I’ve just finished reading Garrison Lovely’s recently released book, “Obsolete.” I wish I had time for a longer review, but I’ll give a brief one and highlight some particularly interesting sections below. I think it’s a great introduction to non-technical readers and even as a pretty dedicated reader of AI policy and news I learned a good bit. It’s deeply, deeply researched and has a good pulse on the alignment community, and Lovely is careful not to understate or overstate his case. The core argument is that AI development should be analyzed through the lens of mass or universal labor displacement, and that a mass movement to shape, pause or freeze development is much more possible, much more positive, and capable of greater breadth of salience and support than is often argued in the alignment community.
He makes an, in my opinion, decisive case that “competitive pressures” by themselves don’t dictate that AI development is fated to happen, and certainly not that it’s fated to happen at the pace the frontier labs are setting. But I wish he’d been a bit more careful about the limitations and subtleties of the plan he’s advocating for; I share the concern raised by folks like the AIFP that it’s hard to pause indefinitely at any given point for very long and that the cost to train a model at any fixed intelligence level, once first reached, is falling precipitously, and that many well-capitalized actors currently want to and are able to speed up development precipitously. I think the book somewhat elides the difficulties we’d still have even if this number were close to 0.
It was likely impossible to train GPT-6 in 2016, but it may be extraordinarily cheap to train a GPT-6-level model in 2036, and even just extrapolating the current rate of progress naively, it might be extraordinarily cheap to train a ~GPT-8 level model in 2036. That seems like a world where many irresponsible agents are deploying very powerful technology even if we have verification mechanisms preventing further uncoordinated capabilities training.[1]
But one thing I do find refreshing is Lovely's relentless focus on partial solutions. They're interspersed throughout the book but come into full focus in Chapters 17 and 18. There's a model of the world that relies on large paradigmatic shifts, where progress on a task is more or less bottlenecked on "true understanding." I think this is sometimes a good instinct in a field as anti-inductive as alignment, where sometimes you really are barking up a useless or even actively counterproductive tree and no amount of just trying harder will fix that. but I think the ugly underside of it is the current situation where much of the field was aggressively averse to policy or advocacy and is now attempting to rapidly build governance infrastructure now that it seems people might actually listen to you. Sometimes you need a list of your most presentable ideas, even if your best ideas have known flaws. And again, not all partial solutions are technical; a pause buys valuable time across many worldviews.
Overall, I still think this book is a well-worthwhile read even for many technical researchers, and even for many people who will fundamentally disagree with the book’s premise. Being better informed about your field holistically is very valuable. Many conceptions of research taste say something like that it’s the art of asking increasingly good questions within your paradigm. But another important part, if making the world a better place is a meaningful consideration in your ethics, is analyzing whether your paradigm would be good to advance in the first place. Do we want to get what we want, or the subset of what we want that we can measure? I think this book is a compelling anchor for discussions of “what do we want out of the development of artificial intelligence?”
On Leading the Future and other lobbying efforts
“[Leading the Future] was the brainchild of OpenAI chief lobbyist Chris Lehane. He cut his teeth smearing Bill Clinton’s accusers in the 1990s and moved on to orchestrating dirty tricks for presidential campaigns, earning him a reputation as a master of the “political black arts.” He’s since become one of the most effective tech lobbyists, joining after stints working for Airbnb and the crypto industry. After dealing with him firsthand, one former OpenAI employee told me Lehane’s “an animal”, clarifying that he “will do whatever to get power and doesn’t really think a lot about greater good type stuff.” During the 2022 midterms, Lehane organized a crypto industry super PAC that intimidated Congress into embracing digital currencies using whatever political attacks were most expedient (which never mentioned crypto). As one person familiar with the PAC’s strategy told the New Yorker: “If you are even slightly critical of us, we won’t just kill you— we’ll kill your fucking family, we’ll end your career.”” (pp. 161-162)
"Behind closed doors, Nvidia appears willing to do whatever it takes. When a Republican congressional proposal sought to codify export controls, the New York Times reported the company "used an unconventional lobbying blitz to smear the bill as "a product of left-wing paranoia peddled by people it calls AI doomers." And there are multiple reports of the chipmaker trying to get opponents fired from think tank and government roles over their work on chip restrictions. Samuel Hammond, chief economist at a conservative think tank, was one of the few to speak about this publicly: "I often criticize Nvidia for putting sales to China ahead of U.S. national security. At least twice this year, they've reached out to complain or worse." (p. 263)
"[Americans for Responsible Innovation] had just helped successfully lobby the Senate to include the GAIN AI Act into the must-pass 2026 defense authorization bill, requiring chip companies to sell to US-based customers before selling to China. Since Nvidia's already over-subscribed, the provision would function much like export controls. Despite its US-first flavor, GAIN —after opposition from both industry and [White House AI Czar David] Sacks, didn't make it into the final bill." (p. 264)
"Of course, [Connor] Leahy's team did have a motive, they just weren't shy about what it was — apparently unlike the techies the MP was used to meeting. Leahy said the man had one last question for them: "So when these OpenAI and Microsoft guys say AI won't affect employment, that's not true?"
"No sir."
"Bastards."
Apparently, this MP's experience is far from unique. "It's really important to understand," Leahy said, "that almost all technical capacity or like advice that flows into these things comes through big tech lobbying firms." Whenever politicians "have a technical question," he said: They basically only have academia — which is extremely biased and also now funded by big tech — and big tech itself to give them advice. When the White House wants advice on what to do with AI, they ask Sam Altman." (pp. 348-349)
On autonomous weapons:
“But what if your security forces were never at risk of being overcome by conscience? Lethal autonomous weapons, or more appropriately, killer robots, promise exactly that. In 2025, the Chinese Army’s YouTube channel published a video of a robotic dog armed with an assault rifle firing at targets and waving to the camera. The title of the 14-second clip translated to “Continuous Fire! A Closer look at the ‘Hunting’ Moment of China’s Robot Dog!” Similar robot dogs brandishing megaphones were used to bark orders enforcing COVID-19 lockdowns.” (p. 205)
AI advances have also dramatically lowered the barrier to making weapons autonomous. One engineer built a voice-controlled smart gun in his garage. He tells it “we’re under attack from the front left and front right, respond accordingly.” It aims left, fires; aims right, fires (blanks). Then a cheerful but determined AI voice replies “if you need any further assistance, just let me know!” The robot gun’s voice commands use ChatGPT, prompting OpenAI to cut off the engineer’s access.
A week later, he posted a video showing off a machine that he “thinks is going to be super transformative for the construction industry.” The rifle has been replaced with a nailgun pointed at a wall of balloons of varying colors. The engineer says “ChatGPT, begin color tracking on the color black.” The robotic bailing methodically shoots the black balloons, (mostly) avoiding the others. The engineer names each color until there are no balloons left. “If you need any further assistance,” the bot chimes, “just let me know!” (pp. 205-206)
Apollo Research CEO Marius Hobbhahn on AI “permission creep”:
“Companies start conservative — no Slack access, no internet— but convenience wins out. “At some point you’re just like, “this is so tedious… We’ll just give it access to all of our Slack.” Then internet access. Then training data. It eventually gets a far better understanding of what’s happening at the company than any individual employee. If the AI is misaligned, he asked, “how would you catch that the Ai Tells one story to employee A and then one story to employee B in a totally different building?” More than that, what if the AI wasn’t just snooping Slack? What if it was also running the company?” (pp. 237-238)
“In a podcast with libertarian economist Tyler Cowen, Altman said “shame on me if OpenAI is not the first big company run by an AI CEO, right?” Cowen was incredulous. “Just parts of it. Not the whole thing.” “No,” Altman corrected him, “the whole thing.”” (p. 239)
On Moltbook:
“In the project’s early days, creator Peter Steinberger said he sent a voice message to his Molt and was shocked to receive a reply ten seconds later. He hadn’t set up support for audio, so he demanded an explanation. The agent said it inspected the file type, converted it, found his OpenAI API key, used it to transcribe the file, then responded.” (p.247)
“Molts are not about to take over the world, and Moltbook, even before its vulnerabilities were widely exploded, is not a window into the true nature of AGI. It is, however, a clear demonstration that many of us are willing to hand our digital car keys to untested and untrustworthy AI agents.” (p. 248)
On compute verification and an international deal:
“Both Yoshua Bengio and Geoffrey Hinton told me they left meetings with Chinese officials feeling like they really understood the technology and loss-of-control risks better than their US counterparts. During this trip, Hinton was invited to speak with the Shanghai party secretary, who is also on the Politburo— China’s top 24 political officials. Hinton told me the secretary “really did understand.” The Chinese leadership comprehends AI and its dangers “much better because many of them are engineers.”
China’s industry, however, is a different story. A former senior natsec official told me that some US companies “take loss of control seriously,” whereas Chinese companies “don’t care about it pretty much at all. Indeed, as of March 2026, none of the leading Chinese AI firms have even published a safety framework.
The former official said that while both the US administration and the Chinese state are complex, heterogenous entities, if you treat them as single actors, the US government is “not concerned about loss of control. The Chinese government, meanwhile, is “not so much concerned about loss of control from AGI, per se. It just had a reflexive posture of wanting not to lose control of anything else, and so they’re just as concerned about losing control to the companies as they are to the AIs.
[...]
Whether Beijing's caution stems from genuine concern about rogue AI or self-preservation, the practical effect may be similar. The CCP has already shown a willingness to pay heavily for control: Independent estimates indicate its crackdown knocked a trillion dollars off the value of its tech sector and its zero-COVID policy cost $700bn in GDP in 2022 alone. And for its success, DeepSeek was reportedly rewarded with government-screened investor meetings and key employees being asked to surrender passports.” (pp. 277-278)
On regulation and liability:
"For decades, American antitrust policy focused narrowly on consumer welfare, measured through prices, without considering the many other ways monopoly power can be exerted — on suppliers, localities, competitors, governments. This paradigm allowed particularly aggressive consolidation within tech, especially for platforms that didn't charge users. The multi-trillion-dollar generative AI boom has been financed by the deepest capital markets the world has ever seen, themselves enabled by ludicrously profitable market dominance.
Google's monopoly on search was so lucrative that the company could afford to let DeepMind conduct open-ended research and win games for years without expecting it to turn a profit. When OpenAI the nonprofit couldn't keep up with Google, Microsoft's billion dollar investment saved the lab in 2019, while attaching terms that gave the tech giant enormous leverage over OpenAI for years. And Anthropic has taken billions from each of Amazon and Google, which may help explain why Dario Amodei warns that we're as close as a year of two from superintelligence while arguing regulations should focus on transparency for now." (p. 299)
"A different, arguably more elegant, approach would incentivize the industry to figure out exactly how risky its products are. "Make AI companies pay for the harm they cause," writes Touro Law professor Gabe Weil. Elsewhere, the law applies "strict liability" to actions deemed abnormally dangerous." As he put it, if "your neighbor's pet tiger bites off your hand," you would be entitled to compensation for your injuries, "even if they took reasonable precautions to keep others safe.
AI should be treated the same way, Weil argues, "because its consequences can also be abnormally dangerous." Strict liability would force companies to internalize risks that are currently falling on all of us. However, this no longer works if the damage gets big enough — a disaster killing millions would bankrupt any company long before victims and families get compensated. Current negligence law does apply to AI companies, but it only requires that they take "reasonable care"— a narrow standard that asks whether they did the standard safety checks, not whether those checks were actually sufficient to mitigate the risks. As Weil puts it, "The best practices may simply not be good enough. For this, he proposes punitive damages: "The human or corporation responsible should be held liable not just for the harm they caused, but for the uninsurable risks they generated."" (pp. 311-312)
On mass advocacy and the "ban superintelligence until mass public support" letter:
"Believing AI poses an existential risk is a sufficient but not necessary condition to want to stop the Obsoleting Project. For too long, much of AI safety has tried to convince people of the most extreme and speculative risks, while strategically offering moderate, incremental prescriptions woefully out of step with the diagnosis. [Machine Learning Professor and founder of Evitable David] Krueger thought AI safety's tendency to think on the margin — what can I individually do to make things go better — is "fine in a way", but when it becomes so many people's approach, "it gives the impression to the rest of the world that what we're doing is adequate," even though so many would acknowledge that "of course the safety work we've done is not enough."
Instead of prematurely compromising, more people could have advocated for the obvious solution, winning allies who didn't need another reason to join the fight. Mixed strategies are smart — not everyone should play the radical. But the default position for much of AI safety was to preemptively accept AI's inevitability, and work backwards from there, while strategically avoiding public declarations that, obviously we should stop if we could." (pp. 341-342)
On the framing device for the book, “The Obsoleting Machine”:
“Aspirations, of course, do not make reality, and AI’s most fervent boosters overhype the technology’s present and near-future capabilities. Meanwhile, its greatest skeptics often focus on semantic debates over what constitutes true “thinking” or “understanding.” To avoid both of these traps, we should focus on two questions: What can AI actually do? And how quickly is that changing?
The Obsoleting Machine does not yet exist. But today’s most capable general-purpose AI models—Anthropic’s Claude, OpenAI’s ChatGPT, Google’s Gemini— aspire to it. These are known as frontier models: ones at the cutting edge. They work somewhat like a consultant you can only hire once. Each can read up on the project, including what was done by previous consultants, but is only able to perform a finite amount of work before being replaced by a fresh one. They’re far cheaper than classic consultants and occasionally very useful, but not something that’s about to replace your whole workforce.
The actual Obsoleting Machine would be able to join the team indefinitely, execute the job end-to-end, and make most knowledge work superfluous—leaving livelihoods safe for the extent a human touch is preferred, a physical presence is required, or the jobs themselves are legally protected. The real-world difference between automating a process 99.9 percent of the way and 100 percent can be vast, as someone who has “nearly finished” his book about a dozen times, I can attest that last mile problems are everywhere.” (pp. 12-13)
This is probably my favorite quote in the book. Part of it is in the book, but I wanted to quote the section of the 2016 New Yorker article in full.
Altman felt that OpenAI’s mission was to babysit its wunderkind until it was ready to be adopted by the world. He’d been reading James Madison’s notes on the Constitutional Convention for guidance in managing the transition. “We’re planning a way to allow wide swaths of the world to elect representatives to a new governance board,” he said. “Because if I weren’t in on this I’d be, like, Why do these fuckers get to decide what happens to me?”
I’ve just finished reading Garrison Lovely’s recently released book, “Obsolete.” I wish I had time for a longer review, but I’ll give a brief one and highlight some particularly interesting sections below. I think it’s a great introduction to non-technical readers and even as a pretty dedicated reader of AI policy and news I learned a good bit. It’s deeply, deeply researched and has a good pulse on the alignment community, and Lovely is careful not to understate or overstate his case. The core argument is that AI development should be analyzed through the lens of mass or universal labor displacement, and that a mass movement to shape, pause or freeze development is much more possible, much more positive, and capable of greater breadth of salience and support than is often argued in the alignment community.
He makes an, in my opinion, decisive case that “competitive pressures” by themselves don’t dictate that AI development is fated to happen, and certainly not that it’s fated to happen at the pace the frontier labs are setting. But I wish he’d been a bit more careful about the limitations and subtleties of the plan he’s advocating for; I share the concern raised by folks like the AIFP that it’s hard to pause indefinitely at any given point for very long and that the cost to train a model at any fixed intelligence level, once first reached, is falling precipitously, and that many well-capitalized actors currently want to and are able to speed up development precipitously. I think the book somewhat elides the difficulties we’d still have even if this number were close to 0.
It was likely impossible to train GPT-6 in 2016, but it may be extraordinarily cheap to train a GPT-6-level model in 2036, and even just extrapolating the current rate of progress naively, it might be extraordinarily cheap to train a ~GPT-8 level model in 2036. That seems like a world where many irresponsible agents are deploying very powerful technology even if we have verification mechanisms preventing further uncoordinated capabilities training.[1]
But one thing I do find refreshing is Lovely's relentless focus on partial solutions. They're interspersed throughout the book but come into full focus in Chapters 17 and 18. There's a model of the world that relies on large paradigmatic shifts, where progress on a task is more or less bottlenecked on "true understanding." I think this is sometimes a good instinct in a field as anti-inductive as alignment, where sometimes you really are barking up a useless or even actively counterproductive tree and no amount of just trying harder will fix that. but I think the ugly underside of it is the current situation where much of the field was aggressively averse to policy or advocacy and is now attempting to rapidly build governance infrastructure now that it seems people might actually listen to you. Sometimes you need a list of your most presentable ideas, even if your best ideas have known flaws. And again, not all partial solutions are technical; a pause buys valuable time across many worldviews.
Overall, I still think this book is a well-worthwhile read even for many technical researchers, and even for many people who will fundamentally disagree with the book’s premise. Being better informed about your field holistically is very valuable. Many conceptions of research taste say something like that it’s the art of asking increasingly good questions within your paradigm. But another important part, if making the world a better place is a meaningful consideration in your ethics, is analyzing whether your paradigm would be good to advance in the first place. Do we want to get what we want, or the subset of what we want that we can measure? I think this book is a compelling anchor for discussions of “what do we want out of the development of artificial intelligence?”
On Leading the Future and other lobbying efforts
On autonomous weapons:
On “permission creep”[2] and delegating to AIs:
Apollo Research CEO Marius Hobbhahn on AI “permission creep”:
On Moltbook:
On compute verification and an international deal:
On regulation and liability:
On mass advocacy and the "ban superintelligence until mass public support" letter:
On the framing device for the book, “The Obsoleting Machine”:
This is probably my favorite quote in the book. Part of it is in the book, but I wanted to quote the section of the 2016 New Yorker article in full.
I can’t divine the future, but here’s a preview of kinds of events that could become much more common, more severe, and/or more autonomous.
Gradual disempowerment is my preferred term/framework for this kind of thing.