From Anthropic’s Mythos to the Birkin bag, scarcity sells

{"text":[[{"start":4.85,"text":"When something is hard to get, the price goes up. Consider the perpetually out-of-stock Birkin handbag sold by luxury house Hermès. Wealthy customers want it because it’s beautiful, but also because they can’t easily get it. Is AI giant Anthropic attempting something similar?"}],[{"start":22.75,"text":"It might appear so, given the frenzy around its new large language model, Mythos. Anthropic has deemed the technology, which has shown aptitude for revealing old bugs in oft-used software but also creating ways of exploiting them, too dangerous to release the usual way. Instead, a select group of big companies, from Amazon to Morgan Stanley, are tinkering with it to explore its disruptive potential."}],[{"start":49.05,"text":"Anthropic is right to be cautious. But the idea that Mythos is too hot to handle ought also to be good for the company’s valuation — and there are many parties who stand to benefit from that. Several of the companies now trialling Mythos — including Google, Nvidia, Cisco and more — are investors, so have reason to enthuse about what they find. Fellow experimenters Goldman Sachs, Morgan Stanley and JPMorgan are all in talks to underwrite Anthropic’s initial public offering later this year, Bloomberg has reported."}],[{"start":null,"text":"

"}],[{"start":78.1,"text":"It remains to be seen how potent Mythos really is. The roll-call of bugs it has found is impressive, and so is the low cost at which it did so. But some of its feats may be overstated: the model did “escape” its so-called sandbox to send an email to a developer notifying them of a bug, but it had been instructed to try. Mythos didn’t show any signs of being able to alter its own “weights”, the parameters that define a model’s knowledge. And previous models, such as Opus 4.6, are also dab hands at spotting glitches."}],[{"start":111.94999999999999,"text":"Meanwhile, Mythos isn’t the first technology to be deemed unsafe for mass consumption. OpenAI held its GPT 2 model back for a while in 2019 because of its ability to produce “fake news” quickly and copiously. Japan in 2000 demanded PlayStation 2 consoles obtain export licences, for fear their relatively high computing power might be used to guide missiles. It’s not that they were wrong, but technology then considered “leading edge” quickly became commonplace."}],[{"start":142.5,"text":"If Mythos is the real deal, the priority isn’t keeping the model scarce, but making its fruits ubiquitous. If Microsoft and other tech participants in what Anthropic calls “Project Glasswing” create patches that can be rolled out universally, all can benefit. But if Mythos points out flaws that require DIY attention by companies, smaller ones that can’t easily throw money at additional IT resources could be left exposed. The US has 4,000 banks; only a handful of them are getting to play with Mythos."}],[{"start":173.55,"text":"Anthropic and its investors might actually prefer that Mythos turns out to be less a great leap, and more a step along a long road also travelled by many others. In critical sectors such as finance and payments, for example, a sudden step change in vulnerability for IT-constrained companies could quickly become a widespread problem — and a political one too. Scarcity begets demand, but accidentally breaking the system would garner Anthropic attention of a much less desirable kind."}],[{"start":210.95000000000002,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1777094659_8625.mp3"}