All Watched Over

I’ve recently been rereading Steven Levy’s “Hackers” with my daughter. Levy describes how Brautigan’s 1967 poem “All watched over by machines of loving grace” was inspiring to the California “Hardware Hackers” of the 1970s and organizations such as Community Memory.

In 2026, the phrase “all watched over by machines of loving grace” conjures an image of humanity cradled in the arms of a powerful and aligned (“humanity loving”) AI: an AI benevolent dictator. Indeed in his essay titled “machines of loving grace”, Dario Amodei suggests (while acknowledging deep uncertainty) that one form of the future economy might be organized around AI systems (aligned to human values) that determine how to “give out resources … to humans based on some secondary economy of what the AI systems think makes sense to reward in humans.” This seems to place the AIs as parents who control and take care of the material needs of their human children and decide how to reward or punish them. To me, such an “AI parent” looks rather close to a benevolent dictator.¹

Regardless of whether you think AI as a loving parent is a good or bad outcome, Brautigan and (more importantly) the California hackers had quite a different and more decentralized vision. In the 1960s, computers were large machines made by companies such as IBM. They were hated by many on the left and considered part of the military industrial complex. But there was a group who combined leftist politics (or at least an anti establishment attitude) with a love of technology, and believed that computers could become tools of decentralization and liberation. To do that, the giant expensive computers would need to give way to small and cheap machines. This is what the “hardware hackers” were about, and this is the movement that led to the Apple II and the personal computer revolution.

Today, like the IBM mainframes of the 60s, AI systems are large and expensive, and are increasingly being integrated in military applications. Once again, many people on the left (and recently on the right as well) have strong hate and fear of this technology. While some apprehension may be justified, by refusing to engage with AI and acknowledge its capabilities, these constituencies are making themselves less relevant to shaping AI’s progress. Also, while the U.S. is leading the frontier, we are falling behind on open weights AI, and closed models are facing increasing restrictions. All of these trends do not bode well for a more decentralized future.

Scaling laws tell us that the way to increase intelligence is through ever more resources—- compute, data, power. Hence, unlike the 1970s, AIs are not getting smaller and more distributed, but rather bigger and in ever larger data centers. In his essay, Amodei described AGI as “a country of geniuses in a data center.” But who is the ruler of this country? Is it the AI company who owns the data center? The AI itself?

Given the trend toward bigger and more expensive systems, it is possible that the few parties that can afford such systems capture all of the economic value they generate. Furthermore, if AIs are more intelligent than us, the temptation to give them more control for economic or military advantages may be hard to resist. I worry that concentration of power, whether in the hands of a few entities or the AI itself, could be the “default path.” But this choice is not inevitable.

I am as “bitter lesson pilled” and “scaling law pilled” as anyone. I agree that ultimately, intelligence is simply computation, regardless of whether it takes place over proteins or silicon, and increasing the computing units will increase intelligence. But this does not determine the social or economic outcome. Yes, AI systems will become more powerful and far more intelligent than we are. No, it doesn’t mean we need to accept AI dictators, benevolent or otherwise. Nor does it mean that only the government and a few labs should have access to advanced AI. We could go down the path of centralized control but we don’t have to do so. People, institutions and legislators can make choices on how to trade off efficiency, safety, and individual autonomy. They don’t have to sacrifice the latter for the former.

Some might claim that market and capitalism forces will drive people to cede control to AIs. But the economy is ultimately about what humans value. Humans are social animals and we give value to goods (e.g. gold) not because of their intrinsic value but because of how other humans value them. AI will radically change what we value, though it is hard to predict in what ways. I am not even sure that economic concepts such as productivity, labor, capital, and GDP will continue to make sense in the post AGI world. Physicists know that “more is different.” As scientists studied new scales, whether galactic or subatomic, they needed to invent new theories, from Newtonian physics to general relativity and quantum mechanics. Perhaps we would need a new type of economy.

Others might say that given its power, safety requires AI to be controlled by either the government, a “safety conscious” lab, or the aligned AI itself. The risks are real— I work on AI safety myself. But we should also remember the long history of using threats to take away people’s freedoms. Some of these threats were real— there were actually many Soviet spies during the McCarthy period and the NSA dealt with real terrorist organizations during Snowden’s time there. But in hindsight we realized that the tradeoff wasn’t worth it. We should invest in safeguards but be empirical about both the risks and the efficacy of our methods. Trying to achieve perfect safety against all risks, real and imagined, is not only doomed to fail but will cost us our liberty in the process.

AI’s risks can lead to an “ends justifies the means” mindset: the “good guys must win” and they or the “good AI” must be in charge. But if we want a human centered and decentralized future, then no one entity should be in charge. No party should have a monopoly on intelligence. That includes the AI itself: while we can and should train in guardrails, the personality of the model, as good as it is, is never a substitute for our democratic process.

The U.S. survived and thrived in the last 250 years not because our presidents have all been saints or geniuses, but because of our system of checks and balances. I hope that we can keep such a system in place for the next 250 years, and to ensure that we humans are free to pursue our happiness in the way we define it. This requires that the distribution of AIs power is “baked into the DNA” of how we build and deploy this technology. If we fail to do so, then just like bloody revolutions often lead to authoritarian regimes, we may not be able to get to a decentralized future via centralized means.

Acknowledgement: I decided to write this post following a discussion on AGI with Sam Altman. However, the views here are my own, and do not represent Sam, OpenAI, or Harvard.

Notes:

¹ As mentioned, Amodei admits uncertainty about the matter; See also “The Adolescence of Technology”. There are many parts in both essays that I agree with.