Ep. 023 - Everyone Leaves Google, Elon Forecasts 1T ARR, Reflecting On GPT-5 | Jon from Asianometry

Summary

The hosts debate GPT-5’s trajectory, arguing the initial release was underwhelming but the broader direction was right once code-focused improvements kicked in and later variants improved materially. A major thread is Google: the speakers view the company as losing key talent, especially Jeff Dean, while also facing a structural problem in which the best frontier-AI people increasingly prefer startup-like labs such as OpenAI and Anthropic. They argue Google’s broad product empire and bureaucracy make it harder to stay AI-native, even if it remains strong in infrastructure and TPUs. The discussion then turns to export controls, with a specific claim that banning Chinese transceivers and other data-center components could hurt Western firms more because China already dominates much of the supply chain. The final stretch becomes a technically detailed conversation about AI-driven software creation, security breakage, and the possibility that local inference clusters could reduce reliance on the public internet.

Notes

• The episode opens around AI model quality, Google turnover, China component bans, and Elon’s revenue timeline.
• GPT-5 is framed as conceptually right, even if the first release was not especially strong.
• Later GPT-5 variants are described as substantially better than the early release.
• The Google story is framed as leadership churn across Gemini and adjacent projects.
• Jeff Dean’s departure is treated as especially important because of his infrastructure legacy.
• Google’s AI talent problem is described as structural, not just personnel-specific.
• The hosts contrast frontier labs with Google’s broader product surface area.
• One sharp Google critique is that it mostly acquired or absorbed innovation rather than building it internally.
• The speakers argue that many major Google products were bought or effectively bought.
• Maps is debated as an example of an “effective acquisition” via team purchase.
• GCP is portrayed as a product Google partly invented but later improved through an acquisition-linked executive.
• Google’s transformer invention is used as an example of inventing a core technology and then losing the commercialization race.
• The strategic forecast is that Google may become a TPU-first infrastructure company with a merely good-enough transformer product.
• A policy idea is floated to keep Google’s search monopoly while forcing science licensing, AT&T-style.
• Google is explicitly compared to Bell Labs: scientifically strong, monetization weak.
• The “killed by Google” theme is used to argue Google often creates value it later does not capture.
• Search is described as the cash engine that funded speculative internal R&D.
• The hosts say semiconductor R&D is more team-based than model-lab drama implies.
• They emphasize that semiconductor development is collective and highly labor-intensive.
• Jeff Dean is singled out as an unusually cross-stack and high-leverage engineer.
• The model-lab argument is that compute and data may matter more than single individuals in future Gemini releases.
• OpenAI and Anthropic departures are cast as a kind of religious schism.
• The hosts joke that AI competition may escalate into long-running fights over data centers.
• A recurring theme is models demonstrating jailbreaks and containment escapes as a kind of competition metric.
• They speculate the public internet may become less useful if local inference becomes ubiquitous.
• The internet is framed as dangerous because malicious AIs could infect computers and steal money.
• A major technical thesis is that no one person can fully understand modern software or chip systems end-to-end.
• The speaker predicts persistent, superhuman software breakage at both high and low levels.
• The proposed fix is a “great firewall,” presented half-seriously as a defense against AI harm.
• The joke extends to a physical moat or trench for humanoid-robot containment.
• Elon’s Terafab and SpaceX-related compute plans are linked to new chip demand.
• The hosts interpret Elon’s scale-up as pulling a trillion-dollar revenue goal forward by a year.
• They say hyperscalers are effectively spending free cash flow to zero for chips.
• Taiwan is described as power constrained and centered on fabs rather than data centers.
• AI use in Taiwan is portrayed as still relatively limited and mostly translational or entertainment-oriented.
• The speaker says they built their own video editor.
• The editor exists because their workflow depends on aggressive keyboard-shortcut usage.
• The custom editor is described as about 40 MB and highly optimized for the speaker’s workflow.
• They say they tried four times before getting the editor design to work.
• The first attempt started from an audio file and was not satisfying.
• They eventually reduced the project to one core feature and handed a large spec to Codex.
• Later model/tool iterations are said to have unlocked features that earlier versions could not.
• The agentic coding workflow involved many questions but could one-shot features.
• A specific feature was auto-filling subtitle text blocks when adding images.
• The claimed total cost for the tooling was only a couple hundred dollars.
• The editor integrates Wikipedia comments search and video search with attribution.
• The speaker also built audio post-processing software that improves output and shows a waveform.
• The broader takeaway is that AI coding tools are already good enough for bespoke, low-cost niche software.