Reid Hoffman's Case Against Doomerism: AI as the Steam Engine of the Mind

Hoffman opens by calling AI somewhere between the largest technology transformation of our lifetimes and of human history. His phrase is "steam engine of the mind": as the steam engine gave us physical superpowers, AI will give mental ones. He expects both amplification of humanity and some substitution of work.

On forecasting, he says people overpredict how much will change next year and underpredict what happens over ten or twenty years. He adds that every technology draws doomsayers first, citing the printing press, which was said to degrade cognition and spread misinformation. What he wants to see more of is the symbiotic amplification loop, a theme of his book *Impromptu*, rather than technologists building autonomous systems alone.

He makes an exception for autonomous vehicles, where effective logistics and saving the roughly 40,000 deaths from human driving matter more than an amplification loop. His example of collaboration is a friend's 15-year-old, interested in organic chemistry, who pastes scientific papers into GPT-4 and asks for an explanation. He says he learned from that and now does the same with dense papers.

Hoffman sees a line of sight to a medical assistant and a tutor for every subject on each of the roughly five billion smartphones, in a world where far fewer than a billion people have access to doctors. He says governments should focus on access, and calls this the biggest potential impact on global health and equity he has seen.

Hoffman says the AI community's pessimism is often well-meaning but conceptually flawed. He finds it amusing when x-risk advocates agree that humans are bad at predicting exponential curves and then make a confident prediction about one. Treating growth in compute as growth in IQ and leaping to superintelligence is hand-waving, he says, and he notes that GPT-4 is already superhuman in some respects without being alarming.

On the six-month pause letter, he recalls asking its authors what they expected, and they said they hoped everyone would pause. He says this misunderstands humanity, since you can't send up a flare and expect eight billion people to stop. The hosts add that doomer scenarios are laid out in great detail while optimistic ones have little color.

Hoffman suggests science fiction inspires technologists with dystopias, because stories need conflict and the machine plays the villain. He says he told Hollywood in 2019 to include good machines too. He also explains why he didn't sign the statement treating AI as an existential risk alongside climate change and pandemics: unlike those, AI has a large positive column, and may even be needed to solve pandemics and climate.

His challenge to critics is that they think trumpeting danger is virtuous, but by only trumpeting negatives they aren't shaping where we could go. He doesn't think technology needs no safety thinking, and says carelessness with viruses or genetic manipulation would be dumb. But the goal is a future much better than the present, and critics should describe how their critiques help get there.

Asked about white-collar workers who fear replacement, Hoffman says he is positive and counters smart AI thinkers who predict rampant unemployment. He does take the medium-term disruption seriously. Steam-engine-era change had huge human consequences, and this transformation will move faster because of the five billion smartphones. Mid-level and some upper-level white-collar jobs will be transformed first and most ferociously.

His advice is to start playing with AI and use it as amplification. People who are 40 or 50 and comfortable may not want their world changed, he says, but the carriage driver and the weavers felt the same way. You don't have to be a perfect student, just engaged in learning the tool the way you once learned Excel.

The good news, he says, is that anything AI creates as a challenge it can help solve. If customer service jobs are displaced, build tools that help those workers work out which jobs they could do, how to find and learn them. He expects AI to take over tasks piece by piece while humans keep learning and adapting.

He also argues there is more time to adjust than the five-alarm-fire narrative suggests. Even if every truck maker began making autonomous trucks tomorrow, he says, it would take about ten years before more than half the trucks on the road were autonomous. That gives individuals, organizations and society time to plan and navigate.

Asked why he isn't retiring, Hoffman says he hates boredom and wants to leave the world better than he found it, and that scale is his thing. He describes how Mustafa Suleyman and Karén Simonyan approached him about Inflection. Nobody could predict what AI would look like in ten years, so the founders picked something that could be a massive independent company: a personal intelligence for everyone.

He describes the product, Pi, as a companion that helps you navigate your life, from fixing a flat tire to debugging a hard conversation with a friend. It has elements of a therapist but isn't one, because it is knowledgeable about the world. If someone says Pi is their only friend, it should help them find other friends, unlike the film *Her*.

Pi is also meant to be compassionate and to have a point of view. If a user declares they are a white supremacist, it shouldn't agree but should kindly push the user to reconsider. Hoffman's examples of unexpected uses include a first-time parent at his firm Greylock asking how to navigate parenting questions on the fly, and someone asking Pi how to fix a flat tire.

His advice for the white-collar worry is to experiment with something that matters to you instead of writing a sonnet. He admits some outputs will be lame, recalling that when he got early GPT-4 access he asked how he could make money investing in AI and got a useless classic-MBA answer. But it is useful for something like an associate asking where to focus due diligence, so the key is finding what it's useful for.

One host raises Western bias in AI ethics, noting that Falcon in the UAE is an open source model partly meant to avoid imposing a Western perspective. Hoffman's baseline is that developers should be honest, open and transparent about what they are designing for. He calls the Silicon Valley claim that technology is value neutral a mistake, and compares it to The Economist or The Atlantic, which state their values and invite accountability.

He adds that as products become ubiquitous in shaping collective mindsets, companies owe responsibility to society as a customer and not only to individual users. Some participation and limits are reasonable, such as not building an AI that enables terrorists or helps advocate genocide. But the uncomfortable truth is that technology is built by small groups, and you can't build AI by committee.

Hoffman says most academics have never built anything and think writing an essay produces a technology. He wants a productive dialogue instead of a witch hunt: say that the product's treatment of race is not good, here is how to improve it and what benchmarks to use. If developers don't listen, he says, he'll go to the streets, because we should be better on bias.

On the argument that we failed to regulate social media, Hoffman says progress comes from deploying, learning and iterating, which means errors are unavoidable. At the UK AI summit he heard that this time innovation would not be allowed before regulating. He calls that dumb because innovation has already happened, nobody knows how to do it, and regulation is "enshrining the past against the future", something he says slows industries down intensely.

Asked about startup whitespace, Hoffman names AI-powered cybersecurity as underdone, along with tools to help white-collar workers through the transition. He thinks the second is neglected because it's not the best economic opportunity, a criticism he admits applies to him as an investor too. He says he thinks of a new AI idea he'd like to co-create almost every week.

For the move from GPT-4 to GPT-5, he expects language-model applications like coding, legal and medical assistants to become more robust. He also expects breadth from scale compute, from protein folding feeding drug discovery to other special-purpose tools. He doesn't know whether those arrive in one or two years, but says the work will intensify, and he is certain some of his predictions will be wrong.

One host gives four areas of improvement: stronger base models with better chains of reasoning, augmentation such as memory and retrieval, bespoke models for areas like biotech, materials and robotics, and experimentation with new architectures beyond Transformers. Hoffman doubts new architectures will matter within one to two years because they need to be scaled. Another host adds that efficiency, as in the models of Mistral, lowers barriers for application developers.

The hosts also look ahead to democratized content creation, with avatars and more controllable video, and to education and healthcare as sectors resistant to improvement. One says that if more people can build useful software, society will change dramatically. Another expects inference platforms and costs to settle within two years, with a clear shakeout in hosting approaches.

Hoffman borrows a slogan from Microsoft's Kevin Scott: the most significant programming language in the next few years will be English, soon followed by Chinese. Someone can already ask an AI agent to build a website they couldn't have coded before. That is part of why he is optimistic about the symbiotic relationship between people and AI.

He urges people not to try to keep the present exactly as it is, but to ask what future we should be making. It is fine to flag a danger and try to avoid it, he says, but where we should be going is the most important context. He also says the way technology is created, by a small group taking a bold risk, is poorly understood outside the industry.

Using a car metaphor, he says only the people in the car have their hands on the steering wheel, so society needs a dialogue with them, as drivers navigating among other cars, not a collective decision about what will happen. He closes by saying the EU AI Act makes him think Europe is holding on to the past so ferociously that it risks sacrificing the future, despite its great AI talent.