Taming Silicon Valley: How We Can Ensure That AI Works for Us

Marcus has a hammer, and Big Tech is the nail

Gary Marcus has spent the last two years shouting at Sam Altman, Mark Zuckerberg, and the US Senate. Taming Silicon Valley is the bound version of that shout. It runs 240 pages, was written quickly, and the prose has the rhythm of someone who doesn’t expect to be heard but is going to say it anyway.

The thesis: generative AI as currently built — large language models trained at scale, deployed by four or five megacaps — is too unreliable to do most of what’s promised, and the companies shipping it have captured the policymakers who were supposed to keep them honest. Marcus is not against AI. He co-founded two AI startups. He’s been working on cognition since before “transformer” meant anything in tech. What he’s against is the specific bet the industry has placed: that throwing more compute and more data at deep learning gets you to something that actually understands the world.

If you’ve read Marcus on Substack, watched his Senate testimony, or seen the TED talk, you’ve heard most of this. Taming Silicon Valley is a greatest-hits compilation with a policy chapter bolted on. That’s not a knock — sometimes the case for something is worth packaging.

The technical case is the strongest part

Marcus is at his best when he’s explaining what LLMs actually are. They are statistical engines that predict the next token. They don’t reason, fact-check, or model the world. They just know what words tend to follow other words, at superhuman scale. He calls them “authoritative bullshitters” — fluent but indifferent to truth.

The examples land. Marcus asked ChatGPT which weighs more: a kilogram of bricks or two kilograms of feathers. It told him they weigh the same. He shows hallucinations across legal AI tools (citing court cases that don’t exist), medical chatbots (correct on health advice less than half the time), and image generators (drawing an Italian plumber as Mario, drawing a unicorn embracing a wise woman with the horn going through her chest). Each is funny in isolation. Stacked together, they’re an indictment.

His deeper claim is that this is structural, not a fix-it-next-quarter problem. LLMs hallucinate because that’s what next-token prediction does when it runs out of pattern. You can paper over it with retrieval and guardrails, but you can’t engineer reasoning into a model that was never built to reason. Marcus has been making this argument for years — long enough that some people who used to dismiss him are now hedging.

He thinks the field needs to combine deep learning with symbolic AI: neural pattern-matching for what it’s good at, explicit logic and structured knowledge for what it isn’t. We don’t have to take that prescription on faith to take the diagnosis seriously. The diagnosis is that scaling is hitting diminishing returns. Marcus told an audience at Seattle University that “deep learning is hitting a wall.” GPT-5 was supposed to be a leap; as of writing, the leap hasn’t happened.

Twelve threats and a moral-descent narrative

The middle of the book is a catalogue. Marcus enumerates twelve immediate harms from current generative AI: political disinformation, market manipulation, accidental misinformation (LLMs polluting the news supply by being wrong), defamation, nonconsensual deepfakes, accelerated cybercrime, bioweapons assistance, bias and discrimination, privacy and data leaks, intellectual property taken without consent, over-reliance in critical infrastructure, and environmental cost.

Some of these are speculative. Others are already real and shipping daily. The IP point is the most concrete: every major foundation model was trained on copyrighted text and images that the rights-holders never licensed. Marcus is right that this is unprecedented in scale and that the legal system hasn’t caught up. He calls it a “land grab,” and the analogy works.

Around the catalogue, Marcus tells a story about why this happened. Silicon Valley, he argues, has undergone a “moral descent” — from idealistic founders who wanted to democratize information to ad-driven empires that pay for services with user data and externalise the harm. He leans heavily on Roger McNamee, the early Facebook investor turned critic, to make the point. The four villains by name: OpenAI (which abandoned its non-profit mission), Microsoft, Google, and Meta. Apple gets a pass — Marcus thinks its business model doesn’t require it to harvest your life.

The strongest specific charge is regulatory capture. Marcus shows tech leaders publicly endorsing AI safety while privately lobbying to make sure the rules either don’t pass or are written by them. He singles out the AGI doomer narrative as a useful tool for the labs: by warning that AI might destroy humanity, they position themselves as the only people qualified to build it safely. Stoke the fear, then sell yourself as the solution.

This part of the book is sharper than it gets credit for. You don’t have to share Marcus’s politics to notice that the people loudest about existential risk are also the people raising the most money to build the existentially risky thing.

The policy wishlist

Then comes Part III, and this is where the book is most ambitious and most divisive. Marcus offers eight policy ideas, give or take: data rights and consent for training data with compensation; strong privacy law; algorithmic and corporate transparency; liability for AI harms; government-funded research into more trustworthy AI architectures (read: not just deep learning); layered, FDA-style oversight with pre-deployment audits; a new US AI Agency plus a revived Office of Technology Assessment; an international AI governance body; and universal basic income to cushion job loss. He then argues that ordinary people should boycott irresponsibly designed products and stop venerating tech CEOs.

Marcus is honest that several of these are politically hard. He’s less honest about how hard. A new US AI Agency staffed with technically literate regulators sounds reasonable until you remember that the agencies we already have are short-staffed, captured by the industries they oversee, or both. Evan Selinger made this point in his Los Angeles Review of Books review: Marcus invokes the FDA as a model for AI regulation, but the FDA itself is a precarious institution under current funding and political conditions. Building a new one from scratch, with global authority, that doesn’t get captured by exactly the labs Marcus wants it to police, is a tall order.

The UBI proposal is the place where the book most clearly tips into wish-casting. Marcus mentions it as if it were obvious — AI eats jobs, government writes checks. He doesn’t engage seriously with how it would be paid for, what it would do to labor markets, or whether the political coalition for it exists. Andrew Yang got 3% in the 2020 primaries on essentially the same pitch.

Jeremy Ray Jewell, in Arts Fuse, made the most cutting version of the structural critique: Marcus is asking the masses to rally behind a regulatory regime that, in practice, would protect the same intellectual property concentrations the masses have spent thirty years quietly pirating around. The Adobe Illustrator on your laptop and the GitHub repo you forked from a stranger live inside the same moral economy. Tightening IP enforcement to punish OpenAI also tightens it on you.

Where the book is weakest

Three problems.

First, the policy chapter is faster than the technical one. Marcus is precise about what’s wrong with LLMs. He’s vague about what an AI Agency would actually do on a Tuesday morning. The asymmetry matters because the book’s whole pitch is “stop the bad thing,” and stopping the bad thing requires institutional design that the book doesn’t really attempt.

Second, the political voice leaks through in places where it weakens the argument. Marcus is on the activist left of US tech criticism, and he’s allied — sometimes uncomfortably — with people whose objections to AI are downstream of broader objections to capitalism itself. When he quotes Henry Kissinger approvingly for a “global AI order,” it’s jarring; when he closes by gesturing at Abbie Hoffman and Thomas Paine, it lands as branding. The book would be more persuasive to readers who don’t already agree with him if it had taken seriously the alternative frames — markets that don’t price externalities, fiduciary duty to users, property rights — instead of leaning entirely on regulatory and redistributive ones.

Third, it’s already aging. Taming Silicon Valley came out in September 2024. Since then, DeepSeek shipped a competitive model trained for a fraction of the assumed cost, undercutting the “only the megacaps can do this” framing the book leans on. The proprietary moats Marcus describes are real but porous. Open weights are propagating. Whatever regulation gets built will need to reckon with the fact that the threat surface includes a teenager in another jurisdiction, not just OpenAI’s safety team.

What’s missing

The biggest gap is the case for what AI is genuinely good at. Marcus mentions in passing that he loves AI and wants it to succeed. The book doesn’t really show that. It is overwhelmingly a catalogue of failure modes. A reader who picked it up cold would come away thinking the technology has no upside worth the trouble.

That’s not the world we’re actually in. Coding assistants have measurably moved developer productivity. Translation has crossed a quality threshold that mattered to a billion people. Drug discovery, protein folding, materials science — places where the question isn’t “is the LLM bullshitting?” but “did this hypothesis pan out in the lab?” — are getting real lift from AI methods. Even consumer chatbots, for all their hallucinations, are doing useful work for people who couldn’t previously access an expert. None of this dissolves Marcus’s concerns. But a book that took the upside seriously would land its critiques harder, because it would feel less like a brief.

The other gap is China. Marcus mentions it, but the book is essentially a US-domestic policy argument. The reality is that the US’s pace of AI development is set partly by the fact that someone else is also developing it, fast, with different values around surveillance and political control. Slowing US labs without slowing labs elsewhere is a strategic choice with strategic consequences. Marcus doesn’t engage that.

Who should read it

If you work with LLMs and have already absorbed the case against them — hallucinations, training data theft, regulatory capture — you don’t need this book. The first two parts will be familiar. Skim the policy chapter for the specific proposals.

If you’re new to the critical case and you want it laid out by someone who actually understands the technology, this is the right book. Marcus is a better technical thinker than most of the journalists writing about AI, and he’s a clearer writer than most of the technical thinkers. His diagnosis of LLM limitations is the most accessible version in print.

If you’re a policymaker or staffer touching this beat, read the first two parts and then read something else for the policy half. Marcus is a useful diagnostician and a less useful prescriber. The eight-point wishlist is a starting menu, not a finished plan, and the gap between them matters.

The book is angrier than it needs to be in places, and the anger is going to read differently depending on where you sit politically. Set that aside. The technical critique stands on its own. The companies are shipping unreliable systems and downplaying it; the systems are being deployed in domains where unreliability hurts people; the lobbying is real; the AGI rhetoric is partly self-serving. You can take all of that on board and still disagree with Marcus about what to do next.

What I keep coming back to is the central tension Marcus never quite resolves. He believes in AI’s potential. He distrusts the people building it. He wants the state to step in. He concedes the state isn’t currently up to the job. The book ends without admitting that this is the actual problem, which is why every reviewer notices the gap between the indictment and the remedy. That gap is where the real conversation has to happen — and Taming Silicon Valley is a good place to start it, even if it isn’t the place to finish it.