May 9, 2024 · 29m · no-priors
No Priors Ep. 63 | With Sarah Guo and Elad Gil
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In No Priors Episode 63, Sarah Guo and Elad Gil analyze the rapid evolution of artificial intelligence, spanning generative music breakthroughs, on-device small language models, enterprise data strategies, massive infrastructure Capex, and nuclear-powered data centers.
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 99.9% of the talking time here. How this is scored →
The hosts as informed peer 6.7 Guest teaching 1.0 Guest disagreement 1.0 The hosts pushing back 0.9
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 7:34 Sarah dismisses Veeva as an apples-to-oranges comparison for OS-level AI
Sarah playfully challenges Elad's historical SaaS analogy, arguing vertical life-sciences compliance is fundamentally different from native desktop AI operating system integration.
Hardest push from the hosts ▶ 6:01 Sarah questions the long-term defensibility of third-party desktop LLM wrappersSarah pushes back against the viability of standalone Mac/Windows LLM indexing tools, comparing them to fragile Android launcher platforms that platforms easily absorb.
Biggest teaching moment ▶ 7:07 Elad explains Veeva's massive market cap built entirely atop SalesforceElad illustrates how an application layer can build a $40B independent business on an underlying platform without being crushed, before eventually swapping backends.
The host holds their own ▶ 18:40 Sarah benchmarks AI capex against historical infrastructure expenditureSarah commands the conversation by citing detailed historic capex metrics from broadband, railroad freight, and oil majors to show $200B AI spend is historically consistent.
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 |
|---|---|---|---|---|---|---|
| Opening Banter and Merchandise Swap | 5 | 1 | 1 | 0 | Sarah and Elad engage in casual opening banter about merch, Bitcoin, and the rapid rise of generative music models like Suno and Udio. The conversation is collaborative and exploratory without friction. | |
| Apple's Small Language Models and Desktop AI | 8 | 2 | 2 | 2 | The hosts debate edge-device LLMs and platform risk, drawing on historical examples like Veeva on Salesforce and early Microsoft Office applications. Sarah mildly challenges whether vertical software analogies translate cleanly to operating-system AI integration. | |
| Balancing On-Device Capabilities and Cloud Compute | 7 | 1 | 1 | 1 | Both hosts analyze the architectural trade-offs between local inference on device processors and offloading compute to the cloud. Sarah draws parallels to traditional client-server compute distribution debates. | |
| AI Hardware Form Factors and Meta AI Scaling | 7 | 1 | 1 | 1 | The discussion covers consumer AI hardware form factors like smart glasses and Meta's massive GPU cluster deployments for Llama models. Sarah points out that Meta's overtraining strategy demonstrates how brute compute scale can outperform standard theoretical efficiency curves. | |
| Model Ownership Strategy for Enterprise Data Platforms | 7 | 1 | 1 | 1 | Elad and Sarah examine whether enterprise data platforms like Snowflake and Databricks need proprietary frontier models. Elad highlights the capital intensity barrier that ultimately concentrates frontier training among hyperscalers. | |
| Global AI Capital Expenditure and Historical Parallels | 8 | 1 | 1 | 1 | Sarah contextualizes the projected $200B annual AI hyperscaler capital expenditure against historical infrastructure spending cycles. She provides concrete figures from oil exploration, broadband rollouts, and railroad expansion. | |
| Expanding Context Windows and Domain-Specific Impact | 8 | 1 | 1 | 1 | Elad discusses expanding token context windows in models like Magic and Gemini 1.5, noting surprising downstream impacts on domain-specific areas like protein folding fidelity. | |
| Data Center Power Constraints and Nuclear Energy | 8 | 1 | 1 | 1 | The hosts break down data center energy bottlenecks, power grid constraints, and the geopolitics of nuclear power adoption. Elad and Sarah both advocate for recognizing energy abundance as an essential AI national security priority. | |
| Episode Conclusion and Audience Call to Action | 2 | 0 | 0 | 0 | Brief outro closing the episode with hat jokes and social channel call-to-actions. |
Statements from this episode (15)
Elad Gil says users will generate custom content via authorized voice clones
“I think I think it was Drake who put out a song, right? Where he had two or three other rappers that he just voice cloned in. And you can imagine a world where you could use anybody's voice, assuming there's permissions and everything else to generate your own…”
Sarah Guo argues easier AI music creation will not guarantee mass adoption
“In like, media platforms in general, they're, like, the ratio varies, right, but there are a lot more readers on X or consumers, like, people who scroll a feed on TikTok, or not TikTok anymore, I suppose, but whatever it is, than creators, and so I think, like…”
Sarah Guo predicts Apple will create interfaces for running local AI models
“And so I think there's a lot of developer demand, and I think it, like, foreshadows we should see Apple creating interfaces for running models locally as part of their ecosystem. That's my general prediction.”
Elad Gil notes major Microsoft Office apps originated as third-party software
“I think people also forget that all of Microsoft Office at some point were third party applications. So in the eighties, there were separate companies like Lotus and others that were providing what turned into Excel. There was a PowerPoint company that was ver…”
Elad Gil says physical size strictly limits on-device LLM reasoning capabilities
“And to some extent, if you look at an LLM, there's like three or four pieces of capability that people care about. There's sort of the reasoning part of it. There's a set of capabilities in terms of what it can do from a synthesis or other perspective. There's…”
Sarah Guo notes overtraining LLMs past optimal compute continues to improve performance
“And if you are meta and you have Somewhere between, you know, 22,000 GPU clusters and 350,000 GPUs available then continuing to train past, like, supposedly optimal points, like, does improve performance apparently and doesn't just fully asymptote as soon as m…”
Sarah Guo says massive frontier AI models are impossible to serve commercially
“Over time, applications are going to want efficient inference, and, like, really large models are impossible today to serve for the vast majority of use cases from a cost and speed perspective”
Elad Gil predicts only hyperscalers will afford subsidizing long-term frontier models
“And so one could argue that in the long run, that's where those types of models should go is the inference platforms probably provide things that are more in that range. And then the hyperscalers and their partners, you know, a handful of them will be at the f…”
Sarah Guo estimates hyperscalers will spend nearly $200 billion on AI compute
“If you think in aggregate, a handful of players in terms of the hyperscalers are spending almost two hundred billion dollars this year on on compute for AI.”
Elad Gil says sovereign nations may drive immense long-term AI compute demand
“And then the one other potential source of immense scale in the long run may be sovereigns as people want to customize models that are specific to their region or customs or language or culture or whatever it may be.”
Elad Gil predicts AI context windows will exceed 10 million tokens soon
“And then it seems lucky that a lot of people will end up in the ten million plus range in the next a year or two or, you know, some reasonable timeframe ahead.”
Elad Gil notes larger context windows significantly improve AI protein folding fidelity
“I think one of the most striking examples of long context window being important is actually some biology models that have come out recently where just increasing the context window for things like protein folding Seems to really make a big difference in terms…”
Sarah Guo points out nobody has built a gigawatt-scale data center yet
“Nobody's built, like, a 500 megawatt gigawatt data center yet, and if you think of it as the equivalent of, like, a nuclear power plant's worth of energy going toward a single data center, it is quite large”
Sarah Guo notes training frontier models requires co-locating GPUs for data transfer
“Today to train these large models, you need all of the GPUs co-located because there is enough data transfer between different chips, right? Between your nodes. And there's a physical constraint on that in that you need to get that much power and to a data, da…”
Sarah Guo believes the AI data wall is solvable via synthetic data
“The number of cheap available tokens on the internet we have used, and now we have to go figure out how to go get more, collect more in the world, or more likely, like, you know or in combination with generating synthetic data. That still feels like a bits not…”