Engineers Become Sorcerers: OpenAI's Sherwin Wu on the Next 12 to 24 Months

Wu says that as a manager he finds AI tools easier than hand-coding, and that his own and other engineering managers' code is written by Codex. OpenAI cannot measure exactly how much code is AI-written, but he estimates close to 100% is generated by AI first. What the company does track is that 95% of engineers use Codex and 100% of pull requests are reviewed by it.

He adds that engineers who use Codex heavily open about 70% more PRs than those who do not, and that the difference has probably grown since he last checked. Some engineers still trust the tool less, but he says people's bar for how much to trust the model keeps rising. He repeats Kevin Weil's line that this is the worst the models will ever be.

Pressed on whether 100% of production code is AI-written, Wu says he would not claim that, since attribution is hard. He does say that almost every engineer uses Codex for almost every task, so the vast majority of code was probably authored by AI first.

Wu says the engineer's job is likely to change substantially over one to two years, and that engineers feel they are in a rare moment when they can define the standards themselves. Individual contributors are becoming something like tech leads, with many on his team keeping 10 to 20 parallel Codex threads going, checking in and steering them. The work shifts from writing code to directing it.

He frames this through *Structure and Interpretation of Computer Programs*, the 1980 MIT textbook known as the wizard book, which describes programming as sorcery with languages as incantations. In his view the current wave makes the metaphor literal, since you tell Codex or Cursor what you want and it goes and does it. He prefers the analogy of the sorcerer's apprentice, where the power is real but things flood if you walk away.

That is why skill and seniority still matter: with 20 threads running, you have to keep the models from going off the rails. Lenny offers the genie-and-wishes and monkey's-paw versions of the same idea, in which a vague request comes back with unwanted side effects. Wu says it is also what makes the job fun, because it feels like casting spells.

Lenny asks about the stress of watching agents stall. Wu says it happens constantly, and describes an internal OpenAI team that is deliberately maintaining a 100% Codex-written codebase with no escape hatch back to hand-coding. They run into the problems people describe, and he says the team will publish a blog post on the lessons.

One finding is that when a coding agent does not do what you want, the cause is usually missing or underspecified context. The fix is to encode tribal knowledge into the repository through comments, code structure, markdown files and skills, so the model can do the task next time.

On code review, Wu says the things OpenAI hands to models first are the boring parts of engineering. Codex reviews every PR, and he says that cuts a review from 10 to 15 minutes to sometimes two or three. CI, lint fixes and the push-to-production process are also heavily automated.

He stresses that attention does not go to zero: people still look at their PRs, but with perhaps 30% of the attention rather than 100%. OpenAI mostly uses its own models for review, including internal variants that offer a different perspective, and is careful about which changes get only a Codex review.

Wu says managing has changed less than engineering, since there is no Codex for managers yet. But he sees AI strongly empowering high-agency people, which widens the spread of productivity within teams. He has always spent most of his time with top performers, and thinks that is even more important now.

He also expects managers to oversee much larger teams, as engineers do with 20 to 30 Codex threads, rather than the current norm of six to eight reports. As an example, he uses ChatGPT connected to GitHub, Notion and Google Docs to summarize a year of someone's work for performance reviews. Lenny relays Marc Andreessen's line that AI makes good people better and great people exceptional.

Wu's concrete rule is to spend more than 50% of his time on roughly the top 10% of performers. He borrows the surgeon image from *The Mythical Man-Month*: the engineer is the surgeon and the manager makes sure the right scalpel is already in hand. In an AI world, he says, bottlenecks are organizational and process-driven, so unblocking matters more.

Lenny suggests asking an AI with company knowledge to list active blockers and anticipate next month's. Wu says he has not tried it, but likes the idea.

Asked what people are not pricing in, Wu points to the second- and third-order effects of the one-person billion-dollar startup. If one person can do so much, starting companies in general gets easier, so he expects a boom of small, vertical-focused startups. To support one giant solo company there could be a hundred small startups building bespoke software.

He pictures a golden age of B2B SaaS, with perhaps tens of thousands of $10 million businesses, which would be great for individuals. The third-order effect is on venture capital: venture-scale returns could shrink if many outcomes are in the $10 to $50 million range, while a handful of large platforms support them.

Lenny is bearish on a solo billion-dollar company because of support load. Wu answers that other tiny companies could build tailored support software for podcasts and newsletters that Lenny would buy, so the company stays small while the ecosystem does the work. Lenny adds that distribution may become the fourth-order effect, since so many things compete for attention.

Wu says he has no hard numbers, but would not be surprised if many corporate AI deployments have negative ROI. One reason is that people in Silicon Valley live in a bubble, while most workers elsewhere are not following model releases and ask the tools very basic questions. Employees often feel AI is being forced on them.

The deployments that work combine top-down buy-in, where executives buy tools and support them, with bottom-up adoption from employees who are excited and share best practices. Work differs across engineering, finance, operations and sales, so the last-mile details have to be worked out by the people doing it. At OpenAI, adoption took off when employees could apply Codex to their own work.

He recommends a full-time internal tiger team that explores capabilities, applies them to specific workflows and spreads knowledge. Because many companies have no software engineers, he says the best members are often technical-adjacent people, such as an operations lead who loves tools and is an Excel wizard. In his words, find the high performers in AI adoption and empower them.

Wu says listening to customers is not always the right strategy in AI, because the field disrupts itself so fast, especially in tooling. He quotes a phrase from Nicholas, founder of Fintool: "the models will eat your scaffolding for breakfast." Agent frameworks and vector stores, which looked central after ChatGPT launched, are examples of scaffolding that has become less essential.

As models improved, he says a better approach often became giving the model search tools, which could be files on a file system or skills and agents.md files, rather than elaborate embedding pipelines. Vector stores still have a place, but customers who only asked for a better vector store or agent framework would have led you to a local maximum. Today's fashionable skills-based context management might also be replaced.

Wu says OpenAI's API team has been guilty of this too, and that everyone is learning a version of the bitter lesson. His advice is to build for where the models are going: the startups he has seen do best build products that work about 80% of the way today, then suddenly click when a new model arrives.

Looking ahead, Wu points to the length of task models can complete coherently. Citing the METR benchmark, he says frontier models are at multi-hour software tasks about half the time and just under an hour at 80%. Products today are mostly tuned for minutes, but within 12 to 18 months he expects multi-hour and eventually day-long tasks that you dispatch and return to.

He also expects significant gains in multimodal models, mainly audio and native speech-to-speech models. He calls audio hugely underrated in business settings, since so many services and operations run through conversation while most attention goes to text and code.

Wu is also bullish on business process automation. Software engineering is open-ended knowledge work, whereas much of the economy is repeatable processes with standard operating procedures, like support lines or utilities. He wants AI applied there with high determinism and integration into enterprise data and systems, and says the opportunity is bigger than people on X assume.

On the common worry about OpenAI squashing startups, Wu says the market is so big that founders should not overthink where the labs are going. Every startup he has seen fizzle did so because customers did not love what it built, not because a lab crushed it. He says even in competitive spaces like coding, a product people love, such as Cursor, can become huge.

He says OpenAI regards itself as an ecosystem platform company, with the API as its first product. Every model released in a product also comes to the API, the platform stays neutral and competitors are not blocked. The reasoning is that a rising tide lifts all boats, and that OpenAI cannot reach every corner of humanity itself, which ties back to the mission of spreading AGI's benefits widely.

He walks through the developer stack: the Responses API as the low-level, unopinionated primitive for long-running agents, the Agents SDK above it for orchestrating sub-agents and guardrails, then AgentKit with widgets for UI, plus an evals product for quantitative testing. Developers can use the whole stack or go as low-level as they want.

Wu closes by saying the next two to three years should be some of the most fun in tech in a long time. He entered the workforce in 2014, had a stretch of five or six less exciting years, and calls the last three the most energizing of his career. He expects the wave to eventually become more incremental, so he urges people not to take it for granted.

His advice is to engage: use the tools, understand their limits, and watch what they start to do as models improve, whether or not you are an engineer. He admits he is chronically online, but says much of the news is noise and that starting with one or two tools is plenty. Installing the Codex client, or connecting ChatGPT to Notion, Slack and GitHub, is a good start.

In the lightning round he recommends the novel *There Is No Antimemetics Division*, Dan Wang's *Breakneck* and Patrick McGee's book on Apple and China. He also praises Ubiquiti's home networking gear and cites "never feel sorry for yourself" as his motto. From his earlier work on pricing houses at Opendoor, he says power lines, floor plans and curb appeal affected prices more than expected.