May 1, 2023 · 38m · another-podcast
Working out AI questions
0:00 / 0:00
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Benedict Evans and Toni Campbell establish a pragmatic analytical framework for generative AI by examining compute economics, user interface design, error management, and safety governance beyond speculative AGI hype.
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 83% of the talking time here. How this is scored →
The hosts as informed peer 7.6 Guest teaching 0.1 Guest disagreement 0.1 The hosts pushing back 0.4
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 13:18 Debating feature awareness versus invisible ML
Toni momentarily pushes back on the utility of device features by asserting that 80% of phone tech goes unused, prompting Benedict to clarify his distinction.
Hardest push from the hosts ▶ 13:26 Reframing ambient computing versus user discoveryBenedict explicitly rejects Toni's framing with 'Well, no, that's not quite what I mean,' redirecting the conversation to invisible background processing.
Biggest teaching moment ▶ 15:08 Explaining manual multi-step podcast productionToni details the specific sequencing and tool fragmentation involved in audio production (Descript, Adobe Sound) to demonstrate where automated AI workflows would eliminate real friction.
The host holds their own ▶ 10:49 Mainframe economics analogy for consumer AIBenedict demonstrates domain mastery by contextualizing LLM subscription costs against the historical evolution of compute architectures, from mainframes to zero-marginal-cost web search.
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 |
|---|---|---|---|---|---|---|
| Parking the Artificial General Intelligence Debate for Pragmatic Analysis | 8 | 0 | 0 | 0 | Benedict establishes a clear analytical taxonomy, arguing that speculative AGI debates must be parked to examine practical near-term questions like compute costs, error rates, and interface design. Toni readily agrees with the framing. | |
| Compute Economics, Scaling Laws, and Emerging Market Structures | 8 | 0 | 0 | 0 | Benedict provides an in-depth breakdown of scaling economics, contrasting commodity machine learning with centralized LLMs and noting the historic return of per-query marginal compute costs. Toni provides affirming commentary. | |
| Embedded Machine Learning and Frictionless Creator User Experience | 7 | 1 | 1 | 2 | Benedict gently corrects Toni's premise regarding unused smartphone features, explaining that modern machine learning operates invisibly in the background on image sensors and audio processing. Toni relates this to audio editing workflows. | |
| The Error Rate Dilemma and the Infinite Interns Analogy | 8 | 0 | 0 | 0 | Benedict details his 2x2 framework for evaluating error rates and shares his 'infinite interns' analogy for tasks needing human verification. Toni reflects on how quickly public perception has shifted from marveling at AI to scrutinizing flaws. | |
| Moving Beyond Text Prompts to Specialized Interface Design | 8 | 0 | 0 | 1 | Benedict critiques the open text prompt as a temporary interface paradigm, arguing that mature software will adopt specialized buttons, sliders, and constrained palettes. Toni agrees, highlighting vocabulary barriers in prompt engineering. | |
| Trust, Safety Filters, Jailbreaks, and Copyright Complexities | 8 | 0 | 0 | 0 | Benedict analyzes LLM trust, safety bypasses (such as grandmother jailbreaks), and complex copyright analogies spanning synthesizers, auto-tune, and music sampling law. Toni adds observations on creative attribution. | |
| Synthesizing Core AI Questions and Concluding Episode Reflections | 6 | 0 | 0 | 0 | Benedict and Toni wrap up by reiterating that foundational questions surrounding cost, reliability, interface, and safety provide the necessary compass for navigating rapid AI ecosystem shifts. |
Statements from this episode (15)
Benedict Evans: AI moves faster than past shifts due to ready infrastructure
“Well, so the sort of meta observation we were chatting about earlier is that this is moving massively faster than most recent things in tech. Cause no one needs, you don't need to buy a new device. You don't need to learn any programming language. Like, well, …”
Benedict Evans: Veteran AI researchers disagree on whether LLMs lead to AGI
“All the people who've actually been working on this for 20 years don't agree. And most of them say we don't know.”
Benedict Evans: Assessing AI's commercial impact requires parking the AGI debate
“To actually kind of make any questions about what companies are going to happen and how we're going to use this, you always have to kind of park the AGI conversation, because if it happens, then all bets are off.”
Benedict Evans: LLMs currently generate one completely wrong paragraph out of four
“The second question is right now, if you ask this a question and the result is three or four paragraphs, one of those paragraphs will be completely wrong.”
Benedict Evans: Fixing LLM hallucinations does not require achieving AGI
“And you could say, well, if that gets fixed, then we've got AGI, but actually I didn't think that's true because one way to fix it might be that it could just say, I don't know. The problem at the moment is it might not know and give you make something up, or …”
Benedict Evans: AI model training costs will drop as efficiency improves
“And those are all, those points on that plot, all of like historically have gone up and to the right, but now they're going to start going down because the models get more efficient and people work out better ways of training them when you optimize all the ind…”
Benedict Evans: Banning facial recognition is absurd because ML is commoditized
“But in general, machine learning has become a commodity very quickly which is sort of the absurdity of people saying, well, we have to ban face recognition, or you can ban it, but that won't work. Cause it's just a commodity technology.”
Benedict Evans: Generative AI adds real marginal costs to consumer interactions
“But the idea that like it might cost, it's like it actually has more, a consumer internet company has real tangible marginal cost every time a consumer does something. That's kind of a new thing.”
Benedict Evans: Generative AI search is risky because errors are undetectable
“The challenge is as I always, I've always been saying, if you use this as a general search, then the error rates might matter and you can't tell.”
Benedict Evans: Narrower scopes and expert users make AI errors highly visible
“The more narrow the product get and the more expert the user, then the more visible the app, the errors.”
Benedict Evans: Professional AI workflows require UI controls, not prompt boxes
“As soon as you move away from make this shot better, you're probably moving away from a prompt box, and you're probably moving to buttons, and switches, and options, and palettes, and lists of possibilities, and a screen that says, okay, here are six different…”
Benedict Evans: ChatGPT's chat UI deceptively implies definitive answers
“To me, this is part of the challenge of ChatTPT versus Google is that Google says here are 10 links that might be the answer. Whereas ChatGPT says this is the answer, and it's, this is the answer. But it's not, you could argue that what it should be saying is …”
Benedict Evans: Users will consistently bypass probabilistic LLM safety filters
“You can try and stop it, but people will find ways of getting around it because it's not just doing a binary database lookup. So you might have a keyword filter on Napalm and people will come up with a way of describing Napalm without using the word Napalm.”
Benedict Evans: Unpredictable LLM outputs challenge China's strict internet censorship
“There's a whole narrative that for China, this is a problem deploying this to, for general purpose, consumer ads. Because China has really, really rigid ideas about what you can and can't say on the internet. And this is not going to be controllable, or that m…”