Podcast
Live in the Future
Naval
- Buy 2028 Productivity Today
- Paying ~$100k/year on tokens today lets you experience a 2028-level AI workflow now.
- Garry Tan estimates inference capacity will rise ~90,000x in 24–36 months, unlocking emergent capabilities each order of magnitude. (Time 0:02:29)
- Will Scaling Produce Superhuman AI
- The key unresolved question is whether scaling leads to AIs surpassing the smartest humans via recursive self-improvement.
- Naval highlights that progress moved from AI being ‘mid’ at tasks to potentially being ‘pro at everything’ including math and science. (Time 0:04:06)
- Drive Costs Down With Agent Fleets
- Use agent fleets and eval harnesses to massively lower per-user cost and scale utility.
- Farbood built an elastic agentic fleet and drove per-user cost from ~$100/month down to $2.84 via evaluation and optimization. (Time 0:07:11)
- Centralized Models Are A Larger Threat Than Ubiquity
- The biggest public risk is concentration of control: a small group controlling frontier models is scarier than ubiquitous access.
- Naval warns central control can lead to defense-style access, surveillance, and regulatory capture. (Time 0:07:51)
- Making GStack Cost Me Sleep
- Garry described sleeping three hours a night for months to build GStack and an open-source coding package.
- That grind converted him from a ‘cloud code cult’ to mixing cloud and local tooling and producing a top-100 open-source library. (Time 0:10:53)
- Use AI For Brainstorming Not Final Human Copy
- Preserve human-authored outputs when writing for people: use AI for brainstorming and editing but ensure a human curates final content.
- Naval argues AI drafts waste readers’ time unless a person compresses and respects human attention. (Time 0:15:23)
- Lateral Stochastic Drift Finds Banger Ideas
- Cross-referencing diverse vector spaces can surface novel high-value ideas.
- Garry’s ‘LSD mode’ collides many vector spaces and re-ranks outputs to find ‘bangers’ from his 400k markdown corpus. (Time 0:17:30)
- Why Open Source Is Closing The Gap
- Open-source models catch up via multiple vectors: local pretraining, distillation from frontier APIs, algorithmic breakthroughs, and leaked or stolen weights.
- Naval maps how Chinese labs, data access, and security lapses accelerate open-source parity. (Time 0:28:13)
- Instrument Your Company With An Internal AI
- CEOs should instrument company data and build an internal AI to score team impact and identify who to keep.
- Garry and others recommend tying AI to systems to compute KPIs and find the 20% to fire or retain. (Time 0:35:29)
- Value Shifts From Software To AI Research And Hardware
- AI commoditizes software, shifting durable value to AI research and compute supply chains.
- Naval argues China owns hardware supply chains while software commoditization makes AI research and datasets the new chokepoints. (Time 0:50:46)
- Humans Become AI Handlers Not Obsolete Workers
- Humans will become AI handlers or ‘robot trainers’ because desire and taste remain uniquely human.
- Naval and guests see people increasing productivity by running many agents, turning the role into supervising and curating AI outputs. (Time 1:02:41)