英文本 我们处于怎样的AI时代

In late 2017, a quiet revolution occurred. AlphaZero, an artificial intelligence (AI) program developed by Google DeepMind, defeated Stockfish — until then, the most powerful chess program in the world. AlphaZero’s victory was decisive: it won twenty-eight games, drew seventy-two, and lost none. The following year, it confirmed its mastery: in one thousand games against Stockfish, it won 155, lost six, and drew the remainder.1 Normally, the fact that a chess program beat another chess program would only matter to a handful of enthusiasts. But AlphaZero was no ordinary chess program. Prior programs had relied on moves conceived of, executed, and uploaded by humans — in other words, prior programs had relied on human experience, knowledge, and strategy.

These early programs’chief advantage against human opponents was not originality but superior processing power, enabling them to evaluate far more options in a given period of time. By contrast, AlphaZero had no preprogrammed moves, combinations, or strategies derived from human play. AlphaZero’s style was entirely the product of AI training: creators supplied it with the rules of chess, instructing it to develop a strategy to maximize its proportion of wins to losses. After training for just four hours by playing against itself, AlphaZero emerged as the world’s most effective chess program. As of this writing, no human has ever beaten it. The tactics AlphaZero deployed were unorthodox — indeed, original. It sacrificed pieces human players considered vital, including its queen. It executed moves humans had not instructed it to consider and, in many cases, humans had not considered at all. It adopted such surprising tactics because, following its self-play of many games, it predicted they would maximize its probability of winning.

AlphaZero did not have a strategy in a human sense (though its style has prompted further human study of the game). Instead, it had a logic of its own, informed by its ability to recognize patterns of moves across vast sets of possibilities human minds cannot fully digest or employ. At each stage of the game, AlphaZero assessed the alignment of pieces in light of what it had learned from patterns of chess possibilities and selected the move it concluded was most likely to lead to victory. After observing and analyzing its play, Garry Kasparov, grand master and world champion, declared: “chess has been shaken to its roots by AlphaZero.”

As AI probed the limits of the game they had spent their lives mastering, mastering, the world’s greatest players did what they could: watched and learned. In early 2020, researchers at the Massachusetts Institute of Technology (MIT) announced the discovery of a novel antibiotic that was able to kill strains of bacteria that had, until then, been resistant to all known antibiotics. Standard research and development efforts for a new drug take years of expensive, painstaking work as researchers begin with thousands of possible molecules and, through trial and error and educated guessing, whittle them down to a handful of viable candidates.

Either researchers make educated guesses among thousands of molecules or experts tinker with known molecules, hoping to get lucky by introducing tweaks into an existing drug’s molecular structure. MIT did something else: it invited AI to participate in its process. First, researchers developed a “training set” of two thousand known molecules. The training set encoded data about each, ranging from its atomic weight to the types of bonds it contains to its ability to inhibit bacterial growth. From this training set, the AI “learned” the attributes of molecules predicted to be antibacterial. Curiously, it identified attributes that had not specifically been encoded — indeed, attributes that had eluded human conception or categorization. When it was done training, the researchers instructed the AI to survey a library of 61,000 molecules, FDA-approved drugs, and natural products for molecules that (1) the AI predicted would be effective as antibiotics, (2) did not look like any existing antibiotics, and (3) the AI predicted would be nontoxic.

Of the 61,000, one molecule fit the criteria. The researchers named it halicin — a nod to the AI HAL in the film 2001: A Space Odyssey.4 The leaders of the MIT project made clear that arriving at halicin through traditional research and development methods would have been “prohibitively expensive” — in other words, it would not have occurred. Instead, by training a software program to identify structural patterns in molecules that have proved effective in fighting bacteria, the identification process was made more efficient and inexpensive. The program did not need to understand why the molecules worked — indeed, in some cases, no one knows why some of the molecules worked.

Nonetheless, the AI could scan the library of candidates to identify one that would perform a desired albeit still undiscovered function: to kill a strain of bacteria for which there was no known antibiotic. Halicin was a triumph. Compared to chess, the pharmaceutical field is radically complex. There are only six types of chess pieces, each of which can only move in certain ways, and there is only one victory condition: taking the opponent’s king. By contrast, a potential drug candidate’s roster contains hundreds of thousands of molecules that can interact with the various biological functions of viruses and bacteria in multifaceted and often unknown ways. Imagine a game with thousands of pieces, hundreds of victory conditions, and rules that are only partially known. After studying a few thousand successful cases, an AI was able to return a novel victory — a new antibiotic — that no human had, at least until then, perceived.

Most beguiling, though, is what the AI was able to identify. Chemists have devised concepts such as atomic weights and chemical bonds to capture the characteristics of molecules. But the AI identified relationships that had escaped human detection — or possibly even defied human description. The AI that MIT researchers trained did not simply recapitulate conclusions derived from the previously observed qualities of the molecules. Rather, it detected new molecular qualities — relationships between aspects of their structure and their antibiotic capacity that humans had neither perceived nor defined. Even after the antibiotic was discovered, humans could not articulate precisely why it worked.