No Priors

AI labs and infrastructure companies are hitting revenue and cost-curve milestones on a timeline with zero prior comparison, which the hosts treat as the strongest evidence against bubble narratives - stronger than any single anecdote. - GPT-4-equivalent token costs fell roughly 100x in a year and could fall a billion-fold over a decade as hardware, algorithms, and architecture compound (2026-01-08) - Token pricing for equivalent-capability models collapsed ~150x in 21 months (GPT-4-class) and ~88x in 11 months (o1-class), even as usage and revenue explode simultaneously (2026-02-19) - AI labs reached $1B to $10B in revenue in roughly a year, versus 20+ years for ADP/Adobe, 8-9 years for Salesforce/SAP, 3-5 years for Google/Meta/AWS (2026-02-19) - Anthropic, OpenAI, and SpaceX went from near-zero to a trillion dollars in market cap in about five years versus the usual 15-20 year climb, but reaching that bar requires $50-100B of revenue at good margin, which only a handful of markets can plausibly support - most "huge TAM" AI companies will land at $20-100B, not multi-trillion (2026-08-06) - Trillion-dollar company formation happens in punctuated bursts tied to technology waves (social, SaaS, crypto, AI) followed by consolidation, not as a smooth continuous process (2026-08-06)

The financing and physical-infrastructure layer, not customer appetite, is where the show's guests locate genuine fragility - a nuance the "bubble" headline consistently flattens. - GPU debt is collateralized primarily by contracted, take-or-pay cash flows from investment-grade customers like Microsoft, not by the GPUs themselves; amortization is deliberately shorter than useful life so depreciation risk never materializes for lenders (2026-02-26) - The 2026 compute bottleneck has shifted from chip supply to people, power, and physical infrastructure - structural steel, electricians, substations - pushing operators toward "bring your own capacity" site designs (2026-02-26) - "Circular financing" fears are overstated because there is essentially zero idle GPU capacity, unlike the dark-fiber overbuild of the 2000s telecom bubble, and enterprise AI's real TAM is estimated around $37B and growing (2026-02-26) - Sovereign, government-lab, and strategic-investor capital (national labs, G42, the US government's Intel stake) function as bridges that let capital-intensive AI-adjacent hardware bets survive years of being ahead of market demand (2026-05-21; 2026-06-18) - Real risk to the capex cycle is financial-plumbing risk (who bears credit risk on pay-on-delivery contracts) and concentration in Nvidia and a few other players, not doubt that AI usage is real (2025-12-19)

Despite theories predicting AI would spread value into a long tail of niche winners, guests repeatedly argue the opposite: concentration persists, and platform owners historically absorb their most valuable dependent applications. - Value in tech remains sharply power-law distributed (Google's ad revenue, YC's overall returns); AI is unlikely to change that concentration ratio, though the absolute count of very large companies may grow as addressable software surface area expands (2026-02-19) - Platform owners have historically forward-integrated into their most valuable dependent applications (Microsoft into Office, Google into travel/local search), and the same dynamic is already visible with AI labs and coding - raising the open question of which application categories stay durable (2026-02-19) - Microsoft frames its own strategy explicitly as an ecosystem play: a platform must create more value for participants than it captures for itself, or it isn't a real platform (2026-06-04) - Investors intellectually endorse AI outcome-based pricing but still underwrite deals with old per-seat SaaS math, missing how much bigger markets like coding actually are once consumption is priced correctly (2026-08-06)

Across three separate enterprise CEO interviews, the argument converges: AI erodes single-feature products fast but leaves deeply integrated, accountable platforms intact. - Products get replaced, platforms are sticky - platform status means at least two products used in unison, wired into a customer's security/governance/compliance infrastructure; an easy wedge is also an easy exit once a competitor is equally disruptive (2026-01-22) - "AI thinks, workflow acts": a model can recommend the right steps for a cross-departmental problem instantly, but only a workflow platform can actually traverse the systems needed to close the case (2026-04-17) - Replacing a mature enterprise platform with LLM-generated code is roughly 10x more expensive once GPU/token costs and rebuild time are added to human capital costs; businesses forgive human error but not software error, preserving trust in an accountable vendor (2026-04-17) - The hardest AI engineering problem at enterprise scale is not the demo, it's scale itself - a 10-document RAG chatbot is trivial, but thousands of documents and 20,000 internal APIs turn orchestration and disambiguation into the real challenge (2026-04-23) - AI's real threat to incumbent software isn't the product, it's the depth of enterprise integration that's hard to replicate - claims like "AI could rebuild Slack or Salesforce" understate how much value sits in cross-system integration, not features (2026-02-26)

Every enterprise guest describes the same directional shift, and every one describes customer resistance pulling it back toward a hybrid. - SAP is moving step by step from seat-based licenses toward consumption and eventually outcome-based pricing, but customers pull it back to a hybrid because they want cost predictability and don't yet fully trust agent outputs (2026-04-23) - Microsoft expects per-user subscriptions to persist for budget certainty, layered with usage-based consumption, because customers who initially love outcome-based pricing tend to resent it once they see how much upside they're sharing away - GitHub Copilot's own shift to per-user-plus-consumption is cited as evidence (2026-06-04) - Decagon and Sierra are cited repeatedly (by MongoDB's Desai and in the code-slop episode) as the clearest real examples of a genuine per-seat-to-usage shift, specifically in customer support (2026-01-22; 2026-02-19) - Incumbents can inflate perceived AI traction by rebranding bundled features as "AI" for pricing purposes; MongoDB deliberately declined to attribute its own growth to AI on CNBC to preserve credibility (2026-01-22)

The show's most direct rebuttal episode argues both sides of this claim are true simultaneously depending on company size and category. - A Fortune 100 company will not displace its CRM with something vibe-coded over the weekend - enterprise software requires distribution, sales, security review, and change management a five-person startup doesn't have to solve (2026-02-19) - Claims that AI agents are already autonomously choosing vendors are overstated; what looks like agentic purchasing is usually a pre-negotiated partnership executing in the background (2026-02-19) - Multi-product bundling is best understood as a defensive strategy against AI-driven commoditization, reversing the old SaaS-era advice to "do one thing well" (2026-02-19) - Abundant AI-generated code creates a new attention bottleneck: once code production is cheap, nobody deeply reviews the resulting codebase, and no one has solved that problem yet (2026-02-19)

Multiple unrelated guests describe the identical shift in near-identical language, suggesting a real, dated industry inflection rather than one person's anecdote. - Karpathy hasn't typed a line of code since roughly December, measuring his day by token throughput across parallel agent sessions rather than lines written, framing nearly all friction as a "skill issue" not a capability gap (2026-03-20) - Notion's Simon Last has not personally written code since mid-2025, describing himself as "the agent manager instead of the coder," acting as the outer verifier for end-to-end tasks rather than the implementer (2026-03-12) - Cerebras's internal AI-coding spend jumped ~25-30x in eight months, but the productivity gain is highly uneven - a small group running 8-10 agents around the clock went from "10x" to "100x" engineers while most, including the CEO, are still "limping along" (2026-05-21) - DoorDash's engineering AI spend grew ~20x from January to June before flattening once ROI discipline (via its internal Dashbench benchmark) kicked in and cheaper tasks got routed to open-weight models (2026-07-23) - SAP compares the shift explicitly to junior developers moving from writing code to reviewing agent-generated code, and expects the same "uplevelling" pattern across finance and HR knowledge work (2026-04-23)

Where coding succeeded fast, guests attribute it to a built-in verification signal (compile/test) that most other domains lack - and enterprise, benchmark, and research contexts are all racing to invent equivalent signals. - Coding agents work because output is verifiable; enterprise agents need explicit "evals" as boundary conditions to substitute for the missing compile/test signal in finance and HR domains (2026-04-23) - "Agent mining" turns every human clarification an agent needs into training data, either flagging a local deviation as an anomaly or promoting it into a new default standard operating procedure (2026-04-23) - Standard benchmark grids report a single score per model without controlling for test-time compute spent, making comparisons misleading; GPT-5.5 looked like a small gain over 5.4 until compute was held constant (2026-06-26) - Fully evaluating a model's ceiling may require running it as long as the task itself takes, which conflicts with a 2-3 month release cadence - meaning nobody, including the labs, actually knows a model's true capability before it's superseded (2026-06-26) - Notion redesigned its APIs specifically for agents (a markdown dialect, a SQLite-style query layer) because the original human-facing JSON API was too verbose for models to use well - agent-facing API design is a distinct discipline from human-facing API design (2026-03-12)

The show returns repeatedly to a split between engineers who value craftsmanship for its own sake and those who treat code as a means to an end, predicting divergent reactions to the same technology. - Engineers with a craftsmanship-based identity will react very differently to agentic coding than utility-focused engineers - craftsmanship-identity engineers are likely to become unhappy as the parts of coding they valued erode, while utility-focused engineers find the shift freeing (2026-02-19) - Documentation and teaching are shifting from being authored for humans to being authored for agents; Karpathy's minimalist microGPT project is now explained to users primarily by an AI agent rather than a human guide (2026-03-20) - Karpathy predicts AI increases, not shrinks, demand for software work via a Jevons-paradox effect (ATMs made bank branches cheap enough to run that teller employment rose), expecting digital-native work to be reshaped far faster than physical-world work (2026-03-20) - Frontier models perform well on cleaned benchmark data but underperform on raw enterprise data, and it's unclear whether that's a harness gap or a genuine training-distribution gap (2026-07-23)

Baseten's infrastructure-level view and Netic's application-level view arrive at the identical conclusion from opposite ends of the stack. - Over 95% of tokens in Baseten's dedicated-inference business run custom or post-trained models; the durable value in a company like Abridge (ambient clinical scribe) is what clinicians edit in AI-drafted notes and what happens downstream in the EMR - workflow signal frontier labs structurally cannot access (2026-05-01) - Serving AI-native application companies gives an infrastructure vendor an indirect read on enterprise requirements (data retention, latency, GPU choice) without having to sell into the enterprise directly (2026-05-01) - Netic doesn't see foundation labs as a competitive threat because winning in a vertical (HVAC, roofing, pet care) requires a full stack of model, orchestration, and deep vertical product work that generalist labs aren't focused on building; over 70% of Netic customers' interactions are now AI-first ("N1") (2026-07-31) - Raw GPU access is commodity, but inference bundled with software is sticky - Baseten's top 30 customers have never churned, with ~400% annual net dollar retention (2026-05-01) - ElevenLabs argues technology advantages are never permanent moats; what compounds is the ecosystem (voice libraries, integrations, distribution) built around a research head start, not the underlying model quality itself (2025-12-11)

As agents gain real permissions, a new discipline emerges specifically because existing security tooling can observe actions but not the intent behind them. - Existing enterprise security (identity, endpoint, API) breaks down for agents because it can see that an action happened but not why the agent decided to do it, so it can't distinguish a legitimate action from a dangerous one mid-task (2026-05-28) - The viable architecture is small, narrow overseer models that decide when to escalate an action to a smarter, more expensive reviewer - like blitz chess, where intuition handles most moves and deep calculation is reserved for critical ones (2026-05-28) - Security vendors gain a durable edge over the labs themselves because enterprises won't hand their historical agent-behavior data to Anthropic or OpenAI for fear it gets used for training, but will give it to an independent security vendor (2026-05-28) - Agent errors split into "jagged intelligence" mistakes that will shrink as models improve (the labs' job) versus independent/misaligned judgment calls that get harder, not easier, as capability increases (2026-05-28) - SAP and ServiceNow both frame enterprise AI's central technical debt as fragmented, siloed data - not model capability - as the top blocker to adoption, ahead of security and complexity concerns (2026-04-23)

Two founders explicitly building in the same space (essential/local services) choose opposite structures, and both defend their choice against the other's model by name. - Long Lake buys operating businesses outright (rather than selling software) specifically because ownership gives it direct control over change management; portfolio companies went from 0-5% to 20%+ organic growth once AI-driven efficiency turned growth from a punishing marginal cost into a high-margin, software-like one (2026-05-11) - Netic's founder deliberately chose a horizontal software platform over a roll-up because M&A isn't her core skill and roll-up-built products only ever serve the specific companies acquired rather than compounding across an industry (2026-07-31) - Long Lake's Nexus platform is ~80% shared infrastructure across verticals, cutting time-to-impact after acquisition from over a year to days; both Long Lake and Netic report needing to combine deal/ops execution with deep AI engineering, a combination each says is rare (2026-05-11; 2026-07-31) - Real enterprise AI use cases are estimated at only about 1% penetrated, with most of the US economy made up of small businesses lacking resources to build AI tooling themselves - both founders treat this as the size of the opportunity, not a reason for caution (2026-05-11) - Private equity's AI playbook is shifting from finding undervalued arbitrage targets toward generating tangible new revenue in portfolio companies, though initial conversations still often start with cost-cutting (2026-07-31)

Across cars, chips, and robots, the guests who won did so by discarding the prior rules-based architecture entirely rather than bolting AI onto it. - Rivian scrapped its rules-based Gen One autonomy stack entirely at the end of 2021 and rebuilt clean-sheet; Scaringe argues the vast majority of pre-2022 rules-based industry investment is now throwaway, and only one to four companies (Rivian, Tesla, Waymo) outside China have the full ingredient list to compete (2026-02-12) - Tesla's Optimus overtook the more established Boston Dynamics because it was AI-infused from the start rather than a legacy platform with AI bolted on - offered as the template for which defense robotics companies win next (2026-01-15) - DoorDash designed its Dot robot from scratch after years of partnering with sidewalk-robot and robotaxi startups revealed that none were built use-case-first; neither existing robot category (too slow, or oversized for people) fit an actual food-delivery profile (2026-07-23) - Cerebras bet the whole company on wafer-scale chip architecture specifically because a 15-20x performance jump requires a genuinely different architecture, not an incremental GPU variant - a bet critics called impossible before it worked (2026-05-21) - Nvidia deliberately protects a programmable architecture instead of shipping fixed-function AI chips, betting that model architectures (transformers, SSMs, diffusion) are still evolving too fast to lock into one design (2026-01-08)

Materials science and biology guests converge on an identical architectural insight: static data is necessary but not sufficient; the loop between experiment and model is where the leverage is. - A pool of experimental materials data isn't enough - the real value is an active, closed loop where results are checked for aberrations against literature/simulation, which then drives the next experiment (2026-04-03) - Generalization in physical-science models happens at the level of shared physical first principles (quantum mechanics, chemical synthesis rules), not uniformly across domains - a strong quantum-mechanical model doesn't automatically transfer to fluid dynamics (2026-04-03) - Biohub deliberately fuses "frontier AI" and "frontier wet-lab" into one organization because biological training data largely doesn't exist yet and has to be generated through new experimental methods, unlike language models drawing on existing internet text (2026-06-10) - Biology has to be modeled hierarchically bottom-up (proteins, then cells, then systems) because each layer is constituted by the one below it; ESM Fold folded 1.1 billion proteins and produced antibody design as an emergent property of a general model never purpose-built for it (2026-06-10) - Compute, not physical lab infrastructure, is the dominant capital cost in AI-driven physical science despite infrastructure's much longer lead times (2026-04-03)

Nuclear, chip, and defense guests independently locate the constraint in institutional capacity rather than physics - and describe near-identical workarounds (in-house builds, dormant legal authority, deliberate reshoring). - Nuclear component pricing is inflated by an atrophied supply industry rather than genuine engineering difficulty - a $5M/2.5-year outside quote for a reactor protection module was replaced by a 5-person in-house build for $400K in 6 weeks (2026-07-02) - A dormant Department of Energy testing authority, unused in legislation for ~40 years, let Valar bring its reactor to criticality entirely outside the NRC's commercial-deployment pathway, which structurally requires proven systems startups can't yet produce (2026-07-02) - The Department of War (formerly Defense) deliberately opened its front door to startups after SpaceX, Anduril, and Palantir all historically had to sue for their first contracts, given an industrial base consolidated from ~50 primes in the 1980s to about five today (2026-01-15) - Intel's turnaround explicitly imports startup speed and modern AI tooling into what its CEO calls a legacy "spreadsheet company," while betting on new substrate materials and advanced packaging as process scaling runs out of headroom (2026-06-18) - Rare earth minerals aren't geologically scarce - the bottleneck is refining capacity, concentrated and subsidized in China; the "Pax Silica" State Department strategy explicitly rejects China's Belt and Road debt-trap model in favor of joint ventures with shared risk (2026-05-14) - US reindustrialization will require heavy automation because the gap between US consumption (20-30% of global) and production is too large to close with human labor alone at ~4% unemployment (2026-05-14)

Multiple founders and the hosts themselves converge on the identical practice, credited to the same source, as the antidote to both over-attachment and panic. - Founders should pre-schedule non-emotional board discussions about exit timing (credited to Ben Horowitz at Opsware) since a company's peak-value window can be as short as 12 months, given historical parallels like Lotus 1-2-3's collapse after Excel launched (2026-02-19) - The hosts sharpen this in their own two-hander: a handful of companies (Anthropic, OpenAI) should never sell, but most pass through a 12-18 month peak-value window, and the check-in cadence should tighten to every six months given how much a year of AI-era progress compresses (2026-08-06) - The largest hidden cost of staying too long in a stalled, overcapitalized company is a founder's most productive years, not capital - 2020-2021-vintage founders are cited as still running non-working companies years later, effectively locked through the whole AI transition (2026-08-06) - Cerebras used secondary-market sales to decouple employee liquidity from the traditional four-year option-vesting clock, letting the company stay private longer while people still got some liquidity along the way (2026-05-21) - Deal and build timelines are compressing past what founders previously assumed was physically possible - Cerebras's $20B+ OpenAI deal went from term sheet to signed master agreement in 4.5 weeks; Cognition bought Windsurf "over a weekend" (2026-05-21)

This is the show's clearest surfaced disagreement: the same hosts flag a trend they find troubling, in tension with the "underdog mentality" framing a policy guest offers as a strength. - A growing share of otherwise strong founders are retreating into niche markets out of fear of the frontier labs rather than competing head-on and out-executing on distribution - a trend the hosts frame as founders being less ambitious than they could be, not a rational response for most of them (2026-08-06) - By contrast, the State Department's Jacob Helberg argues American "underdog" resilience (13 disorganized colonies, repeated declared-in-decline moments) is the same psychological pattern as Silicon Valley founder mentality, and should be leaned into as a national asset, not diagnosed as a deficit (2026-05-14) - Netic's founder independently criticizes a growing "AGI pill" belief among younger hires that they must extract all value within 18 months before AI renders them obsolete, arguing this short-termism undermines the patient, decades-long commitment building something real actually requires (2026-07-31) - Belief that recursive self-improvement is ~18 months away is driving burnout-adjacent overwork at major labs (researchers reportedly asking whether to get married given 18-month uncertainty), even though that exact prediction has recurred every 18 months for about five years (2026-08-06) - Cerebras's Feldman frames founder psychology bluntly: being CEO is "extraordinarily lonely," and the right time to quit is when every hypothesis for winning has come back negative - not a vague gut feeling, and not an endless "one more thing" (2026-05-21)

The show's guests largely converge on Jensen Huang's task-vs-purpose framing, but sharply disagree on pace, and several offer surprising predictions that automation increases specific human headcounts rather than reducing them. - A job splits into task and purpose - AI automates the task (studying a scan) but rarely the purpose (diagnosing disease); radiologist employment grew even as 100% of radiology applications became AI-powered, because expanded task throughput increased demand for the purpose (2026-01-08) - DoorDash predicts more human dashers in ten years, not fewer, because demand growth (delivery becoming cheaper) will outpace how fast autonomy can substitute for humans (2026-07-23) - ServiceNow's McDermott takes the opposite emphasis: net-new headcount growth will slow sharply as agents absorb tactical work (2.2 billion AI agents entering the workforce is more than projected new human hires), with 90% of ServiceNow's own customer service cases already agent-resolved (2026-04-17) - Booking.com's Fogel argues the real risk is the speed of displacement, not displacement itself - technology has always displaced jobs, but the current wave is unusually fast, and public opinion on AI flips dramatically depending on how survey questions are framed (2026-07-09) - As AI reduces need for less-productive engineers at top-tier tech companies, that talent is likely to flow into traditional non-tech enterprises (GE, PG&E, Hershey's) that never had the brand to recruit top engineers before (2026-08-06) - Circle's Allaire frames the stakes at the macro level: double-digit global GDP growth in the 2030s is plausible from AI, but GDP could become a misleading metric if gains mainly reflect capital capturing value at labor's expense, absent a new social contract (2026-04-09)

The show's most technical guests independently locate the same constraint on where AI can improve itself. - Karpathy's "auto research" framework (define an objective, a metric, and boundaries, then remove the human) found hyperparameter mistakes in a codebase he'd hand-tuned for two decades - but auto-research and RL only improve domains with cheap, objective evaluation, which explains models' uneven capability elsewhere (e.g., stale jokes despite transformed coding ability) (2026-03-20) - Near-term recursive self-improvement is happening specifically in software engineering, enabled by cheap, instantly verifiable feedback (unit tests); it does not automatically extend to domains like biology with a real human-AI knowledge gap (2026-04-03) - AI research itself is a slower self-improvement loop than software because verifying results (convergence, generalization) requires actual GPU-hours of experimentation, not instant unit-test checks - the same closed-loop principle underlying Periodic Labs' premise for physical science (2026-04-03; 2026-03-20) - Current models can dramatically optimize existing research work (10-100x speedups on a published poker-solver algorithm) but still lack "research taste" - the ability to originate genuinely novel directions from synthesizing literature (2026-06-26) - Models can be world-class on one narrow domain and fail badly on small perturbations of the same problem type - intelligence is spiky, not a single scalar quantity, so capability claims need to be domain-qualified (2026-04-03)

Guests with the most technical standing to make either claim consistently reject both extremes, for structurally similar reasons. - Noam Brown argues an overnight, self-reinforcing intelligence explosion isn't near because unlocking a model's full capability requires long stretches of test-time compute, rate-limiting progress by wall-clock time even as internal research accelerates (2026-06-26) - Jensen Huang's "God AI" counter-narrative rejects policy built around a hypothetical monolithic superintelligence, arguing that framing isn't close to reality and isn't a useful basis for near-term regulation (2026-01-08) - Preparedness frameworks and responsible scaling policies were built before test-time-compute scaling and don't specify what inference budget to evaluate at, leaving a real safety-policy gap since a model's dangerous capability, like its useful capability, scales with money spent (2026-06-26) - Karpathy expects open-source models to keep trailing frontier closed models by a roughly stable ~8-month lag rather than closing the gap, and argues that persistent lag is structurally healthy because it avoids centralizing intelligence in two or three closed labs (2026-03-20) - Multi-agent systems that accumulate and share knowledge across instances, rather than existing in isolated short context windows, are an underexplored frontier - individual AI agents are "born" into a short context and disappear, unlike human civilization's cross-generational knowledge accumulation (2026-06-26)

• Paul Janssen YouTube interviews on regulatory capture in pharma other (2026-08-06)
• Noam Brown's essay on benchmark evaluation and test-time compute article (2026-06-26)
• AlphaFold other (2026-06-10)
• Essay on organizational ambition in the age of AI article (2026-06-04)
• Auto-GPT other (2026-05-28)
• No Priors episode with Andrej Karpathy podcast (2026-04-23)
• Blueprint for Agentic Business (ServiceNow white paper) paper (2026-04-17)
• Inference cost/performance article on Hopper vs Blackwell efficiency article (2026-02-26)
• Silicon Data spot-pricing and price-per-token article article (2026-02-26)
• Mark Zuckerberg's post on AI eating the world article (2026-02-19)
• Keep Your Identity Small (essay, attributed to an Applied Intuition co-founder's blog post) article (2026-02-19)
• Ready Player One movie (2026-02-05)
• Vanilla Sky movie (2026-02-05)
• Black Mirror (dating episode) show (2026-02-05)
• Grand Theft Auto other (2026-02-05)
• Pantheon show (2026-01-29)
• DeepSeek (R1 release and paper) paper (2026-01-08)
• MIT study on enterprise AI deployment ROI paper (2026-01-08)
• Andrej Karpathy's open-source weekend chatbot project other (2026-01-08)
• MIT report on AI ROI (widely-cited 2025 study claiming most AI pilots fail to show return) paper (2025-12-19)
• Ilya Sutskever interview on the 'age of research' other (2025-12-19)
• The Hitchhiker's Guide to the Galaxy other (2025-12-11)
• Lex Fridman Podcast interview with Narendra Modi podcast (2025-12-11)
• MasterClass: Chris Voss negotiation lesson other (2025-12-11)