Belmont Club: Stealing You

The cyberworld has been disrupted by the launch of a computationally cheaper but high quality AI platform from China: DeepSeek. "It’s been arguably the most discussed company in Silicon Valley ... thanks to the release of DeepSeek-R1, a new large language model (LLM) that performs 'reasoning' similar to OpenAI’s current best-available model o1 — taking multiple seconds or minutes to answer hard questions and solve complex problems as it reflects on its own analysis in a step-by-step, or 'chain of thought' fashion." DeepSeek-R1 is open source and may be found in Github here.

It's chief advantage lies in how much fewer resources it uses compared to comparable American products. But that is unsurprising. Everyone is building on the foundations someone else built. "OpenAI stole from the whole internet to make itself richer, DeepSeek stole from them and give it back to the masses for free I think there is a certain british folktale about this," one publication wrote. Ironically, it was because DeepSeek was constrained by hardware sanctions that it was forced to innovate in software. MIT Technology Review writes that necessity was the mother of invention:

DeepSeek’s success is even more remarkable given the constraints facing Chinese AI companies in the form of increasing US export controls on cutting-edge chips. But early evidence shows that these measures are not working as intended. Rather than weakening China’s AI capabilities, the sanctions appear to be driving startups like DeepSeek to innovate in ways that prioritize efficiency, resource-pooling, and collaboration.

Which brings up an unavoidable Hollywood reference. The new idea is sometimes born out of duress.

Within days DeepSeek was reportedly beeing hit with 'large scale cyberattacks', presumably from its commercial rivals or hostile state powers. Interestingly the reported attacks have not been directed against the service itself but on the new registration process. This forces one to consider what is the prize in this rivalry. An AI platform extracts a lot of information from those that join. It learns about them, and when those new registrants are prominent personalities in politics, finance, technology and defense that can be important. Keystroke patterns or rhythms. Location information. It tells the platform who you are. More importantly it learns about what their customer base is thinking and the way they are approaching complex problems they are trying to solve. It maps the inside of their mind and even looks into capturing user generated data, photos and information domain linkages. Later the AI platform becomes a source of suggestion. It can steer its users in certain directions, plant opinions, recommend products.

The cyberattacks are probably aimed at slowing this migration to the Chinese platform and delaying the consequent harvest of knowledge. In the longer term the cyberattacks might install man in the middleware between the user and DeepSeek. They could be eavesdropping, even modifying the exchange between a user and his AI app. Readers will recall that suspected Chinese hackers recently executed Operation Salt Typhoon. The Salt Typhoon cyberattack, also known as GhostEmperor, FamousSparrow, or UNC2286, infiltrated at least eight major U.S. telecom companies, compromising backdoors created by the Communications Assistance for Law Enforcement Act (CALEA). These intentional weaknesses were designed so the US government could spy on the public. The Chinese then stole the information from the US government in turn. Perhaps the US government is even now stealing the data China stole from you. Everybody is stealing data from you.

In a very real sense the information inside of the human brain is the most valuable training set of AI. The world of drones, space travel, biotech and AI really drives home the fact that a country's defense industrial base is in its human capital. Vivek  Ramaswamy's unpleasant rant against the prevalence of mediocrity of American public education underscores a problem on whose solution the public may differ but ultimately must solve to remain competitive. The path to excellence is not necessarily through the crammers or the focus on academic monomania but surely it is not also in the mindset that produced Karen Bass and Gavin Newsom. No amount of swotting is going to fix the problem of living in a world of words detached from reality. No amount of resources, no coming together will enable a train bound for nowhere to get anywhere. And yet the Chinese train is chugging of that there is no doubt.

To the extent that the toolmaker interacts with his tools AI has to feed back upon its creators. The Chinese Communist Party, the EU or the US government can try to control it, but AI must fundamentally align with the external world or departures from reality will be punished by innovation. You can enforce artificiality but only temporarily before innovation leapfrogs you and puts you in your place. That disillusion can happen quickly. UK Prime Minister Keir Starmer's promise to make the UK an AI superpower, made shortly before DeepSeek was announced, saw it as a way to turbocharge government. Like many other world leaders he thinks he's still in control but maybe he's not even in the game.