As aired
Prakash Narayanan introduces Justin McCarthy, founder and CEO of Diffusion, a company that helps large organizations build “software factories” in which people describe desired outcomes and AI agents implement and evaluate the resulting software; McCarthy previously co-founded StrongDM, where he spent a decade building cybersecurity infrastructure before forming its AI team in July 2025 with Jay Taylor and Navan Shahan. Nathan Labenz opens by disclosing that Diffusion is a sponsor of the Cognitive Revolution — that's how the two first connected — and asks about the service Diffusion is best known for: flying corporate leadership teams to Silicon Valley for an intensive week meant to get them to truly “grok” how to use AI.
McCarthy frames the company's name as the thesis: frontier-model capability is extremely concentrated (“in the data warehouse”), and the hard, unsolved problem is diffusing it into incumbent firms that don't have a rulebook for the transition. Asked by Prakash how to handle the common blocker of messy or missing data, McCarthy says “existence is the correct format” — don't wait for clean data, because working with it is exactly what surfaces the right evals and benchmarks. If a company genuinely lacks the data, he says, buy it, pointing to the well-known example of Spirit Airlines' bankruptcy: after the FTC blocked JetBlue's acquisition and Spirit went under, its operational data went to a bankruptcy auction where, per McCarthy and Prakash's account, a bidder identified as “Merkur” offered $7.5 million and Google won with a $10 million bid. McCarthy argues that kind of historical operational data — pilot load calculations, weight-and-balance tricks, and other tacit knowledge never written down on the public internet — is directly useful for building validation loops, even outside the airline itself.
On why AI initiatives stall inside companies, McCarthy points to a Diffusion principle he calls “make success inevitable”: engineer the agent's token/context environment — which files it reads first, how the mission branches — so that a plausible path through the information reliably leads it to the right answer, the way water finds the bottom of a valley. Unprepared, human-oriented codebases, by contrast, cause agents to load the wrong context. That environment design can itself be evaluated and automated with loops running on loops, he says, to the point that questions like whether a legal-team handoff to a paralegal is ready for automation can be answered objectively today.
Prakash raises Six Sigma and asks whether “five nines” of reliability is achievable with today's models. McCarthy answers that unreliable components can be composited into highly reliable systems — he's speaking from the sixth floor of a steel-and-concrete building made of imperfect parts, and invokes the Golden Gate Bridge — and notes that the previous “intelligence source,” humans, was never five-nines reliable either. He cautions against chasing nines indiscriminately: businesses should match the reliability bar to the actual stakes (an ACH deposit needs far more nines than a brand campaign) and avoid “gold-plating” agent output with unnecessary precision before they know what they actually want.
On interpretability and compliance — raised by Prakash via the example of a bank whose FICO-based lending decisions must not encode illegal redlining — McCarthy says the regulatory statute has to be treated as a first-class constraint (“your physics”) that the system is built directly around, with human managers, not the model, setting risk thresholds in legally untested territory. He also argues SOC 2 compliance, an accounting-derived process that has become “hyper-gameable,” should be renegotiated with auditors and regulators around new, agentic-loop-based checksums, drawing an analogy to how Walmart closes its books through cascading rounds of top-down “are you sure” verbal verification — the same pattern, he says, that lets Nvidia's Jensen Huang manage well over a hundred direct reports, and that Steve Yegge described in stories of presenting to Jeff Bezos, who could “drill down to your shoelaces.” McCarthy calls his own version of this “demand understanding.”
Asked when the ROI of AI transformation becomes undeniable, McCarthy insists on measuring revenue and market share, not cost savings, which he calls “too easy.” He cites the global shortfall in air conditioning — by his estimate only about a quarter of people on Earth who'd want AC have it — as an example of latent demand large industries could meet far faster than they assume. On budgeting, against Nathan's cited estimate of roughly 3% of human labor cost for ongoing inference, McCarthy says absolute spend may rise even as unit costs fall, because more revenue and value are being created (his back-of-envelope: CapEx/OpEx going from $1 to $1.50 alongside revenue going from $3 to $5). On the human side, responding to Prakash's question about “attention load” and rubber-stamp approvals people keep in the loop out of habit, McCarthy applies ordinary management-hierarchy discipline to agents — demanding the right “elevation” of detail rather than getting pulled into line-by-line minutiae — and describes checking in with a swarm of working agents around a normal human schedule: overnight work, a bedtime check-in, a morning brief, even a drive-time brief.
On competitive moats, McCarthy tells incumbents worried about both adjacent rivals (citing how Brex, Ramp, and Mercury keep matching each other's features) and AI-native startups to inventory what they actually have that a challenger can't get overnight: physical assets (“does Nathan have a lithium refinery? No.”), regulatory approvals, patents, brand affinity, and customer inertia. Closing on the design of the Silicon Valley training week itself, McCarthy tells Nathan — who is reconsidering his own past “overwhelm” style of AI evangelism — that Diffusion spends comparatively little time on inspiration and far more on “the how”: getting a real win to production and, ideally, showing tangible results within twenty-four hours, which is what actually converts skepticism into momentum.
Cameron Berg
Justin McCarthy