Dec 8, 2024 · 42m · another-podcast
Ai eats the world
0:00 / 0:00
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Technology analyst Benedict Evans and Tony Cameron explore Evans's annual presentation deck 'AI Eats the World,' analyzing generative AI through historical platform shifts, enterprise adoption constraints, model commoditization, and the practical dynamics of probabilistic computing.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 85.4% of the talking time here. How this is scored →
The hosts as informed peer 8.2 Guest teaching 0.5 Guest disagreement 0.2 The hosts pushing back 1.7
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 29:47 Tony counters enterprise skepticism with successful visa drafting example
Tony offers a slight counter to Benedict's premise that regular people do not find unprompted software workflows by describing how ChatGPT unexpectedly drafted legal paperwork for her O-1 visa.
Hardest push from the hosts ▶ 30:53 Evans counters visa success with probabilistic error rate realityBenedict pushes back on relying uncritically on LLM drafting by citing his own Indian visa paperwork experience where ChatGPT produced believable but inaccurate results.
Biggest teaching moment ▶ 28:50 Tony shares unexpected bottom-up legal drafting use caseTony educates Benedict on a practical workflow where a non-technical user successfully employed conversational prompts to generate legal authorizations accepted by an immigration attorney.
The host holds their own ▶ 10:25 Evans lays out hyperscaler capex and semiconductor cyclicalityBenedict demonstrates authoritative industry knowledge by dissecting the $200B hyperscaler data center buildout, marginal cost curve mechanics, and investment bank bubble dynamics.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| The Feeds and Speeds Phase versus Practical Enterprise Adoption | 8 | 0 | 0 | 2 | Benedict Evans drives the discussion by comparing current AI benchmarking to the feeds and speeds era of PC magazine hardware tests in the 1980s and 90s. Tony Cameron acts as an interviewer and agreeable facilitator. Evans cites historical internet analogs like local mall websites and running web servers on PCs to frame AI adoption challenges. | |
| Bucket One: Scaling Frontiers, Massive Capex, and Model Commoditization | 9 | 0 | 0 | 2 | Evans delivers a deeply analytical overview of capex spending across cloud hyperscalers, cyclical semiconductor dynamics, and model commoditization. Tony poses an initial prompting question about whether broader tech trends live in a separate vortex from AI. Evans provides extensive industry data, referencing Bain surveys, Nvidia supply constraints, and Amazon Nova model releases. | |
| Bucket Two: Probabilistic Computation, Error Rates, and Disruption Dynamics | 8 | 1 | 0 | 2 | Evans outlines the nature of probabilistic computing, comparing LLM limitations to early Apple II systems and mid-90s web limitations. Tony asks thoughtful framing questions on present versus future trajectories and agrees with Evans's observations. Evans uses vivid analogies like the Yoda quote and mid-90s internet predictions to detail error rates. | |
| Bucket Three: Enterprise Deployment Realities versus the Maximalist Vision | 8 | 0 | 0 | 1 | Evans deconstructs how enterprise IT adoption works in waves, citing enterprise cloud penetration metrics and Okta SaaS application data. He challenges bottom-up enterprise AI assumptions with humorous references to The Office and Home Depot electric motors. Tony supports the pacing by chiming in with an anecdote about fax machine phase-outs in banking. | |
| User Experience Evolution, the Infinite Interns Model, and Core Industry Answers | 8 | 2 | 1 | 2 | Tony shares a personal anecdote using ChatGPT to draft legal corporate authorizations for an O-1 visa application, illustrating surprising end-user discovery. Evans responds by contrasting her positive experience with his own trip visa documentation errors, reinforcing his infinite interns thesis. Evans walks through product history including early web browsers, Tim Berners-Lee, Ajax, and key product hires at Anthropic and OpenAI. | |
| The Lineage of Tech Shifts and the Inevitable Invisibility of AI | 8 | 0 | 0 | 1 | Evans explains the lineage of Marc Andreessen's Software Is Eating the World essay and his own 2013 Mobile is Eating the World presentation. He details how AI might either become an ambient background layer or sit atop the stack, referencing historical census data on Otis automatic elevator attendants. Tony harmoniously wraps up the episode by directing listeners to the published slide presentation. |
Statements from this episode (13)
Evans: Big enterprises are running AI trials, but few use it broadly
“Every big company's got a trial. No one's really using it very much, or not many people are using it very much. There's a small number of fields where you had this immediate obvious use case, like it's great for coding, it's great for marketing, it's great for…”
Evans: Amazon generates $55 billion in annual advertising revenue
“Amazon has a three, has a fifty-five billion dollar ad business.”
Evans: Instacart generates almost all of its profits from advertising
“Almost all of Instacart's profits come from advertising.”
Evans: Only 2% to 4% of developed-world consumers use generative AI daily
“Every consumer has tried this, but well, a very large portion of consumers, something like a third to a half of all consumers in the developed world have now tried this, but only like two, three, four percent of people are using it every day.”
Evans: Top four cloud platforms will spend over $200B on data centers this year
“And we will build the five, four platform companies, four big cloud companies will spend something over two hundred billion dollars this year, which is up 80 or ninety billion dollars from last year on building data center.”
Evans: LLMs generate probable answers, unlocking value only where exact correctness isn't required
“These models are very good at automating certain kinds of things, but not others. They get stuff wrong, or rather what they do is they say, what would an answer to this question probably look like? That's about the best way I can come up with formulating it ri…”
Evans: Testing LLMs on complex logic is like testing Apple II for five-nines uptime
“This is, you know, classic disruption that the new thing tends to be crap at the stuff that was important for the old thing, but then it does something different that's new. And if you test an LLM by asking it to do complex logic problems or masquerade problem…”
Evans: Accenture generates $1 billion per quarter in generative AI bookings
“And so Accenture is now doing a billion dollars a quarter in what they call generative AI bookings, doing pilot projects and building stuff for big companies.”
Evans: Cloud computing accounts for only 30% of enterprise workflows
“All these kind of basic procurement, IT procurement questions, which are why IT, why cloud is still only sort of 30% of enterprise workflows.”
Evans: Typical large enterprise runs 400 to 500 SaaS applications
“Even at only a third of their workflows, it's sort of four or 500 SaaS apps. If you look at data from Okta or Productive or companies like that.”
Evans: Bottom-up software sales only captures 5% of enterprise users
“There's this whole sort of thesis of bottom up enterprise software sales that you don't need to sell to the CIO. You make this amazing cool tool and the users will discover it and use it. And what actually happens is that most that that always works for like t…”
Evans: Perplexity does nothing OpenAI cannot do, winning solely on UX
“Theoretically perplexity isn't doing anything that you couldn't do with OpenAI, but it said, let's narrow it and make it more specific and build product around that and make a UX around that and communicate to the users, this is what it's for, not those other …”
Evans: AI will become invisible background infrastructure within five years
“It just, like, it just disappears. You don't think about it anymore. You know, when's the last time you did autocorrect? You know, you don't, you just, it just happens. When's the last time you used the cloud? Like it just happens. And so that's sort of the ba…”