Stratechery's five frontier overhangs; Warp's factory that ships 2,000 pull requests a month

Stratechery: the five overhangs behind the call to pace the frontier

"Frontier Overhangs"

1. Those problems have names: capability, product, pricing, capital and safety. Each is a burden created by models getting better faster than the business around them can absorb. The capability overhang is that the models already do more than customers will pay frontier prices for. The product overhang is that the labs have to own a user touchpoint before their capability advantage erodes. The pricing overhang is that they can charge well above the cost of serving because demand exceeds compute. The capital overhang is that the buildout depends on money that has to be repaid out of revenue still to arrive. Slowing model improvement relieves all four, with safety as the fifth and the one argued in public. Thompson's sharpest observation is about who is asking: the call to pace the frontier arrived at the moment the frontier was, for the first time in a while, being set by OpenAI. Anthropic is fine with Anthropic in the lead, and a competitor in the lead is where an argument for government intervention starts. 1
2. Customers have already stopped paying for capability alone. Microsoft's Copilot Cowork now offers a model picker covering OpenAI's GPT models and Anthropic's Opus 5, Claude Sonnet 5 and Claude Fable 5.1, so the model and the harness can be chosen separately. 2 Anthropic's own retreat supplies the second piece of evidence. Thompson predicted that enterprises would hand over their data for access to Anthropic's best models; according to his account, usage stayed low until Fable 5.1 dropped the data-retention requirement. He reads both through the cycle Clayton Christensen set out in The Innovator's Solution: while performance falls short of what customers need, interdependent designs win, and once capabilities are good enough for everyone, modular ones do. The product side runs on the same logic. Meta's Muse is powered by Muse Spark 1.3, which is behind the state of the art, and Thompson still calls it the best personal agent he has tried. A personal agent is much stickier than a chatbot: your code sits in a repository a rival's harness can be pointed at, while your family calendar lives inside one product. 1
3. The money explains the impatience, and the security argument is the one he takes seriously. On price, OpenAI and Anthropic charge what they do because compute is scarce, and much of their compute goes into training and research instead of serving users, so a slower frontier would free capacity they could sell. On capital, the Financial Times reports that Anthropic has told backers it will be profitable this quarter on adjusted operating income that strips out stock-based compensation, with gross margins above 80% before revenue-sharing and training costs; Thompson flags the training exclusion as the number that matters, then adds a correction that he has since been told Anthropic is profitable including training costs. Either way, the labs and the hyperscalers need AI revenue to rise sharply before the capital funding the buildout runs out. On safety, he separates the concrete risk from the alignment debate: an attacker's automated effort has positive expected value and a defender's has negative, models with weights anyone can download are already capable of attacks, and a slower frontier widens the window in which those attacks succeed. 1

Lenny's Newsletter: Warp's software factory, and the numbers it runs on

"How Warp ships 2,000 PRs a month with AI factories | Zach Lloyd (CEO, Warp)"

1. A factory, in Warp's sense, is the whole development lifecycle running in the cloud and in public. Someone types a request into a Slack channel and tags Wilson, the factory, which triages it, opens a Linear issue, implements the change, creates the pull request, runs computer-use quality checks that record a video of the finished feature, and merges it. Work also starts on its own: crash reports from Sentry feed the same pipeline. Everything the factory is made of — its repositories, its configurations and its agents — is defined in code, and every task runs in a shared channel, so a junior engineer can watch how a senior one steers the agent. 3
2. Its headline measure is human effort per pull request, and the bottleneck it exposes is human review. Kickoff to pull request averages 35 minutes; the first human review arrives three and a half hours later, and the team shipped more than 2,000 pull requests last month. Warp counts "human interactions per pull request" — every Slack reprompt, every clarification on a ticket, every correction in code review — on the argument that a person who has to steer the agent is the real limit on how much the factory produces. Every pull request still gets a human review, though the rule that someone other than the person who prompted the agent must do it is gone. Spending is tracked the same way, per pull request, and the largest lever on it was model configuration, with context management a distant second. 3
3. Three practices are worth copying whether or not you build a factory. Warp scores every agent run with an AI judge kept deliberately mid-tier so that scoring costs less than it saves; one failure mode it watches for is redundant tests, which agents write to confirm current behaviour instead of guarding against future regressions. Once failures accumulate into a real sample, roughly 20 to 25 runs, an observer agent proposes changes to the factory's own agent definitions, which is possible because those definitions are code. For model choice, Warp replays its past tasks under different configurations and plots cost against quality, using the company's own workload where a public benchmark such as SWE-bench measures generic tasks. Lloyd's caveat is that a faster factory makes building cheaper, while deciding what deserves to be built still comes from user interviews, design sessions and product judgment. 3

Where to start