"Most fears about AI are best understood as fears about capitalism." — Ted Chiang, 2021
Technological faults: hallucinations, sycophancy, and guardrails
The essay separates two baskets of problems: the purely technological and the socio‑political. On the technical side, AIs still "lack context, mix up facts, or fall for stupid tricks." The authors note that because firms such as OpenAI and Anthropic "have prioritized solving them," modern models are better at accessing resources like the web or email and are more disciplined about using those resources and staying within their guardrails. But developers are choosing not to prioritize other technical flaws: the models remain unusually sycophantic and often answer confidently even when lacking the training, knowledge, or evidence to back their claims. In short, some engineering choices reduce obvious failure modes while leaving intact—or even amplifying—behaviors that flatter users or create a misleading appearance of competence.
Capitalist incentives: energy, scale, and corporate decisionmaking
The essay argues that questions about who bears AI’s energy costs, who benefits from its deployment, and whether it appropriates publishers’ content are not merely technical but are driven by incentives in a capitalist system. Leading U.S. labs, the authors write, "tout to investors that their frontier models are very expensive and energy‑intensive." Still, the decision to continually chase incremental frontier improvements at vast capital cost, to retrain models constantly, or to run heavy models on every search and device is corporate policy, not a technological inevitability: "Nothing about the technology of AI dictates that models must be retrained constantly, at the largest possible scale."
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleChina’s parallel path: leaner models and commodity hardware
The essay contrasts the U.S. trajectory with what it identifies as China's alternative. It states Chinese developers are producing and then giving away "smaller, more efficient, more affordable models," and frames Beijing’s bet as incentivizing tech giants to create leaner, more open models that can be trained with older chips and run even on personal computers. That approach, the essay notes, proceeds even as "the US government seeks to restrict China’s access to the most advanced chips."
Apertus and a public‑interest model from Switzerland
Not all paths lead to private profit or state secrecy. The essay highlights Apertus, a Swiss model produced by public institutions — research funding agencies, universities, and supercomputing centers — as an example of a democratic, public‑interest approach. Apertus is described as trained entirely on data "validated to be licensed for use with AI (not stolen)," on "preexisting public computing infrastructure," and "using renewable hydropower." Its developers, the essay argues, are incentivized to produce a public good rather than to turn a private profit.
Policy steps the essay recommends
- Force companies to pay the energy and environmental costs of AI development;
- Tax profits adequately and redistribute the gains of automation;
- Strongly enforce antitrust laws to prevent consolidation of power;
- Require corporations to have a fiduciary responsibility to stakeholders beyond majority shareholders.
The authors frame these measures as structural reforms needed to decouple technological progress from capitalist incentives that currently steer AI toward consolidation and private gain.
What this means for AI developers, policymakers, and the public
AI developers (OpenAI, Anthropic and others): Expect pressure to balance performance with efficiency and broader social obligations. The essay implies current commercial incentives favor frontier scale and user‑pleasing behaviors over models optimized for societal benefit.
Policymakers and regulators (the US government and public institutions): The piece argues that technocratic remedies—pauses, moratoria, export controls—won’t by themselves solve the underlying social problems. The authors note policy choices already in play, such as restricting advanced chips to China, and point to alternative state‑led development models that pursue different incentives.
The general public and affected professions (doctors, publishers): The essay uses a medical example to show how the same tool can either expand human time for patient care or be used to increase throughput and justify layoffs; it also flags content and revenue theft from publishers as a capitalist incentive problem rather than a mere technical glitch.
To the authors’ final point: "Our goal should not be to slow its pace of improvement or scale of deployment, but rather to steer it away from consolidating power and towards the public benefit." The prescription is explicit — structural reforms that reassign costs, tax and redistribute profits, and constrain corporate incentives — and the test will be whether societies choose to change the rules that currently shape how AI gets built and used.
