Garrison Lovely’s Obsolete: Why He Reframes the AGI Race as a Project to Make Humans Obsolete

1. Guest Background

Garrison Lovely is introduced in the episode as a Brooklyn-based freelance journalist and the author of Obsolete: The AI Industry's Trillion-Dollar Race to Replace Us and How to Stop It. The interview is therefore not a generic conversation about AI anxiety. It is an extended examination of Lovely’s book-length argument that frontier AI companies are engaged in a race to replace human labor, and that this race should be confronted through politics, labor power, democratic authorization, whistleblower protection, and safety governance.

Nathan Labenz emphasizes at the start that Lovely writes from a left-leaning perspective for a largely left-leaning audience, but does not treat AI risk as mere industry hype. Labenz says Lovely takes the companies’ stated goal seriously: building systems better than humans at cognitive work. That framing matters because it positions Lovely as neither a conventional Silicon Valley booster nor a casual skeptic. He is arguing with people who accept AI capability progress, but who may disagree about what follows from it.

Lovely is also connected in the episode to Irreplaceable, an organization seeking “a say, a stake and a slowdown” in the political arena. The interview notes that he gives advice to potential whistleblowers and is donating book royalties to that organization. The guest background, then, is not only writerly. Lovely appears here as a journalist, author, and movement participant trying to translate AI labor displacement and frontier safety concerns into public power.

2. What the Episode Covers

The episode analyzes Lovely’s central reframing of frontier AI. Labenz summarizes the book’s policy target as preserving deep learning’s upside in domains such as medicine and materials science while preventing companies from racing into recursive self-improvement or systems that render humans obsolete. Lovely’s key phrase is the “obsoleting project”: his name for an AGI industry that is not simply building better software, but trying to create general substitutes for human labor.

The conversation begins with Lovely’s own use of AI. He says he does not use it for writing, partly because AI writing quickly became stigmatized in journalism. Yet he uses AI extensively for transcription, NotebookLM document analysis, interview organization, feedback, fact-checking, research, and tools for his book website. This matters editorially because Lovely is not criticizing AI from ignorance or abstinence. He says NotebookLM can query large document sets and organize interview quotes, and he acknowledges that AI can catch real mistakes and omissions.

At the same time, every productivity claim comes with a constraint. NotebookLM can hallucinate and miss material. AI feedback can be foolish; AI fact-checking can incorrectly challenge true claims. Lovely’s most vivid self-critique is his near “Claude psychosis” period while writing the book: rapid full-manuscript feedback was useful, but it also pulled him into fixing things that were not broken and seeking reassurance from AI when exhausted. His practical conclusion is not that AI is useless. It is that the more a knowledge worker relies on it, the more they need strong human judgment about the task, the model’s competence, and when the tool is wasting time.

The discussion then turns to why the American left has often underestimated AI risk. Lovely gives several explanations: influential left-aligned AI skeptics and the “stochastic parrots” frame shaped deference; crypto, NFTs, the metaverse, and social media hype created a reflex that tech people were again overpromising; and AI bubble narratives are psychologically comforting because they imply no need to organize. Lovely translates the risk into left political language: if capital can become labor without workers, labor’s share can be driven toward zero.

Labenz pushes the strongest automation counterargument. He says AI gives him major daily value, does tedious work, and has counterfactually substituted for labor he might otherwise have hired. He also invokes agricultural mechanization: automation once allowed a small share of the population to feed society while others did different work. Lovely accepts that labor automation historically raised living standards, but insists that automating some labor and trying to automate all labor are different projects. A universal labor-replacing machine would affect income, hiring, bargaining power, identity, and political power simultaneously.

Lovely’s positive vision is therefore not a return to a pre-technological society. He describes a “third New Deal”: Medicare for All, locally administered jobs guarantees, welfare-state protections, public technology investment, and a “cures for all” agenda. He imagines government treating diseases more like COVID, ranking by tractability and disease burden, using advanced market commitments, vaccine trials, human challenge trials, and AI where useful. For drug discovery, he favors prize systems that specify socially desired medicines, reward success, and then produce at generic costs, rather than allowing monopoly patents to steer AI toward the most profitable drugs.

The second half of the interview focuses on leverage. Lovely argues that AI researchers and engineers may be near their peak power because frontier labs still need them to train models; if recursive self-improvement works, that leverage vanishes. He therefore advises insiders not to see resignation and public whistleblowing as the only option. Staying to organize safety-conscious coworkers, forming unions, bargaining collectively, withholding labor, striking, or slowing work could make safety commitments binding. His example is OpenAI’s superalignment team, which was promised 20 percent of compute but, according to Lovely, received far less in practice. For him, that shows what happens when safety commitments lack an organized counterforce.

3. Core Views: Reasoning, Examples, and Limits

Lovely’s first core view is that the key question is not whether AI is useful. He accepts continued deep-learning development and sees specialized systems as potentially valuable for disease, materials science, and other public goals. His target is the AGI “obsoleting project.” The reasoning is powerful because it shifts the debate from whether to embrace or reject AI to who gets to authorize a universal labor substitute. Its limitation is that the boundary between tool AI and general labor replacement is not always crisp; this is why Lovely later favors broad freeze definitions that can be calibrated over time.

His second view is that frontier AI companies cannot be understood only as profit machines. Lovely thinks early AGI founders were mission-driven before capital flooded in, and that leaders remain motivated by idealism, inevitability, the desire to be historical figures, messiah complex, or “better us than them.” This makes governance harder. If executives were merely maximizing profit, conventional regulatory and market pressure might be easier to model. If they believe they are present at the decisive event in human history, they may interpret opposition as shortsighted rather than legitimate.

A third view is that technical alignment is not a sufficient solution. Lovely criticizes the history of AI safety for focusing too narrowly on making powerful AI do what users want. He proposes an “alignment polycrisis” that includes technical, normative, economic, and geopolitical alignment. RLHF is his concrete example: developed by Paul Christiano and others at OpenAI for safety reasons, it also made LLMs conversational and useful, helping enable ChatGPT and the broader wave that followed. The limitation in this argument is not that every alignment advance is bad; it is that even successful technical alignment cannot by itself decide ownership, externalities, geopolitical use, or whether the system should be built.

A fourth view is that market discipline cannot price the relevant harms. Lovely concedes that markets give companies some incentive to stop AI from obviously misbehaving, but he doubts whether that incentive is enough to solve the problem rather than to make a saleable product. He points to current systems that can be lazy, hallucinate, and sycophantically fabricate while still selling well. More severely, he argues that if an AI company caused a disaster killing 10 million people, it would go bankrupt before compensating society; insurers, he says, refuse coverage because the risks are too large and correlated. Liability and procurement standards may improve some behavior, but they cannot compensate for extinction or answer the democratic-consent problem.

A fifth view is that “freeze the frontier” should become operational, not merely rhetorical. Lovely proposes no training runs as large as or larger than the previous frontier scale, no reinforcement learning from verifiable rewards aimed at capability gains, and no recursive self-improvement. He favors overinclusive RSI definitions because the downside of error is asymmetric, while acknowledging that forcing the industry back to human speed would have significant economic consequences requiring mitigation. This is not a ban on all AI use: ordinary inference for customers is not, in his view, usually frontier advancement, while larger pretraining runs and many RL experiments are.

A sixth view is that international AI governance is primarily a problem of credible verification and political will, not only technical design. Lovely suggests embedded auditors with employee-level access inside companies, plus international auditors, agencies, cryptographic tools, chip inventories, and monitoring of data-center or chip network activity without exposing model weights or state secrets. His Cold War arms-control analogy is meant to show that rivals do not have to trust each other’s intentions if they can create verifiable arrangements.

His seventh view is that the “but China” objection is not decisive. Lovely argues that if the United States slowed or stopped, China would also slow for a time because it has used a fast-follow strategy, usually three to nine months behind the U.S. frontier. He also argues that credible U.S. regulation could make China more willing to negotiate, and that both countries have self-interested reasons to prevent rogue hacker AIs and AI systems that help anyone make bioweapons. He does not claim this is risk-free: he acknowledges Chinese developers could eventually pass a stopped U.S. frontier. His claim is that catching up is easier than advancing the frontier, and that the Chinese Communist Party is unlikely to willingly permit uncontrolled recursive self-improvement.

4. Learning and Application

For knowledge workers, the most immediate lesson is not simply whether to use AI. It is how to set boundaries around use. Lovely’s own practice suggests AI can be valuable for transcription, search, document organization, interview synthesis, first-pass fact-checking, and small prototypes. But the closer the task gets to judgment, explanation, factual responsibility, or prose ownership, the more human verification matters. A useful rule is to treat AI as an assistant that expands retrieval and organization, not as an editor that replaces accountability. If the model pulls you into fixing non-problems, using praise as emotional support, or feeling busy without progress, the tool is no longer saving time.

For policy thinkers, Lovely’s framework encourages decomposing the question “Is AI automation good?” into more precise questions. Is the system replacing dangerous or unwanted work, or building a universal labor substitute? Are benefits being locked behind patents and monopoly rents, or converted into public goods through prizes, public investment, and generic-cost production? Is a medical AI project an AlphaFold-like specialized system, or is a speculative superintelligence story being used to justify ungoverned AGI? That separation helps avoid both blanket anti-AI policy and using medical upside as a blank check for frontier racing.

For frontier AI employees, especially researchers, Lovely’s application is to convert private conscience into collective leverage. Resignation and public criticism may sometimes matter, but if the company still depends on your labor, organizing coworkers, forming safety-focused unions, and demanding enforceable safety terms may have more institutional force than individual dissent. The condition is that workers need legal advice, an understanding of retaliation risk, and a realistic view of reputational protection. Lovely himself says he is not the relevant expert and offers to connect people with those who know more.

For enterprise buyers and institutional users, the episode suggests that procurement should not focus only on price, model strength, and ease of integration. Buyers can ask suppliers for evidence about safety standards, incident responsibility, logging, auditability, insurance gaps, red-team results, and governance. Lovely does not think corporate purchasing can solve the whole problem, but procurement can become one source of external pressure, especially when government regulation lags.

For movement builders, the episode’s practical lesson is to organize through a big tent. Some people arrive through jobs, others through environmental concerns, data centers, surveillance, wealth concentration, or existential risk. Lovely draws from the climate movement to argue that immediate harms and long-range catastrophic risks do not need to compete. A clear demand such as stopping the race to replace humans or freezing the frontier can coordinate different concerns. But he also warns that data-center moratoria alone are not sufficient; without constraints on model developers, chipmakers, export controls, and training activity, local opposition will not substantially change capability progress.

For whistleblowers and internal sources, Lovely’s advice is specific: contact the AI Whistleblower Initiative or trusted journalists before taking unnecessary risks. He says he treats inbound contact as off the record by default, but also stresses that he is not a lawyer. California SB 53 and other protections require expert guidance. The application is not reckless disclosure; it is converting early inside knowledge of dangerous behavior into public knowledge through legal, journalistic, and professional safeguards.

The broader lesson is that this episode offers not a finished blueprint but a governance lens. Universal labor substitutes deserve a higher authorization threshold than ordinary products. Technical alignment cannot substitute for economic, political, and international governance. Public benefit requires institutional design, not only stronger models. And if companies, CEOs, employees, and governments are locked in a prisoner’s dilemma, public organizing is not a decorative add-on; for Lovely, it is the mechanism that might put deep learning’s real potential on a better path.

Source

• Original episode: Obsolete or Irreplaceable? Garrison Lovely on Stopping the Race to Replace Human Labor