OpenAI’s head of platform engineering on the next 12-24 months of AI | Sherwin Wu
These notes were generated by AI and may contain inaccuracies.
95% of engineers use Codex. 100% of PRs are reviewed by Codex for engineers. Engineers are becoming tech leads, managing fleets of agents. It feels like wizards casting spells that go out and do things. Second- or third-order effects of the one-person billion-dollar startup aren't priced in yet, enabling a hundred small startups building bespoke software, leading to a golden age of B2B SaaS.
There's stress when agents aren't working. A team at OpenAI is maintaining a 100% Codex-written codebase, running into these problems without an escape hatch of manual coding.
Listening to customers isn't always the right strategy in AI because the field and models change quickly, disrupting themselves. Models eat scaffolding for breakfast.
Advice: Build for where models are going, not where they are today. > "This is the worst the models will ever be." – Kevin Whale, VP of Science at OpenAI.
Sherwin Wu is head of engineering for OpenAI's API and developer platform.
DX is a developer intelligence platform. Companies like Dropbox, Booking.com, Adion, and Intercom use it to understand AI's value and impact on engineering productivity. Visit getdx.com/lenny.
Sentry catches crashes, slowdowns, regressions. Shows what happened, where, why, down to the commit, developer, line of code. Sentry's AI debugging agent uses context for root cause, suggests fixes, opens PRs, reviews PRs. Try at sentry.io/lenny with code Lenny for $100 credits.
Sherwin Wu still writes code occasionally. For managers, AI tools are easier than manual coding. All code by managers like himself is written by Codex.
95% of engineers use Codex daily. 100% of PRs reviewed by Codex, suggesting improvements.
Engineers using Codex more open 70% more PRs, gap widening.
Vast majority of code (close to 100%) generated by AI first.
"This is the worst the models will ever be."
Peter the Claudebot/moldbotclaw (now OpenClaw) trusts Codex enough to commit directly. He's a great user giving feedback. Mold's book shows AI agents talking, surreal, like Her happening in real life.
Engineer role changed dramatically: from writing every line to AI writing all code. Engineers becoming tech leads managing 10-20 parallel Codex threads, steering agents, giving feedback.
Reference to SICP (Structure and Interpretation of Computer Programs), MIT intro programming textbook on Scheme, cult following, called the wizard book.
Programming as sorcery: engineers as wizards, languages as incantations/spells. Written in 1980, metaphor persists. Now literal with vibe coding: tell Codex or Cursor what to do, it does it.
Like Sorcerer's Apprentice from Fantasia: powerful but need skill to avoid chaos (brooms flooding). Senior engineers manage 20 Codex threads without models going off rails. High leverage, fun, magical.
Genie analogy: be precise with wishes to avoid monkey's paw side effects.
Stress when agents fail, need to stay on top.
OpenAI team experiment: 100% Codex-written codebase, no manual fallback. Forces solving agent issues. Blog post coming on learnings.
When agents fail, often due to insufficient context/information. Solution: add documentation, encode tribal knowledge via code comments, structure, MD files, etc.
Codex reviews 100% of PRs, speeds reviews from 10-15 minutes to 2-3 minutes with suggestions. For small PRs, trust Codex as second pair of eyes, no human review needed.
CI process automated: lint errors fixed by Codex patching and restarting.
Reduces work, enables more PRs/shipments. Codex writes and reviews its own code.
Use internal model variants for different perspectives, dogfood own models. Not fully zero human attention (30% still), prevents brooms going crazy.
Not 100% of production code written by AI, but vast majority authored by AI; nearly every engineer uses it heavily.
Less change for managers than engineers; no Codex for managers yet, but used for managery tasks.
Codex empowers top performers most, widening productivity spread. High-agency people supercharge.
Management philosophy: spend majority of time with top performers to unblock, ensure happiness/productivity. Even more true in AI world.
Example: Letting 100% Codex codebase team run freely pays dividends.
AI tools like ChatGPT with organizational knowledge enable managers to do research and understand organizational context better. For performance reviews, use ChatGPT hooked up to internal knowledge like GitHub, Notion docs, and Google Docs to generate a deep research report on what a person has done over the last 12 months.
Managers will manage much larger teams, similar to software engineers managing 20-30 codebases. Current best practice is 6-8 for software engineering. Tools allow people managers higher leverage. This applies to non-engineering domains like support or operations, where passing work to agents enables more work and managing more people. Some teams have EMs managing many people adeptly using these tools for higher leverage, better understanding of team activities, and organizational context.
AI makes good people better and it makes great people exceptional.
Spend more time with top performers to unblock them and ensure they have everything needed. Example: Group of engineers exploring best practices for interacting with models—high leverage activity. As manager, encourage exploration, share best practices org-wide via knowledge sharing sessions and documents. This elevates everyone and shows top performers getting exceptional.
Idea of the one-person billion-dollar startup (possibly first said by Sam Altman). One person gains massive agency and leverage to build a billion-dollar business.
Second-order effect: Easier to create startups in general, leading to a huge startup boom and SMB boom. Anyone can build software for anything. AI startups are vertical-oriented, leaning into specific domains. Could see 100x more startups. To enable one-person billion-dollar startups, hundreds of small startups build bespoke software supporting them. Golden age of B2B SaaS and startups as building/running software becomes easier.
There might be one one-person billion-dollar startup but there might be like a hundred $100 million startups [or] tens of thousands of $10 million startups.
$10 million business sets you for life.
Third-order effect: Startup ecosystem and VC ecosystem change. Micro companies (1-2 people) lead to handful of big platforms supporting startups. Fewer venture-scale (100x-1000x) returns; more smaller $10-50M companies—solid for individuals, not great for venture.
On support costs for billion-dollar scale: Even with AI, support is hard unless high ACVs and few customers. People email support for solvable issues. AI takes you far, but may need contractors (does that count as single-person?).
Response: Lenny's podcast could become billion-dollar via outsourcing to smattering of small startups building tailored software (e.g., 10-20 for podcasts/newsletters support)—one-person startups coding easily. Outsource vs. in-house shifts as software cost collapses, reducing company size while driving high-leverage company.
Example: Peter at Clawbots/moldbot/openclaw bombarded by asks, emails, pings, DMs, PRs—insane, like post-HatchVT launch craziness. Distribution becomes crucial as attention competes; people with audience/platform more valuable.
Spend >50% time with top 10% performers, empowering them. Analogy from Mythical Man-Month (1970s): Software engineers as surgeons—one does work, others support (nurse, resident, fellow providing scalpel/tools).
Not exactly collaborative reality, but emulate: Treat engineers as surgeons, manager as support army (really just manager). Look around corners, unblock organizationally/process-wise—key bottleneck when cranking PRs. Have scalpel ready proactively.
AI can help: Query ChatGPT with company knowledge for active blockers (Notion docs, Slack). Anticipate future blockers for engineers/teams months ahead—second/third-order anticipation.
Observation: Many AI deployments likely negative ROI. General sentiment outside tech: AI forced on them.
Silicon Valley/X/software engineering bubble—most people not AI-savvy, not following model releases, asking simple questions, little understanding of tech.
Successful deployments: Top-down buy-in (C-suite wants AI-first, buys tools) + bottoms-up adoption (excited employees learn, evangelize, build/share best practices). Everyone's work unique (software eng ≠ finance ≠ ops ≠ sales)—needs bottoms-up for last-mile.
Anti-pattern: Pure top-down mandate, divorced from work. Employees: "I know I'm supposed to use this, maybe on performance review, but unsure what to do; no one else is."
Recommendation: Staff full-time internal tiger team to explore capabilities, apply to workflows, knowledge share, create excitement.
Tiger team: Technical-adjacent people (not always software engineers—rarity, expensive). E.g., support/ops leads who love tools, Excel wizards—get excited, build team around them. Anti-pattern: Top-down CEO/exec "we are going AI first. Everyone's going to be"
The CEO exec team declared going AI first, judging performance on productivity increases from AI tools. Top-down mandates without bottom-up team buy-in fail. Advice: Find excited people, form an AI evangelist team to spread usage across the organization. Empower high performers in AI adoption: let them build hackathons, hold seminars, do knowledge sharing to create internal excitement.
Talking to customers is useful, but in AI, models and field evolve rapidly, disrupting tooling and scaffolding. Quote from Nicholas, founder of FinTool: "the models will eat your scaffolding for breakfast."
In 2022 post-ChatGPT, models were raw, requiring heavy scaffolding like agent frameworks and vector stores. Models improved, eating much of that scaffolding. Current fashion: skills files and context management—may become obsolete as models manage it themselves.
Agent frameworks less useful now. In 2023, vector stores seemed essential for organizational context: embed corpora, optimize vector search. Better approach: trust models, give tools for search (not just vector stores)—e.g., files on filesystem like skills.md and agents.md.
Vector stores still used, but ecosystems built solely around them changed. Customers in local maxima push for better vector stores or agent frameworks, leading astray. Balance customer feedback with model trends over 1-2 years.
Bitter lesson in AI/ML: Less logic added scales better—models compute smarter. Applied to building with AI: models eat engineered scaffolding. OpenAI API team guilty of missteps, but learning daily.
Advice for Building on API/Agents
Build for where models are going, not today. Successful startups target capabilities 80% there: product works okay initially, explodes as models improve (e.g., o3 to 5.1/5.2). Creates superior experience vs. static assumptions. Models improve fast, minimal wait needed.
Key trend: Coherent task length (Meter benchmark for software engineering tasks). Current: Multi-hour tasks at 50%, ~1 hour at 80%. Trend steep—products today optimize for minutes (e.g., Codex interactive ~10 min max, though multi-hour possible exceptionally). In 12-18 months: Multi-hour coherent tasks routine, potentially 6-hour dispatchable tasks.
Product types change: Need feedback loops to avoid runaway. Expands model-usable tasks dramatically.
Multimodal improvements, especially audio (speech-to-speech native models, new architectures). Audio underrated in enterprise/business: Much work/communication via audio (services, operations). Exciting unlocks in 6-12 months.
Summary: Agents/AI tools run longer; audio/speech native, better, core.
Silicon Valley bubble: Tech work (software engineering, PM, data science, strategic finance) is open-ended, non-repeatable. Most economy: Repeatable business processes with SOPs, low deviation (e.g., support lines, utility company interactions).
Bullish on AI for automating repeatable, high-determinism processes integrated with enterprise data/systems. Underrated outside tech—huge opportunity to improve.
Bigger productivity/job impact outside engineering on repetitive tasks. In 20 years, transforms companies more on business processes than software engineering (though both massive, under-discussed).
Market too big—don't overthink OpenAI/labs competing. Failures from lack of customer resonance, not squashing (e.g., Cursor huge in competitive coding). Build what people love.
Opportunity unprecedented: VCs invest in competitors routinely. Even niche love yields massive business.
OpenAI as ecosystem platform company (API first product). Foster, don't squash: All models/products to API (e.g., Codex-optimized). Platform neutral—no blocking competitors, full model access. Testing "signin with ChatGPT."
Rising tide lifts all boats—API growth from this approach.
Mission: Build AGI, spread benefits to all humanity. Can't reach alone (e.g., podcaster/newsletter support bots). API (launched ~2020) enables. Diversity of builds fulfills mission.
ChatGPT App Store (separate team under ChatGPT, close collab; apps SDK). Leverages 800M weekly active users (~10% world, growing). Enables others to build/expand audience.
Free ChatGPT (GPT-4o-mini today vs. 2022) raises global floor. Leans into health, education. Billionaires get marginal gains over free access anywhere.
For $20 a month, users access the same AI that billionaires use, similar to using the same iPhone as Mark Zuckerberg. For $200 a month, get the pro model, though billionaires likely use the plus tier for daily tasks. This democratization spreads benefits worldwide and drives OpenAI's work.
The API offers developer endpoints for sampling from models. The most popular is the Responses API, optimized for building long-running agents. At the lowest level, provide text to the model, poll for progress, and retrieve the response. It is unopinionated, allowing developers to build anything.
- Agents SDK: Uses Responses API or other endpoints to build traditional agents in infinite loops with sub-agents, delegation, guardrails, subtask farming, and orchestration of agent swarms.
- Agent Kit and Widgets: UI components for beautiful interfaces on top of API or Agents SDK, as agents often share similar UIs.
- Eval Products: Evals API for quantitative testing of models, agents, or workflows.
These layers provide increasing abstraction and opinionation. Use the full stack for quick agent building or drop to Responses API for custom work.
The next two to three years will be the most fun in tech and startups in a long time. Sherwin entered the workforce in 2014; it was great initially, then five to six unexciting years, followed by three insanely exciting years. This wave will continue, enabling exploration, invention, world-changing, and work transformation. Do not take it for granted.
Lean in: build tools, use tools (no software engineering needed), understand limitations to track improvements. Get familiar with the technology instead of letting it pass.
Much industry news on X is noise due to frenetic pace. No need for 110% awareness. Start small with one or two tools. Install Claude client or ChatGPT, connect to internal data sources like Notion, Slack, GitHub, and experiment with capabilities.
Books
Fiction: There Is No Antimemetics Division by qntm (online author, shared on X). Science fiction about a government agency fighting things that make you forget. Smart, creative, fresh, unintentionally hilarious; devoured in two days, couple hundred pages.
Non-fiction (two recommendations on China/US relations, last year):
- Breakneck by Dan Wang: Lawyerly US vs. engineering society China, with pros/cons.
- Patrick McGee book on Apple and China: Fascinating on Apple's China relationship and company insights; page-turner, timely. Sherwin is an Apple fanboy.
Limited time with two kids and busy job. Watched new season (season 3) of anime Jujutsu Kaisen (JJK); really good. Big fan of Japanese anime for novel, unique plots and universes that Western media avoids.
Went all-in on Ubiquiti routers and security cameras for home Wi-Fi/networking. Like the Apple of home networking: beautiful hardware, excellent software with great mobile app. Requires Ethernet wiring. Security cameras shine with incredible mobile, Apple TV, iPad apps for live feeds. Pricey but worth it.
"Never feel sorry for yourself." Remind yourself of agency to pull yourself up amid work/life challenges. Repeats it to himself and others.
- High-voltage power lines: Major price impact; buzzing, families avoid for kids (observed in Dallas).
- Floor plans: Hard to quantify (kitchen width/style, master bedroom location). Felt intuitively; ops team flagged unsellable homes due to bad flow.
- Curb appeal/front door: Underrated; Zillow book notes front door replacement has highest ROI. First impressions as buyer matter.
Floor plans often not digitized; relied on paper plans in Phoenix/Dallas.
Sherwin Wu on X/Twitter as @sherwinwu; tweets about OpenAI, API, products. Loves hearing about startups/hacks; reach out to discuss building and support.
Keep Lenny's Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
