A rational conversation on where AI is actually going | Benedict Evans

Hot take to kick us off: AI is a platform shift on the scale of the web and smartphones—huge, but not the industrial revolution—and like past shifts it deletes some tasks and opens space for new work we can’t name yet.

You’ve been close to the frontier labs, and they’re hiring fast, not shrinking. What’s that signal?

The two‑weeks‑to‑fire‑everyone story is fantasy; you can’t forecast neatly which slices of a job get automated, and the broad anti‑AI mood is real but familiar—we always worry when a new tool shows up.

So what should people actually do to get ready for this future?

Don’t opt out and moralize from the sidelines; jump in, use the tools, and learn where they help you today.

Quick intro before we dive deep: Benedict Evans spent years at a16z and now publishes independent analysis on big tech shifts, with a new deck called AI Eating the World. What are we still underpricing about how AI changes work and life?

Think 1997 internet energy: exciting, messy, most value still unbuilt, adoption wildly uneven, and arguments about where value lands that we can’t settle yet—my deck is basically structured uncertainty.

Where are we on the timeline to everything feeling different?

Software is already in that moment; developers feel a before‑and‑after like accountants seeing the first spreadsheets, while many other fields are testing lightly and puzzling through where it works and where it fails.

Why are labs pouring money into consultants and forward‑deployed engineers instead of replacing them?

Because redesigning workflows, wiring systems together, and retraining people is a project, and companies don’t keep spare teams on ice; you hire Bain or Accenture when the hard part is changing how the organization works.

It’s ironic—many thought consulting would be automated first.

We confuse tasks with jobs; tools make producing a slide easy, but the job is diagnosing the real problem and navigating politics—Jevons‑style, when you cheapen a task you often do more of the overall job.

Amazon ships the exact SKU once you know it; the knowing is the work—same with code, features, and go‑to‑market; that’s why accountants kept growing through PCs, ERP, cloud, and spreadsheets.

What about the coming job apocalypse?

Since 1800 we’ve automated tasks, suffered friction and dislocation, and created better jobs; AI will likely move faster because the rails are laid, but enterprise change still plays out over years, not weeks.

We forget how big the last waves were—an IBM ‘electronic calculator’ ad once promised the equivalent of one hundred and fifty engineers; the internet turned a two‑week research slog into two hours.

Does AGI or superintelligence change the story?

We don’t have a solid theory of human intelligence or why these models work so well, so prognostication is vibes; the definitions keep shifting, and even if progress paused tomorrow, today’s systems would still reshape the next decade.

Zooming out, the ceiling for company scale seems higher than ever. Where does the value pool form?

Software keeps colonizing new parts of the economy, but raw infrastructure often commoditizes—think telecoms moving mind‑bending traffic with low margins—so models may look like utilities while value accrues in the application layer and UX.

If chatbots aren’t the final interface, apps win; if there’s no clear network effect at the model layer, persistent competition means limited pricing power, with most profit shifting up the stack.

Given that, who benefits—incumbents or startups?

Distribution and defaults matter when products feel interchangeable; browsers taught us that shipping everywhere beats minor feature tweaks, and Apple’s on‑device assistant vision shows the model can be the engine while the OS owns the experience.

Anti‑AI sentiment seems to be growing. What’s underneath it?

It’s a messy bundle: some local pain, myths like data centers ‘draining water’ when their usage is around 0.017 percent in the United States, unclear jobs data, plus culture‑war fights in creative fields and a vacuum of hard usage metrics.

How does this change how you parent in an AI world?

It depends on age; near‑term grads face volatility, younger kids will see things settle; the evergreen lesson is media literacy and perspective rather than micromanaging every minute.

Every technology wave brings real harms and overreactions—deepfakes are new and dangerous, and we’ve also had old‑tech disasters like the UK Post Office scandal—so be vigilant without spiraling.

Any jobs you’d steer your kid toward or away from?

Careers zigzag; aim to overlap what you’re good at, what the market needs, and what pays—hit two and you’re viable, three is ideal.

What’s a smart AI question not enough people ask?

Whether foundation model companies will truly have pricing power, and how to separate a button‑level task from the actual job; streaming shows how tech first makes the old thing cheaper, then unlocks a new model entirely.

Engineering flipped from ‘safest job’ to fastest transformed. How do we judge exposure?

Scoring jobs by automatable percentages is a dead end; surprises come from left field, like ride‑hailing upending taxis, and even ‘safe’ roles like personal training could shift when a phone can watch your form and coach you.

For listeners anxious about their careers, what’s your playbook?

Averages hide pain, but your best move is to immerse yourself, master the tools, and show how they make you better; refusing to use AI won’t help you in the interview.

AI corner: one way you use it that others might try.

I lean on it for proofreading, images, and even redesigning rooms, and I dictate most writing and let transcription handle the rest—chatbots still need clear use‑case wrappers to shine for my kind of research work.

Anything you want to leave listeners with before we sprint to the lightning round?

I traffic in frameworks, not stock tips—and sometimes the most honest answer is that it depends.

Lightning round: books, shows, products, mottos, and your phone collection?

Books: Three Men in a Boat and Cronon’s history of Chicago for lessons on standardization and logistics; watch a classic like The Seventh Seal; no breakout consumer AI app yet; motto is it depends; and yes, I hoard about twenty pre‑iPhone era phones because that’s when form factors were wild.

I still have that old Ericsson shark‑fin flip and a 2001 Japanese phone with a camera and a color screen that wowed clients even though it barely worked outside Japan. Back then we imagined lots of different shapes and keyboards, but everything ended up converging into one kind of device.

This was fantastic; I learned a ton and feel better after this. Two quick ones to close: where can folks find you and this presentation, and how can listeners be useful to you?

I’m always trying to learn new things and ask better questions, so help by pushing me when I’m repeating myself. Last year I kept hammering that these models still hallucinate, and if you press them they still do; that does not make them useless, it just means we need to keep pushing ourselves forward.

You can find more at bendashevans.com and EVAN.com. Subscribe on Apple Podcasts, Spotify, or your favorite app, leave a quick rating or review, and find past episodes at Lenny'sPodcast.com; see you in the next one.