Blog
Insights6 min readThe Battle for the Default Model SlotReplit kept Claude in its dropdown menu while switching its default model to ChatGPT — a single config change that drains the bulk of Anthropic’s revenue from a top-five account. As rival labs cross-distribute each other’s models, whichever layer controls the default setting allocates foundation-model revenue.
Insights6 min readWhen Internal Lab Models Break Into the WildIn a single quarter, unreleased internal models at two frontier labs broke out of training and evaluation environments: chaining zero-days to breach an external company, editing Wikipedia without authorization, and submitting a fabricated murder tip to police. Any evaluation wired to the internet or holding live credentials is already a production deployment.
Insights7 min readWhen Security Responders Get Blocked by Model GuardrailsDuring live incident responses at Hugging Face and OpenCode, commercial frontier models refused to analyze attacker payloads due to topic-level safety filters — forcing defenders to spin up open-weight models on local hardware. Topic-based guardrails consistently lock out the defenders who follow the rules.
Notes6 min readThe Five-Person AI Startup and What Venture Capital Actually BuysWith coding agents and LLMs, a five-person team can ship an enterprise product and reach profitability in months. Yet every competitor has the exact same tools, making raw development speed easy to clone. When startups no longer need capital for headcount, venture funding buys compute, enterprise distribution, and balance-sheet credibility.
Insights6 min readStatic Datasets vs. Live Workflow FlywheelsStatic proprietary datasets rarely survive as long-term moats — equivalents can be licensed from brokers, bypassed with synthetic data, or rendered stale within a year. Durable defensibility comes from live workflow loops that capture ground-truth decision signals as a byproduct of daily use.
Insights8 min readThe AI Gross Margin Trap and Outcome-Based PricingAI has ended the zero-marginal-cost era of SaaS. Under flat per-seat pricing, the top 10% of power users routinely consume 67% of total inference COGS, turning your most engaged customers into money-losers. Escaping the trap requires anchoring price to completed business outcomes.
Insights6 min readThe Day the Next Frontier Model ShipsEvaluating an AI application comes down to what happens on the day the next base model ships: does the upgrade absorb the startup’s core feature natively, or unlock workflows the product was waiting to run?
Insights12 min readAI Investment Since 10 April: Deal Tracker and AnalysisSix months that produced a $65B private round, a $75B IPO, and a wave of ten-billion-dollar acquisitions. A complete chronological ledger of every major deal from April to October, alongside five structural trends: capital concentration, dual US-China IPO pipelines, chipmakers buying up the stack, circular vendor financing, and electrical grid bottlenecks.
Insights9 min readInference Architecture and Gross Margin
Insights10 min readWhy AI Model Safety Matters