In one task the implicit goal is to complete a symmetrical pattern, relying on a prior concept of symmetry. In another it is to count unique objects and select the one that appears most often. Fridman notes that such tasks are hard to generate, but finds many interesting technical and philosophical ideas in the benchmark.

In his takeaways, Fridman asks whether the test measures intelligence or humanness, since conversation also carries irrationality, laziness and boredom. It judges external appearance, not internal processes, which he thinks is all we can quantify for poorly understood concepts like consciousness. He adds that the interrogator's skill matters, both in challenging the bot and in identifying humanness.

He raises two further issues. Anthropomorphism may be cheating, or may be essential, since we might need to anthropomorphise something before we judge it intelligent. And why judge a system on such a narrow window of time and expect a bot to feel like talking? He imagines an *Ex Machina*-style test in which a relationship builds over days.

Fridman's view is that natural-language conversation is close to the ultimate test of intelligence, and he disagrees with François Chollet, Stuart Russell and others who see the Turing test as a distraction. In his view it keeps researchers honest about where they stand, and Loebner- and Alexa-style competitions deserve more research attention. He adds that AI researchers tend to avoid the messiness of humans, in fields from human-robot interaction to semi-autonomous vehicles, and should embrace it.