| > Tulips were real shipping physical products. Railways were real. Housing was real. Whether or not the demand is speculative is largely disconnected from the actual subject of the bubble. I think this is definitionally wrong, but okay. We're talking about the price of supply chain components, not assets or speculative commodities. Tulip valuations were entirely driven by speculation. Housing bubbles are largely driven by speculation with cheap credit. Railway mania was about a stock market bubble, not the price of rail itself. > Forecasts do not make it so. In publicly traded companies, they tend to be accurate. Are you suggesting that NVIDIA has given improper forward guidance and is about to be subject to major shareholder lawsuits? That's a major short sell opportunity, and I suggest you get on that! > Inference is massively subsidized. The demand is fictitious just because of that. Once prices go up, especially once free or cheap inference dries up, demand will collapse. This is, kindly, incredibly wishful thinking. It may be true for individual companies that are stretching themselves too thin. On the other hand, Anthropic claims they are not subsidizing inference (they have raised prices - it hasn't slowed demand) and will become profitable around 2028 as training costs for LLMs have likely peaked. We'll see. > It hasn't. For all the claims that AI has made any given job so much easier, developers who claim "It'd have cost me a billion years to do so" (next time bring a counterfactual), the actual economic benefit appears to be a big fat zero. We're right back to the Solow paradox. ... I think you may misunderstand the Solow paradox. Paul Strassmann's work on information productivity and "The Squandered Computer" covers a lot of this: there is no correlation between IT investment and productivity or profitability on the firm level, and a serious lag in overall productivity statistics until the 90s, because such gains are only achievable through alignment to meaningful business goals and structural change on processes. It's only when impactful practices become ubiquitous that they reflect in the statistics. The same will be for AI. Which is why I don't think that AI will be as fast an economic transformation as its proponents believe. But it will take place, unevenly to start. > So much of the demand for inference is driven by hype. We agree there. > Companies using AI in the expectation of an ROI that has not materialized, and in many places, is very unlikely to. We agree there too. > Where will these tens of trillions come from if the aggregate economic benefit doesn't exist? Joe Slopman making a dozen CRUD apps a week for half a million in revenue, but there ain't a million of him. Compressed project capital requirements and schedules, and compressed operational expenditure & time for many business processes. I've seen this myself before AI, 25+ years ago, with business process management and enterprise resource planning - think Peoplesoft, SAP, etc., replacing armies of paperwork processors. Same for accounting systems in the 60s and 70s. Another area I have deep familiarity with - "precision railroading" pioneered by Hunter Harrison (a mix of locomotive distributed power technology, classification yard technology, IT systems, lean thinking, and management insight) has led to the most effective freight delivery times we've seen and productivity gains that many thought would be impossible (operating ratios in the sub-60s at times). This took decades to tweak and "get right" and is still subject to to many problems (safety issues, crew stress, and reliability delays to smaller customers due to a lack of slack in the system). Of course, it takes insight, discipline, and design to extract technological gains on an individual firm or agency basis, which is why this stuff doesn't show up in the aggregate statistics until many years afterwards. Business intelligence systems also improved production and market decision making in the 90s and 00's, but that's hard to trace back to IT improvements. I think with AI it is another round of major IT improvements in industries that were previously less susceptible to such change, and as such it will definitely be a multi-trillion dollar impact. But... It will be a mixed bag for a while - much longer than the AI CEOs would have you believe. I don't think this will lead to a crash, but I do think component shortages will eventually end (towards 2028-2029). Some will find significant benefits with it through expressive prompting, appropriate agent structure & guardrails, and deep contexts... and it will take a while for that know-how to spread. > In no small part because any "efficiency" or "productivity" gains realized by AI immediately drives down the cost of the good or service produced. I think there's significant lag on that point. |