Impromptu: Amplifying Our Humanity Through AI

A book co-written with GPT-4, by one of GPT-4’s funders

The premise is the gimmick: a book about GPT-4, written with GPT-4, in early 2023, by an OpenAI early funder who also sits on Microsoft’s board. Hoffman calls it a “travelog” — a snapshot, not a treatise — and that framing is the most honest thing about it. The book has a known half-life. He says so himself, more than once.

What you get is ten chapters where Hoffman prompts GPT-4 about education, creativity, justice, journalism, social media, work, his own workflow, hallucinations, “public intellectuals,” and a closing essay on technology and humanity. He pastes the prompts and the outputs verbatim. Sometimes he edits for length. Often he asks GPT-4 to “write in a less wooden style than usual.” (It usually doesn’t.) The book is short, cheap, and was written fast — released within months of Hoffman getting GPT-4 access. The Kindle edition was priced at under a dollar, which tells you the goal was reach, not royalties.

The book Hoffman is overtly trying to write is a “what could possibly go right” counterweight to the doomer framings dominating early-2023 AI coverage. The book he’s actually written is closer to a demo reel — a record of how an enthusiastic early adopter prompts a frontier LLM. Those two books overlap, but they aren’t the same. The demo reel is more interesting than the argument.

The AHA framing, and the three principles that survive

Hoffman’s central claim is that GPT-4 is a co-pilot, not a replacement. He calls the moment of seeing this an “AHA!” — Amplifying Human Abilities. The phrase is corny but it’s load-bearing: the whole book is an extended argument that the right way to think about LLMs is as productivity amplifiers paired with continuous human oversight. As Hoffman puts it, “GPT-4 doesn’t replace human labor and human agency, but rather amplifies human abilities and human flourishing.”

The most practically useful pages in the entire book are the three principles he lays out for using GPT-4 in your own work. They are:

1. Treat it like an undergrad research assistant, not an omniscient oracle.
2. Think of yourself as a director, not a carpenter — you’re coaxing performance out of an actor, not swinging a hammer.
3. Just try it. Generate three variations and throw two away. The cost of a bad output is near zero.

This is genuinely good advice, and it has aged better than almost everything else in the book. The director-not-carpenter metaphor in particular is worth keeping. It captures something real about why people who treat LLMs like deterministic tools get worse results than people who treat them like collaborators with a temperament.

He also opens with a memorable demo — the lightbulb joke that became his personal AHA moment. Asked “How many restaurant inspectors does it take to change a lightbulb?”, GPT-3 produced a generic “Only one, but the light bulb has to want to change.” GPT-4 wrote a Seinfeld bit and a Wittgenstein gloss. You see why he was excited. You also see, two and a half years later, that the gap between GPT-3 and GPT-4 was small compared to the gap between GPT-4 and what we have now. The lightbulb joke does double duty: it makes the case Hoffman wants it to make, and it stamps a date on the book.

Where the book earns its keep

Two chapters are doing real work. The first is on education. Hoffman tells the stories of Steven Mintz, a UT Austin professor who immediately required his seminar students to write essays collaboratively with ChatGPT, and Cherie Shields, a public-school English teacher who started using ChatGPT to give students faster feedback on essays. These are real people doing real things. The contrast with NYC public schools’ early ChatGPT ban does the argumentative heavy lifting Hoffman wants — and the Shields example in particular, where she describes ChatGPT’s ability to give detailed feedback on a student essay in seconds, lands. English teachers grade hundreds of essays a week. The “twenty hours of extra work per week” math is concrete in a way most of the book’s claims aren’t.

The other chapter that earns its keep is the closer, “Homo techne.” Here Hoffman argues that humans have always co-evolved with our tools — stone tools, fire, the wheel, writing — and that the cognitive expansions we credit to “human nature” were in fact downstream of tool use. The phrase “Homo techne” itself is a small bit of branding, but the underlying observation is real: there is no clean line between human capability and the technologies humans use, and there never has been. The implication, which Hoffman states clearly, is that worrying about AI making us “less human” gets the causality backwards. Tools are how we became us.

It’s also where Hoffman’s optimism finally has some grit to it. He warns against an AI future where “human beings play less and less of a role in determining their own destinies,” and writes that “as today’s imperfect LLMs improve, requiring less and less from us, we will need to demand more from ourselves.” This is the sharpest sentence in the book. The whole rest of the optimism Hoffman is selling depends on us actually doing this. He doesn’t say much about how.

The “Hallucinations” chapter is also better than I expected. Hoffman’s move is to compare early-2020s panic about LLM errors to mid-2000s panic about Wikipedia errors, and to make a “good enough knowledge” argument: distribution and accessibility matter, error rates trend down, and humans are not exactly trustworthy narrators either. The argument is partial — it skips lightly past the question of whether confidently-wrong outputs at scale are categorically different from a bad encyclopedia entry — but the historical comparison is fair, and the chapter dates well.

Where it embarrasses itself

The “Public Intellectuals” chapter is bad. It is a long sequence of GPT-4-generated imaginary dialogues between paired thinkers — Iris Marion Young and Habermas on the public sphere, Walter Rodney and Fernand Braudel on AI and history, Galileo and Turing on dialogues, and so on. Hoffman calls these “possible interviews,” which is the careful hedge of someone who knows what they are: bland Wikipedia-summary versions of these thinkers, generated to a prompt, indistinguishable from one another in voice, and asserting positions the named authors may or may not have actually held. Anyone with real familiarity with Habermas reads the Habermas character and immediately notices it’s a smart undergrad’s pastiche. Anyone unfamiliar with Habermas now has a counterfeit Habermas in their head.

This is the chapter where Hoffman’s enthusiasm for his tool pulls him into a trap. The whole book is supposed to be about co-pilot use of GPT-4 — human judgment plus machine output. The “Public Intellectuals” chapter is just the machine output, lightly framed. There’s no judgment being applied. It reads as if Hoffman discovered GPT-4 could generate convincing-looking dialogues and decided that was sufficient. It isn’t.

The other chronic problem is that GPT-4’s output is often boring. Hoffman knows this — he repeatedly prompts the model to be “less wooden” or “more lively” — and yet he keeps pasting the boring versions in. Big chunks of the chapters on management consulting, sales, and law are sequences of bulleted lists that GPT-4 produced in response to fairly generic prompts. They read as filler. A more ruthless edit would have cut these by half and let the strong examples breathe.

There’s also a conflict-of-interest issue Hoffman handles unevenly. He discloses, once and in a footnote, that he sits on Microsoft’s board. He discloses that he was an early OpenAI funder. He doesn’t dwell on what this means for the book’s perspective. The chapter on his own work is a list of his investments — Tome, Coda, Adept, Nauto, Cresta, Nuro, Aurora — explained in upbeat terms. It’s not hidden, exactly, but the book is a sales pitch for a category Hoffman is heavily invested in, and the absence of any serious counterargument tells.

The blind spots

What’s missing from Impromptu is the part of the conversation Hoffman is least equipped to lead.

The book’s treatment of labor displacement is thin. The “Transformation of Work” chapter gestures at “training and retraining programs” and “safety nets,” then moves on. Hoffman’s view is that previous technology revolutions created more jobs than they destroyed, and AI will too. This may be true. But the people who lost their livelihoods in those revolutions weren’t comforted by aggregate statistics, and Hoffman’s confidence here isn’t argued, it’s asserted. The Sales chapter contains a striking passage where Hoffman acknowledges AI will likely shrink the sales profession overall — and then immediately pivots to how the surviving salespeople will be more productive. The displaced ones don’t get a chapter.

The training data question gets even less. Karla Ortiz and Glaze get a paragraph in the Creativity chapter. The broader question of whether models trained on the open web without explicit consent are extracting value from creators who never opted in goes essentially unaddressed. This is a fair fight to have, and Hoffman doesn’t have it.

The regulation question is handled with what Hoffman calls “smart precaution” — basically, don’t regulate too fast, let the technology mature, learn from real harms before legislating. This is a defensible position. It’s also the position you would expect from someone with Hoffman’s investments. He doesn’t engage seriously with the case for moving faster.

Most consequentially, there’s almost no engagement with adversarial use. The Journalism chapter discusses disinformation at scale. The Social Media chapter discusses bots and deepfakes. But Hoffman’s solution in both cases is some version of “flood the zone with truth” — make verified content easy to find, build better fact-checking tools — which is more aspiration than mechanism. Anyone who has watched social media platforms try to do this for fifteen years knows how hard it is. Hoffman gestures at the difficulty and moves on.

What’s missing entirely

Two and a half years on, the book reads as a document of its moment, and a few absences are striking.

There is no real engagement with AI safety as a research field. Hoffman mentions OpenAI’s safety work in passing — toxicity classifiers, RLHF — but doesn’t engage with the alignment community’s arguments at any depth. He also doesn’t engage with the China question, or with the specific labor-market distributions across countries, or with the energy and infrastructure costs of training and running these models. The book is written as if AI is happening in a frictionless American context.

The economics are hand-waved. Hoffman cites studies about AI productivity gains in writers and programmers. He doesn’t grapple with what happens when those productivity gains accrue mostly to capital, or with how a workforce reorganizes when one person plus an LLM can do what five people used to do.

And the model itself is gone. Impromptu is, very specifically, about GPT-4 as it existed in early 2023, with its hallucination rate and capability profile and personality. Almost everything Hoffman demos in the book has been superseded — multiple times — by later models. The hallucination rates he’s defending are no longer the rates you’d see today. The “less wooden style” he keeps asking for is what you get by default now. Reading Impromptu in late 2025 is like reading a 1996 book about web search engines: the frame is right, the specifics are antiquarian.

Who should read this

Read Impromptu if you want a snapshot of how an enthusiastic insider was thinking about LLMs in early 2023. It’s a primary source for the period — useful for the same reason early-internet books are useful, which is that they record what people thought before the dust settled. Read it for the three principles, which are still good. Read the Education chapter and the Homo techne chapter; skim or skip the rest.

Don’t read it for an introduction to LLMs. Anything written in 2024 or later will give you a more accurate picture of current capabilities, current failure modes, and current debates. Don’t read it for a balanced view — Hoffman is committed, and the book reflects that. Don’t read it for the GPT-4 outputs, which are mostly dated and weren’t great even at the time.

The book Hoffman wanted to write is the case for techno-humanism — the case that tools have always extended us, that AI is the next iteration, and that the right response is to engage rather than retreat. He makes that case, more or less. But the case is stronger in the closing essay than in the demos. If you read only the introduction and the Homo techne chapter, you’d get most of what’s worth getting, in about a tenth of the pages.

What survives, ultimately, is the meta-experiment. Hoffman wrote a book with GPT-4 in early 2023, before most people knew what that meant, and let readers watch the process. That’s the artifact that matters. The arguments inside it have been made better elsewhere by now. The act of writing it remains its own thing.