All-In PodcastSpaceX’s $2T Case, Nvidia’s Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?
CHAPTERS
- 0:00 – 2:26
Gavin Baker returns: big week setup and why Karpathy’s move matters
Jason welcomes Gavin Baker (Atreides) and sets the agenda: SpaceX, AI company moves, and Nvidia. The conversation quickly centers on Andrej Karpathy joining Anthropic and what that signals about the next phase of model development.
- •Gavin Baker joins as guest; Sacks is out
- •Show roadmap: SpaceX, OpenAI/Anthropic, Nvidia, macro
- •Karpathy’s reputation (OpenAI founding member, Tesla FSD, “vibe coding”)
- •Anthropic’s plan: a new pre-training team led by Karpathy
- 2:26 – 8:02
Recursive self-improvement and continual learning: the next AI frontier
Chamath and Gavin frame Karpathy as a “wave rider” of AI breakthroughs and argue recursive self-improvement could accelerate model gains dramatically. They connect this to the “bitter lesson” (compute + scale) and the open question of when AI will substantially improve AI with minimal human input.
- •Karpathy’s history: Tesla data labeling, OpenAI, and fast prototyping
- •Recursive self-improvement as an “overdrive + autopilot” path for models
- •Gavin: continual learning + recursive improvement as key remaining frontiers
- •Expectation of faster-than-linear quality improvements (a new Moore’s Law)
- 8:02 – 18:04
Small models, new architectures, and “don’t breathlessly track every release”
Friedberg argues the most important developments are end-user achievements, not incremental model headlines. The group debates model architectures (networks of smaller models), falling inference costs, and the tension between product rollouts and public perception—sparked by talk of Gemini Nano appearing inside Chrome.
- •Friedberg: focus on end-user outcomes (math, drug discovery) vs model gossip
- •Architectural shift possibilities: smaller models cooperating to reduce cost/token
- •Gemini Nano in Chrome becomes a flashpoint for privacy/communications
- •How phrasing and framing can unintentionally create an “AI boogeyman”
- 18:04 – 25:11
Why America is turning on AI: inequality, foreign influence, and anti-human framing
The panel digs into the backlash against AI, including boos at commencement speeches and rising public skepticism. Friedberg lays out a multi-causal explanation: perceived power concentration, geopolitical information operations, and a deeper psychological discomfort with non-human-centric intelligence.
- •AI seen as leverage accruing to a small group; diffusion feels slow/asymmetric
- •Foreign state actors may amplify anti-tech narratives (historical parallels)
- •AI can feel “anti-humanist,” triggering identity/ego threat reactions
- •The need to communicate benefits clearly and credibly to the public
- 25:11 – 31:53
Should we slow down? Executive orders, KYC, and frontier model testing debates
A pulled Trump AI executive order becomes the springboard for regulation talk. Chamath emphasizes minimal but meaningful guardrails (like KYC) to prevent extreme misuse, while Gavin warns about the one-way ratchet of government power and notes existing legal liability already pressures responsible behavior.
- •Discussion of the scrubbed/pulled AI executive order and what it implied
- •Chamath: KYC-style controls to reduce bio/terror misuse risk
- •Gavin: concern about permanent expansion of government oversight
- •Existing enforcement via courts and liability as a de facto regulator
- 31:53 – 34:00
Automation, jobs, and safety: self-driving bans, robot taxes, and local governance
The group evaluates proposals like paced self-driving rollouts and taxing humanoid robots. They argue safety (reduced traffic deaths) will overwhelm job-protection arguments, and predict a patchwork rollout where cities/states compete on policy—setting up a broader debate about surveillance and public safety tech.
- •Questioning whether workers actually want certain high-churn jobs
- •Self-driving: wrongful-death liability and safety incentives push adoption
- •Municipal/state patchwork as a de-risking mechanism for new tech policy
- •Transition from national rhetoric to local, practical implementations
- 34:00 – 37:11
AI-enabled policing and privacy: Flock Safety, gunshot detection, and Vegas as a case study
AI surveillance tools become the concrete example of “tech vs privacy vs safety.” Gavin and Chamath cite gunshot detection and Las Vegas policing operations as evidence that crime can be deterred with modest budgets, while Jason emphasizes audit trails and retention limits as pragmatic privacy protections.
- •Cambridge turns off gunshot detection; debate over unintended consequences
- •Las Vegas policing: drones, mission control, rapid response as deterrence
- •Jason: privacy-by-design options (retention limits, audit logs, policy controls)
- •Claim: ‘crime is a choice’ when detection and response become reliable
- 37:11 – 45:13
AI layoffs and the PR crisis: Cloudflare memo, Zuckerberg’s approach, and messaging failures
Jason argues layoffs are now credibly tied to AI-driven efficiency, and that CEO messaging is inflaming public fear. The segment highlights Cloudflare’s “measurers” framing and Meta’s training/monitoring rhetoric, with Chamath blasting tech leaders for communications that stigmatize workers and accelerate backlash.
- •Jason: layoffs no longer just ‘post-COVID bloat’—AI is a real driver
- •Cloudflare: strong financials but major cuts framed around AI efficiency
- •Meta: fear that employees are ‘training their replacement’
- •Chamath: CEOs/PR teams are mishandling the moment and fueling revolt
- 45:13 – 48:02
SpaceX S-1 teardown: three businesses and the path to a $2T+ valuation
The show pivots to SpaceX’s IPO filing and a valuation debate centered on business mix and growth. Jason breaks down Starlink, the launch business, and the AI/compute segment—highlighting how rapidly the AI line item could reshape SpaceX’s revenue profile.
- •IPO scale and valuation framing (largest-ever IPO ambitions)
- •Starlink as current “money printer” with major subscriber upside
- •Launch/space segment economics vs growth and losses
- •AI segment revenue growth but heavy operating losses and capex intensity
- 48:02 – 49:39
Elon Web Services: Colossus build speed, Anthropic off-take, and compute as a product
Gavin spotlights the most surprising S-1 implication: SpaceX/xAI’s data-center build capability and how quickly it can be replicated. They discuss why GPU allocation rewards whoever can energize racks fastest, and how a large offtake partner can catalyze a new hyperscale compute business.
- •Data centers built faster over successive iterations (speed as advantage)
- •Why GPUs flow to teams that can convert electrons into tokens quickly
- •Anthropic as a massive compute customer (with cancellation flexibility)
- •Compute services could become a major, repeatable revenue engine
- 49:39 – 1:06:49
Cursor, Grok Build, and ‘harness + model’ co-design: the agentic AI arms race
The panel argues the tooling/runtime layer is now as important as the model itself. Gavin cites Cursor’s Composer 2.5 leap and Grok Build’s emergence as evidence that proprietary data, reinforcement learning, and agent runtimes can rapidly reposition winners at the frontier.
- •Composer 2.5 jump framed as Pareto-dominant progress
- •Proprietary coding tokens + RL as a fast path to step-function improvement
- •Grok Build introduces a runtime/harness (state, memory, integrations)
- •Claim: frontier competition is now model + environment co-development
- 1:06:49 – 1:11:17
Orbital compute and Starship economics: rapid reusability as the unlock
After weighing near-term earthbound compute growth, the group explores the feasibility of space-based data centers. Gavin explains rapid reusability as Starship’s defining technical hurdle and provides a timeline estimate for orbital compute, noting that GPUs have already operated in space today.
- •Rapid reusability vs reusability: multiple flights/day as the real breakthrough
- •Starship learning loop: failure as iteration fuel
- •GPU in space already demonstrated; radiation/launch constraints discussed
- •Gavin’s orbital compute timeline estimate: late 2028–2030 window
- 1:11:17 – 1:22:11
Nvidia blowout quarter, stock skepticism, and why chip share-loss stories persist
Nvidia posts extraordinary growth and capital return, yet the stock reaction reveals continued market doubt. Gavin argues cross-sectional AI valuations are inconsistent, challenges opaque ASIC performance claims, and points to Nvidia’s expanding CPU business as a major underappreciated signal.
- •Nvidia financials: massive revenue growth, margins, buybacks, dividend raise
- •Valuation mismatch across AI supply chain (chips vs power/cooling/optics)
- •Benchmark opacity: why some ASICs aren’t submitted to standardized tests
- •Nvidia CPU business targeting ~$20B as a meaningful new pillar
- 1:22:11 – 1:32:53
Macro warning lights: inflation, yields, debt spiral fears, and how to invest anyway
The conversation turns to oil-driven inflation, rising yields, and international bond stress—prompting Friedberg’s darker ‘credit crisis’ framing. Chamath advocates concentration in a few long-term winners, while Gavin balances rate risk with the strength of AI fundamentals and America’s energy advantages.
- •Oil and inflation expectations rising; yields pushing higher globally
- •Friedberg: debt dynamics and carry-trade unwinds as crisis catalysts
- •Chamath: own a handful of durable ‘future’ businesses; avoid broad speculation
- •Gavin: rates are a risk, but AI fundamentals and US energy position matter
- 1:32:53 – 1:41:59
US–China trip aftermath: no grand deal, chip sales debate, and avoiding the Thucydides Trap
The episode closes on the US–China diplomatic trip, with Friedberg seeing mostly symbolism and continued great-power tension. Chamath suggests real progress may have happened behind closed doors, while Gavin argues dialogue is inherently stabilizing and that controlled chip sales may reduce China’s incentive to build a separate ecosystem.
- •Friedberg: limited concrete outcomes; rivalry arc continues
- •Chamath: public optics vs private bargaining—possible ‘game board’ alignment
- •Gavin: selling deprecated GPUs may slow rival ecosystem formation
- •Shared emphasis: communication reduces escalation risk (Thucydides Trap)
