Erik Torenberg

Erik Torenberg × Benedict Evans

Agentic coding has crossed unmistakable product-market fit, with customers pulling products from vendors’ hands.Foundation models appear headed toward commodity economics absent durable differentiation, while today’s token scarcity meets $1 trillion–$2 trillion of capex and “100x, 200x” efficiency gains.Cheaper development should create more software, but the unresolved question is which SaaS incumbents survive and whether models capture infrastructure-like returns.

Erik Torenberg × Nathan Labenz × Logan Kilpatrick × Tulsee Doshi

Gemini 3.5 Flash leads with Tulsee citing roughly 280 tokens per second, about three times faster than other large models and significantly cheaper, prioritizing intelligence per dollar and second.Google is shifting its AI stack toward Antigravity’s shared agent harness, while preserving harness diversity; the payoff is a cross-product feedback loop, but co-optimization could still raise switching costs and pricing power.

Alex Rampell × Erik Torenberg × Mike Cannon-Brookes

AI is repricing SaaS before proving universal impairment, with Zendesk-like seat models exposed to agent substitution while Workday’s employee-based pricing and Adobe’s middle position may be underappreciated.Atlassian’s three great quarters, accumulated process knowledge, Teamwork Graph and extensibility strategy could make core systems stickier, but value depends on fair pricing and product design that earns trust as agents enter workflows.

Benedict Evans × Erik Torenberg

Generative AI may match the internet or smartphones in scale, but uneven adoption—ChatGPT’s 800 or 900 million weekly users with only about 5% paying—makes workflow integration, not model novelty, the central commercial test.Specialized products that encode validation and institutional knowledge may capture value above increasingly comparable models, while OpenAI’s distribution remains a fragile moat and falling compute costs, overbuilding and uncertain capability keep the bubble’s timing and infrastructure demand unresolved.

Erik Torenberg × Marc Andreessen × Ben Horowitz

AI need not match Beethoven to transform productivity: clearing 99.99% of humanity at intelligence and creativity could unlock recombination across domains.Current models already show commercially useful theory of mind in Socratic dialogue and simulated focus groups, though intelligence alone does not confer leadership or human connection.Demand, talent, and chip shortages are attracting supply, while the still-unformed interface and US–China robotics race leave the platform outcome open.

Dylan Patel × Erin Price-Wright × Guido Appenzeller × Erik Torenberg

GPT-5’s router is an economic release, directing simple queries to mini models while reserving “ungodly amounts of compute” for transactions OpenAI could monetize through agentic commerce.Flat-rate subscriptions face heavy-user losses, while custom silicon threatens Nvidia mainly if demand stays concentrated among hyperscalers; powered sites, grid equipment, and Intel’s capital needs remain near-term constraints on broader deployment.

Erik Torenberg × Balaji Srinivasan

Newspaper revenue fell from roughly $70 billion around 2000 as Google and Facebook rose, helping turn a former tech ally into an adversarial media institution after 2013.Srinivasan’s state-versus-network framework favors direct distribution, founding creators, and a cryptographic ledger of record, while jurisdictional competition and the unresolved economics of replacing legacy reporting remain the key tests.

Erik Torenberg × Steven Sinofsky

Apple’s China advantage rests on accumulated manufacturing expertise and line-level knowledge transfer, turning a capability moat into strategic concentration risk as COVID exposed global supply-chain single points of failure.Apple’s AI retreat may reflect a return to its “first integrator” model, while edge, privacy, and inference economics could support several large platforms rather than a winner-take-all market.

Vincent Weisser × Johannes Hagemann × Erik Torenberg × Nathan Labenz

Prime Intellect is building an asset-light marketplace across fragmented clouds and data centers, while Intellect-1 demonstrated 10-billion-parameter distributed training using DiLoCo and 8-bit updates to cut communication roughly 400x.R1-style reinforcement learning could improve decentralized scaling by shifting compute toward inference-heavy rollouts, but efficient participation beyond about 16 workers, fault tolerance, safety, governance, and the planned tokenized foundation remain unresolved.