Overview of A rational conversation on where AI is actually going | Benedict Evans
Benedict Evans argues that AI is a major platform shift on the scale of the internet or mobile, but not necessarily bigger than either. His core message is that the technology is real and transformative, but the near-term narrative is being distorted by hype, doom, and simplistic job-loss predictions. The conversation focuses on how AI will actually diffuse through companies, where value will accrue in the stack, why adoption will take time, and why the biggest winners may be the application and distribution layers—not the model labs.
Key Thesis: AI Is a Platform Shift, Not Magic
- Evans’ “controversial” view: AI is as big a deal as the internet or mobile—but not obviously bigger.
- He compares the current moment to the internet in 1997:
- exciting,
- messy,
- incomplete,
- and still lacking many of the products and workflows that will eventually define the era.
- Most people are not using AI deeply yet:
- some power users are far ahead,
- many others barely use it weekly or at all.
- The main lesson: we know it matters, but we don’t know exactly how it will settle out.
Why the Job Apocalypse Narrative Is Overstated
Automation has always removed jobs—and created new ones
- Evans rejects the idea that AI is uniquely job-destroying.
- Historical pattern:
- technology automates existing tasks,
- that lowers costs,
- new uses emerge,
- new jobs appear that didn’t previously exist.
- He argues that people can usually see the old jobs that will disappear, but not the new ones that will be created.
Why current “X% of jobs can be automated” claims are weak
- He dismisses attempts to score professions task-by-task as mostly nonsense.
- Jobs are not cleanly decomposable into automated vs. non-automated pieces.
- The real question is often:
- What is the task?
- What is the job actually hiring a human for?
Adoption will be slower than doomers think
- Enterprise adoption is constrained by:
- procurement cycles,
- integration complexity,
- workflow redesign,
- trust and reliability concerns.
- AI won’t instantly replace entire departments.
- More likely: 3–10 year transformations by sector, not overnight disruption.
Where AI Actually Changes Work
Software development is the clearest near-term example
- Evans says software is already in the “before and after” moment.
- Developers can now generate large amounts of code quickly.
- But this does not mean the software industry disappears.
- Instead, it changes:
- how software is built,
- how much software gets built,
- who can build it,
- and what the bottlenecks are.
The hard part is often not the task
- Example: consulting.
- AI can generate slides.
- But clients hire consultants for:
- diagnosis,
- organizational politics,
- interviews,
- strategic judgment,
- implementation planning.
- Example: software.
- AI can write code.
- But it does not automatically know:
- what product to build,
- which features matter,
- who the customer is,
- how to go to market.
- Example: Amazon.
- It can get you the SKU.
- But choosing the right SKU is another job entirely.
Value Capture: Model Labs vs. Application Layer
Model labs may not have durable pricing power
- Evans is skeptical that foundation model companies will capture most of the long-term value.
- Why:
- models appear increasingly commodity-like,
- there may be little winner-take-all network effect,
- competition may persist,
- and users may not care which model powers a feature.
The likely winners are higher up the stack
- If models become interchangeable, value shifts to:
- applications,
- workflows,
- distribution,
- brand,
- and integration.
- His comparison:
- model labs may end up more like cloud infrastructure than Windows.
- cloud is important, but the biggest profits often accrue above it.
“Selling intelligence on a meter” may not work as advertised
- Evans questions whether AI will behave like a utility with stable pricing power.
- He notes that other foundational infrastructures (telecom, electricity) often become:
- essential,
- huge,
- and low-margin commodities.
Distribution Is Becoming More Important
- As products become easier to build, distribution becomes a bigger moat.
- This favors:
- incumbents,
- companies with default placement,
- platforms that already have user traffic.
- Examples discussed:
- Google using distribution to push Gemini,
- Meta using its surface area to keep AI products visible,
- Apple potentially embedding AI into devices and OS-level workflows.
- The key point: if the model is just the engine, the real battle is who owns the interface and customer relationship.
Why He’s Not Panicking About AI Backlash
Anti-AI sentiment is real, but messy
- Evans sees a blend of:
- legitimate concerns,
- cultural backlash,
- labor anxiety,
- environmental fears,
- and online moral signaling.
- Some complaints are grounded in reality:
- job displacement,
- misuse,
- deepfake harms,
- local infrastructure strain.
- Others are overstated or factually weak:
- especially around water usage and data center panic.
This resembles the social media backlash
- He likens the current moment to the social media era:
- some criticisms were true,
- some were exaggerated,
- some were simply wrong,
- and the public conversation often collapses them together.
What People Should Do Now
Don’t avoid AI out of moral superiority
- His advice is blunt:
- don’t just declare AI evil,
- don’t opt out psychologically,
- don’t pretend it won’t matter.
- That stance may feel good online, but it won’t help your career.
Learn by using it
- The best response is to:
- dive in,
- experiment,
- learn where it works and where it doesn’t,
- and figure out how to become valuable in a world where AI is normal.
- For job seekers, especially in professional services:
- saying “I hate AI” is probably a losing interview strategy.
- better to show you understand how to use it.
Practical Uses Evans Actually Has for AI
- He says AI is not yet best for the exact kinds of tasks he most wants automated.
- But he does use it for:
- proofreading,
- image generation,
- apartment redesign / interior visualization,
- voice dictation and transcription.
- His broader point: many AI use cases are embedded and invisible rather than flashy standalone apps.
Important Mental Models from the Conversation
“1997 moment”
- We are early.
- The infrastructure exists, but the full product ecosystem doesn’t.
“Jagged frontier”
- AI works extremely well in some places and poorly in others.
- It is hard to predict in advance where the boundary will be.
“Task vs. job”
- AI may replace tasks before it replaces jobs.
- Sometimes the task is the job; often it isn’t.
“Old thing, but more”
- New technologies first amplify old behaviors.
- Only later do they enable entirely new categories of products and work.
Notable Takeaways
- AI is likely to create major economic change without necessarily causing mass immediate unemployment.
- The model companies may be powerful, but they may not own most of the profit pool.
- Distribution, brand, and application design are likely to matter more over time.
- The most important thing for individuals is to learn the tools and adapt, not posture from the sidelines.
Closing Advice
Evans’ message is broadly optimistic:
- things will change,
- some disruption will be painful,
- many predictions will be wrong,
- but we have lived through this pattern before.
His core recommendation is simple:
- engage with AI directly, understand it deeply, and stay flexible.
If you want, I can also turn this into:
- a shorter executive summary,
- a bullet-point “key lessons” version,
- or a quote-heavy recap.