{"text":[[{"start":6.65,"text":"The writer is a professor emeritus at New York University and author of ‘Taming Silicon Valley: How We Can Ensure That AI Works for Us’"}],[{"start":15.200000000000001,"text":"Between now and 2030, US hyperscalers like Meta, Microsoft, Alphabet and Amazon are expected to spend over $5tn on compute. It is a huge bet on technology — one that has already led some tech companies to reduce buybacks and issue new debt and stock. What is it that they hope to get for their money? "}],[{"start":35.25,"text":"“Scaling compute” is AI industry terminology for spending more on data centres and the chips that go inside them — GPUs made by Nvidia, for example, TPUs in the case of Alphabet. These power the large language models used in generative systems such as chatbots like ChatGPT, Gemini and Claude. Compute is shorthand for how much computation a given system can do. More compute means adding more chips better able to compute at high speeds in parallel, allowing for the training and use (known as inference) of ever larger neural networks."}],[{"start":69.05,"text":"The result is scale — the supposed magic fix for all that ails LLMs. Adding more data and more compute has resulted in significant AI model improvements. OpenAI’s GPT-3 was much better than GPT-2, for example. Since 2022, computing capacity from AI chips has grown by approximately 3.3 times per year, according to research institute Epoch AI."}],[{"start":94.75,"text":"Unfortunately, there are at least two problems in continuing with this trajectory. "}],[{"start":99.45,"text":"The first is technical. Scaling has not solved some of the core problems that plague LLMs, including hallucinations and occasional reasoning mistakes. Despite a huge increase in compute, outputs remain vulnerable to error. "}],[{"start":114.25,"text":"The second problem is that the pursuit of scaling has become so widespread that there is almost no technical moat distinguishing the leading companies. This has led to price wars coexisting with high operating expenses (needed to run bigger data centres to train and operate new models) and low or even negative margins, since all are building more or less the same product. Meanwhile, more of corporate America appears to be turning towards cheaper, open-source AI models created in China."}],[{"start":142.9,"text":"If there is a deflation of the AI bubble, the optimists say that the new infrastructure will remain even if the companies do not — just as railways survived the 19th-century railway bust. However, this fails to reckon with the reality of depreciation (few pieces of silicon hold their value for very long because better chips inevitably come along) and the possibility that LLMs could be displaced by more efficient models less dependent on massive numbers of expensive AI chips. "}],[{"start":173.4,"text":"In placing massive hyperscaling bets, investors are setting lavish expectations about future earnings. But LLMs are not likely to replicate the near monopolies that have made the market power of current tech giants hard to assail. A better analogy for them might be airlines, which are hobbled by small margins, intense competition, high expenses and dependence on hardware created by outside vendors. "}],[{"start":198,"text":"We could wind up with a lot more data centres than we need. "}],[{"start":202.65,"text":"But the real problem is what the blast radius might be if the hyperscaler bets on compute do not pay off. What will the potential collateral damage to pension funds, banks and the global economy look like? In the worst-case scenario, governments might be called on to bail them out. OpenAI has already tried to float a version of this idea by briefly calling for government loan guarantees for data centre construction, until it was met with public outcry. "}],[{"start":230.45000000000002,"text":"Some form of AI — one that is reliable and efficient and compatible with human safety — might well be worth the investments that are being poured into data centres. To make the bet on the version we have now is premature."}],[{"start":249.10000000000002,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1782438111_2000.mp3"}