Dario Amodei became famous for comparing the best language models with nuclear weapons. The company he runs has also regularly said that it supports regulating access to artificial intelligence, while for several months it has been convincing the world that its latest language model, called Mythos 5, is too dangerous to be immediately made publicly available to everyone. These statements and calls at the end of last week brought results – although probably not quite what was expected in Anthropic.
This is what the US administration did at the end of last week ordered Anthropic to suspend access to the latest models of Fable 5 and Mythos 5 for all non-US citizens — including those residing in the United States, and foreign employees of the company itself. Anthropic decided that this is what it means in practice the need to disable both models for all customers. Access to other Anthropic models has not been blocked.
Lock history of the latest Anthropic models
The story of how Anthropic's latest language models came to be disabled is a bit like a story from a spy movie.
According to the description of events published by Politico and Reuters' reports, it began with talks between US administration representatives and the head of Amazon, Andy Jassy. From them they learned that at least a partial bypass is possible security measures sewn inside the Fable model, i.e. a limited public version of Mythos that should not be used for, among other things, cybersecurity work. After deliberations, they contacted Dario Amodei to discuss the matter and, in short, the parties did not reach an agreement.
According to available reports, government representatives wanted Anthropic to do something about the found ways to bypass the limitations of the Fable model and turn it off for some time, while Dario Amodei explained that the problem was not so big that the entire model would have to be turned off. Ultimately, it ended with an export directive, the effects of which we can see now.
There is also information in the media that the White House may have been afraid of groups associated with China gaining access to Mythos, although this has not been officially confirmed anywhere yet. However, these fears would not be completely groundless, because Anthropic has already had a mishap, as a result of which a certain group operating within the Discord messenger had unauthorized access to this model for two weeks.
Shock therapy for the world of the latest technologies
The practical effect of disabling access to the most powerful Anthropic models is minimal for now. Mythos was made available only to selected companies and institutions, while Fable was available only for about three days in a rather limited form, so no one has managed to actually implement it into their processes yet.
Much more important, however, is the shock caused by the fact that something like this happened. The mechanism used for this purpose is also important. So far, American export restrictions have been associated primarily with chips, the machines needed to produce them and China. Now direct access to the AI model is in the spotlight, and the results are felt around the world. Something that was only theorized about has now become a permanent part of the risk management manual for institutions and businesses.
Winners and losers
Without a doubt, it is very a strong gust of wind in the sails of initiatives related to the so-called digital sovereignty. It has been talked about for a long time, but one often gets the impression that behind the many words there has been little concrete action – perhaps because those who said them hoped that they still had some time. Now, developing projects related to alternatives to the Anthropic and OpenAI models, as well as our own, fully local infrastructure to support these alternatives in many corners of the world, has become a necessity in practice.
At least for now, Nvidia looks like one of the biggest potential beneficiaries of this situation. Jensen Huang's company has long seen its pursuit of technological independence as the source of its greatest future growth, which is why it changed its reporting method in the last fiscal quarter, separating hyperscaler orders from other orders.
Nvidia is also working intensively on its large language models from the Nemotron family and has even created a group working on its next generations, which includes, among others, one of the European AI leaders, the French Mistral. By default, this is supposed to be the main choice for companies and institutions that want to have control over their artificial intelligence and data, but do not want to use popular open models developed in China.
For similar reasons The potential beneficiaries of this situation include the largest companies dealing with network infrastructure and custom chip designsuch as Broadcom and Marvell.
Many additional orders will probably also reach system integrators such as Dell and HPE in the near future.
Sovereignty has layers
However, two problems overshadow the above assumptions.
The first is that For now, there is little chance of fulfilling additional orders for AI infrastructure. All computer memory production is booked for this year and next, and much of this is thanks to hyperscalers and leading AI labs such as OpenAI and Anthropic. The situation with Mythos and Fable may lead to a change in plans related to infrastructure expansion by some of these entities, because certainly at least some of the potential recipients of cloud services and buyers of “American” tokens have just started to seriously think about alternative solutions.
However, it is difficult to say how big these changes would be, how many of these orders could be canceled in practice (in the case of memory, they are often associated with several-year contracts and prepayments) and therefore what part of them could be taken over for the implementation of projects related to digital sovereignty.
Another problem is that in practice, digital sovereignty, which involves ordering chips from Nvidia, AMD, Intel or Broadcom, even if you have your own language model, may in many situations have little to do with sovereignty and everything to do with reducing the degree and form of dependency. At last The US administration has already implemented global export restrictions on the sale of accelerators.
Dario Amodei – founder and CEO of Anthropic
|
Kimberly White/Getty Images
However, this does not change the fact that the global blocking of access to some of the most advanced language models in the world has set off the first domino. It's hard to say where the last one will fall, but the Anthropic thing won't end with Anthropic. Even if Fable and Mythos return to customers quickly, the reputational damage to the entire model of using closed, remotely hosted AI systems has already been done.
