The contributors to this anthology share a general consensus that nothing in principle prevents us from creating artificial intelligence -- not quantum gravity, not some secret mystical élan vital. Other than that, there are numerous disagreements. Naturally, this is healthy for the field -- let a thousand flowers bloom! I confess that my own background and sentiments are close to Dennett's and Lenat's, although I might stress the role of learning somewhat differently than they do. It will be hard indeed to go from low levels and up through the enormous range of levels to create the exquisitely organized systems needed for language or scene analysis. In my own area of expertise, pattern recognition (admittedly a small part of the AI puzzle, though an essential one), we seem to understand the central principles. It is applying them that has proven extremely difficult. Even if Kurzweil is right in believing that we can someday reverse engineer a brain, such an effort would tell us little about how the brain learns or represents information.

The fact that we have not achieved AI (for whatever reason) should not, however, blind us to the fact that in some domains we have met and even surpassed the vision of a HAL. As David Kuck (chapter 3) points out, recent advances in computer design and the spectacular and continuous rate of improvement in hardware summarized by Moore's law tell us that we could soon build a computer the size and power of HAL. Ravishankar Iyer, reviewing the progress in computer reliability, shows how we could make the hardware of a massive computer like HAL reliable, even for a prolonged space mission. Alas, the techniques for insuring reliable hardware are more effective than those for software, and this is surely an imposing, but not insurmountable, barrier. In principle, there seems to be nothing to prevent us from making a large computer that is fault-tolerant. Does this sound too much like the hubris of HAL and his designers? As Frank points out in that context, making a computer with no errors "sounds a little like famous last words."

Still, in limited-application domains, we have made steady improvements. We have computer chess systems that beat all but a few dozen human grandmasters, and they are improving every year. It seems all but certain that the world's best chess player will soon be a computer -- perhaps even by the year 2001. In chapter 5, Murray Campbell, a member of IBM's Deep Blue team that challenged Garry Kasparov in February of 1996, analyzes 2001's chess scene in fascinating detail. Campbell also reflects on the changing public perception of chess machines, as illustrated in the film.

We have made several important strides toward automated speech recognition, especially in the initial stages of transcribing raw sound into phonemes. Kurzweil summarizes that progress and applies his own commercial VOICE speech-recognition system to the 2001 soundtrack. He finds that the system's simple phonological transcription is good, and can be particularly accurate when restricted to just two talkers (e.g., Frank and Dave in the quiet spaceship). General speech recognition, however, relies very heavily on semantics, common sense, context and world knowledge. (For example, how do we give a computer the ability to distinguish among such homonyms as their, they're, and there?) We are still far from solving problems like this one.

Similarly, in my own research (chapter 11), I have developed speechreading (lipreading) systems that use both sight and sound. These systems outperform purely acoustical speech recognizers in noisy rooms. Alas, no current system even remotely approaches HAL's proficiency at speechreading in silence. In fact, the lipreading scene in the pod is the only one Clarke thought was technologically implausible for the year 2001. Automatic speech recognition and speechreading will always be limited by the problems of representing semantics, common sense, and world knowledge -- profoundly difficult issues that will occupy science for many decades to come.

As humans, we take our faculties of vision for granted, but making computers see has proven to be extremely difficult. In chapter 10, Azriel Rosenfeld discusses some of the successes of research in "early" vision, such as edge and motion detection, face tracking, and the recognition of emotions (which he illustrates with images from 2001). Full vision, however, would include the ability to analyze scenes. Imagine, for example, a computer that could look at an arbitrary scene -- anything from a sunset over a fishing village to Grand Central Station at rush hour -- and produce a verbal description. This is a problem of overwhelming difficulty, relying as it does on finding solutions to both vision and language and then integrating them. I suspect that scene analysis will be one of the last cognitive tasks to be performed well by computers.

Other capabilities have proven remarkably difficult to develop, including one that Clarke wasn't sure we could solve by 2001: making a computer produce natural-sounding speech. In chapter 6, Joe Olive reviews the development of computer speech generation, such as the systems needed to convert text to speech for the blind. Although these experimental systems work adequately for short utterances or single words, with sentences (let alone entire speeches) they cannot convey the human subtleties of stress and intonation. A convincing artificial speaking system would require the system to understand what it is saying -- again, an extremely hard and unsolved problem.

Reading Roger Schank's discussion of language in chapter 8 is like seeing the Wizard of Oz behind the curtain. With a few simple tricks, we could duplicate some of HAL's linguistic performance. For instance, it would be a fairly straightforward task to record a large number of canned stories and have HAL play them back when he hears appropriate "trigger" questions. Such a program might even persuade a casual or unsuspecting person that HAL understood language. Indeed, a noted computer program (ELIZA) around the time of 2001's release, sought to mimic a Rogerian therapist; in limited dialogues it convinced naive users that they were conversing with a real person. Such demonstrations tell us more about the novelty of computer discourse of the time and the gullibility of users than they do about true machine intelligence. ELIZA and its descendants are a far, far cry from true language understanding and general intelligence.