Benedict Evans says AI is as big a deal as the internet or mobile, and only that big, comparing today to the internet in 1997.
Evans argues that predictions of mass layoffs ignore how work is structured, since tasks can be automated while jobs persist, as elevator operators showed.
Evans says model companies may face little differentiation, so value will likely flow up the stack to applications and distribution rather than to model makers.
AI Is as Big as the Internet, and No Bigger
AI is a platform shift on the scale of the internet or mobile, and history is a better guide to it than apocalypse talk. Work is made of tasks inside slow-moving organisations, so jobs shift more than they vanish. Models may become interchangeable, which would push value toward applications and distribution.
Takeaways
AI is as big as the internet, no bigger. Evans compares today to the internet in 1997, when most things did not yet work and most of what people would build did not exist. Even if model progress stopped, AI is useful enough to keep rolling out for ten years.
Tasks get automated while jobs mostly persist. Accountant numbers grew through the 20th century despite spreadsheets, because cheaper tasks raise demand through price elasticity. Enterprise sales cycles are also slow. Elevator operators are the exception, since there the task was the job.
Mass layoff forecasts ignore how work is structured. Evans says Dario Amodei's entry-level jobs commentary is outside his field, and he rejects putting an automation percentage on a senior law partner's work. Needed organisational change explains why AI firms invest in consultancies like McKinsey and Accenture.
Model makers may lack pricing power. A normal person sees no difference between Gemini and Claude. Evans says Altman's electricity meter idea misreads thin utility margins. Value may sit higher in applications, more like cloud than Windows. Mobile telecoms spend about $200bn a year on capex for low margins.
Distribution may beat model quality. Google is pushing Gemini hard, and Apple Intelligence on iOS is expected to run on Gemini. Meta was behind in AI even before the Llama news. ChatGPT's 900m chat users mirror the 900m people already online.
Nobody can define or predict AGI. The definition keeps moving, and AI is whatever has just started working. Adoption is still early, with survey data showing many teenagers are not regular users. Evans likens developers today to accountants first seeing VisiCalc.
Being early is not the same as right. Evans says there is a huge difference between being right and being early, and between the right company and the right price. In 1997, asking whether Excite or Yahoo would win got the answer generally wrong.
The anti-AI backlash, sorted by Evans
Water worries are overstated. A Lawrence Livermore study put US data center water use at about 0.017 percent of US water use.
The effects on energy and employment remain unclear, and there is no clear consensus yet that AI is affecting jobs.
Deepfakes are a real new harm. A 15-year-old could now fake pornographic images of every girl in their high school in one afternoon, which Photoshop did not allow.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Every technology can ruin lives. His example is the UK Post Office scandal involving Fujitsu software.