Missing from the Future
Reading Karen Hao's "Empire of AI" (part one)
Finishing Hao’s excellent book around the same time that Pete Hegseth declared war on Anthropic amplified for me the way that the left is missing from the struggle over the future. Dario Amodei looks like one of the good guys, sincerely wanting to prevent a horrible administration from using AI for widescale domestic surveillance based on commercially available data as well as legitimately concerned over the negative potential of unsupervised AI in military weapons and targeting. The US government, clearly the bad guys, was saying that private corporations should not dictate US policy, which actually seems reasonable enough. And then another bad guy, the odious and self-serving Sam Altman, weasels his way in as the seemingly principled but actually complicit in imperialist operations such as the kidnapping of Maduro Amodei gets slapped with a “supply chain risk” label. Amodei now looks doubly virtuous, a principled victim who goes down trying to save us not just from the unknown unknowns of AGI — that he is helping create — but from the known knowns of Trump.
Where is the Left? Absorbed in meaningful local struggles against data centers that can’t articulate anything more than arguments anchored in environmentalism and affordability. The direction of the future is being set by feuding tech lords and their billionaire backers, with varying degrees of support from the state. This past week the choice has been about which billionaire capitalist builder of a technology already announced as economically transformative for its impact on cognitive labor will supply weapons to US imperialism for its assaults on people at home and abroad.
The Left isn’t even present in the Silicon Valley battle between boomers and doomers. AI is naturalized as “technology” or “automation,” the conditions that are determining the specific ways it is developing excluded in advance from myopic conversations that boil down to “why shouldn’t we want to make our jobs easier, you Luddite?”
Given the rapid development and adoption of AI, the very alternative of “for or against” is off the table. What we have now is: ignore, refuse, adapt, embrace, regulate, or seize. The first option seems impossible at this point, while the last option depends on the success of a communist revolution unlikely in the near term —although with widespread economic disruption it could be closer than we think. Even beginning to assess any of the middle ones requires thinking through what we are talking about when we are talking about AI. This is where Hao’s book on OpenAI is helpful. She illuminates the conditions under which AI is developing in the US, making us attend to the ways that there is not some kind of neutral or pure AI but, for us, here and now, the development of a specific set of technologies driven by a vision, competition, and funding. For those of us who are not in tech, getting a sense of the specificity of the LLMs released by the frontier labs and how much they are the effects of a billionaire ego-project may be a step toward formulating a better politics.
In 2015, OpenAI is founded as a research-focused non-profit by Elon Musk and Sam Altman (Musk leaves in 2018). They each put up a billionaire dollars for its operations, as does Peter Thiel. Their commitment was to develop artificial general intelligence (AGI) “for the benefit of humanity.” This changes pretty quickly as they start to worry about the competitive race to AGI and confront the enormous costs of its development. In 2019, Altman created a for-profit arm, Open-AI LP, and got a billion dollar investment from Microsoft. The company thus triggers “the very race to the bottom that it had warned about, massively accelerating the technology’s commericialization and deployment without shoring up its harmful laws or the dangerous ways that it could amplify and exploit the fault lines in our society” (14).
Hao emphasizes that AI is a “multitude of technologies that shape-shift and evolve, not merely based on technical merit but with the ideological drives of the people who create them and the winds of hype and commercialization” (15). It’s not all AGI: not only is there significant debate over what intelligence entails, what artificial general intelligence would look like, and whether it’s even possible, AI is also not the same as Chat GPT and Claude. These LLMs are just particular forms that were not inevitable. They were the result of specific choices made by specific people. These choices have costs and benefits. If current trends are any indications, the benefits flow upward: this is why hundreds of billions of dollars is being invested in them. The costs, already present in the impact on people involved in cleaning the training data, are born by the least well off. For Hao, the order being ushered in —forced on us — by OpenAI, its investors, and competitors is colonial. The very fact that the richest tech giants are all-in on this competition is preventing the development of alternative AIs. Far from natural or inevitable, the technology shaping our future is itself being shaped by a war among the great houses. And the effect is ever greater concentration of wealth in their hands.
Hao suggests that smaller, distributed AI models are more promising. This reminds me of the participatory democracy and dotcom promises of the Web, promises which actually ushered in the massive wealth inequalities, capture of data in communication networks, the circulation of rage, and the power of billionaire tech lords driving this new round of enclosure. It’s hard to see how dispersed AI combats any of this.
Hao describes OpenAI as initially sort of clueless: high ideals — get to AGI first and save humanity — but no plan how to do this. The idea that emerged centered on increasing compute, the most calculations, running on the most chips, for the longest time. Combined with ever more data, increasing compute would, it was thought, ultimately lead to AGI. Scale up, scale up, scale up.
The repercussion was the need for more chips/GPUs, which would cost money. Hence the push to become a for-profit company. That meant that OpenAI had to have something investors would value, something monetizable. Its GPT-2 model could generate human-sounding text. This was attractive to Microsoft, who had the money and compute necessary to scale it (as well as an intense desire to compete with Google). In July 2019, Microsoft invests a billion dollars and OpenAI gets locked in to using Azure (Microsoft’s cloud-computing platform), still while trying to create something when no one knows what this something will look like, even as they know making it will take enormous amounts of compute that, incidentally, might cover the whole world in data centers and require unfathomable amounts of energy. What matters, above all else, is getting there first.
Hao demonstrates that AI’s development has been shaped by what she calls a powerful elite but which are clearly predatory tech lords searching for new opportunities for accumulation (94). Generative AI grew up out of the first decade of AI commercialization. Part of the choice for neural networks/connectionism/machine learning over more accurate rules-based symbolic systems was how much more easily the former could be commercialized, especially given the massive stores of data that accumulated through communicative capitalism. This commercialization had effects of its own, choking off alternatives as “companies pumped unprecedented sums into deep learning and connectionism, overshadowing all other sources of funding” and remaking “the landscape of research around their priorities” (105). By 2020, 91 percent of the best-performing AI models would come from industry, concentrated in just a few corporations (106).