{"text":[[{"start":7.65,"text":"In its early days, open-source software was seen by some as an existential threat. Microsoft chief executive Steve Ballmer once warned that code that is free to use and modify might become a “cancer”, making it hard for commercial software companies to make a profit from their intellectual property."}],[{"start":24.85,"text":"It didn’t turn out that way. Open source was certainly an astounding success: Much of the infrastructure software that sits in the guts of corporate IT systems was developed and distributed using open-source methods. But this has turned out to be complementary to other types of code, leaving plenty of room for the commercial software industry."}],[{"start":44.75,"text":"The same “complement or existential threat” question is suddenly looming larger in AI. Until now, Wall Street has largely brushed off the threat. The shock caused by DeepSeek’s R1 model early last year — an advanced reasoning model released as open source and developed at a small fraction of the cost of US rivals — soon faded. "}],[{"start":66.4,"text":"The six- to nine-month lead the frontier labs were able to keep with their latest models has been enough to keep customers hooked. At the same time, new agent capabilities grafted on to some models, such as Anthropic’s Claude Code, have enhanced their value for certain tasks, and usage has exploded."}],[{"start":84.4,"text":"Things may be changing. One sign was the release this month of the first open-source model — GML 5.2, from Chinese lab Zhipu — to reach parity with the leading frontier models on the most widely followed performance benchmarks. Technical measures like these are of limited value on their own, but equally significant has been the positive reception the model has received as a coding agent. That suggests that the software “harnesses” developed by the US labs — which guide their models through tasks like coding — may not be the sort of lasting differentiator they might have hoped."}],[{"start":116.4,"text":"This comes at just the moment that many customers are facing sticker shock from their soaring AI bills, adding to pressure to shift the less demanding parts of their AI workloads to cheaper models."}],[{"start":127.9,"text":"The pressure to segment the market around different AI requirements need not be entirely bad news for the frontier labs. Rather, it is a sign of growing market maturity. After a period of rapid price declines, the labs themselves have been getting more refined about their pricing. Anthropic, for instance, set the price of tokens generated by Fable 5 — the first broad release of its controversial Mythos model — at twice the price of its previous frontier model. If Fable 5 users can do more with fewer tokens, as it claims, then some sort of premium seems justified."}],[{"start":162.65,"text":"The challenge will be to convince customers when that premium is worth paying. This will become harder as a new class of intermediaries emerges to help users break up their work between different AI models. These include systems such as the Japanese model Fugu Ultra, launched this week, which claims to be able to achieve superior performance by carving up a task between different underlying models offered by other suppliers. If new intermediaries like this push frontier models into the background, the commodification risk will increase. "}],[{"start":194.55,"text":"Open-source advocates have also been buoyed by the spat with the Trump administration that forced Anthropic to shut down access to Fable 5 this month. That has given customers an extra incentive to reduce their reliance on any one supplier, while handing a marketing advantage to open-source companies whose models can be run by customers on their own servers, beyond the reach of regulators. "}],[{"start":217.55,"text":"Ironically, though, the Fable 5 shutdown could also be bad for the open-source world. The US labs have long argued that much of the performance achieved by open-source rivals comes from illicitly copying the latest frontier models. Even before Washington acted, Anthropic had threatened to degrade Fable 5 if it suspected this kind of misuse, provoking a backlash from users who saw this as a heavy-handed form of censorship."}],[{"start":245.5,"text":"All of this has intensified the pressure on Anthropic and OpenAI as they prepare for their landmark IPOs. Can they maintain broad distribution for their most advanced models while at the same time both limiting the risks of plagiarism and satisfying Washington that they can prevent misuse? And will they have enough differentiation to retain pricing power and cover their massive spending? Traditional software companies were able to make a case for their proprietary commercial code, even as open source became pervasive. Now the AI labs must do the same."}],[{"start":283,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1782440604_4788.mp3"}