LLenny's Podcast
← All episodes
The future of software development with OpenAI’s Sherwin Wu12 February 2026

“Engineers are becoming sorcerers”

5Frameworks
14Insights

Frameworks in this episode

Strategy4 steps

Build for Where the Models Are Going

Design products for the capability one to two years out, not today's model ceiling.

Read →
Productivity4 steps

Context-First Agent Debugging

When a coding agent won't do what you want, treat it as a missing-context problem, not a model problem.

Read →
Productivity3 steps

Hand the Toil to the Model First

Delegate the boring, repetitive parts of engineering to agents before anything else, and collapse the path to production.

Read →
Leadership4 steps

The Surgeon Management Model

Spend over half your time making your top 10% feel like a supported surgeon who can't be blocked.

Read →
Leadership4 steps

Two-Sided AI Adoption

Pair top-down executive buy-in with a bottom-up tiger team of excited power users.

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 4

Hot Take07:30

Engineers are becoming sorcerers casting spells at fleets of agents

Sherwin frames the changing role of an engineer through the SICP 'wizard book' metaphor: programming languages as incantations that have become literal with AI. Individual engineers are turning into tech leads who steer 10-20 parallel agent threads at once. He invokes the Sorcerer's Apprentice as a warning that this leverage requires real skill or the brooms go wild.

  • IC engineers are becoming tech leads managing fleets of agents
  • Engineers run 10-20 parallel Codex threads, steering rather than typing
  • The SICP 'wizard book' cast programming as sorcery back in 1980
  • The Sorcerer's Apprentice is the risk: high leverage but you must know what you're doing

“They're managing fleets and fleets of agents.”

Sherwin Wu · 08:00

“Like it literally feels like we're wizards now.”

Sherwin Wu · 11:30
#future of engineering#agents#vibe coding#metaphor
Hot Take24:00

The one-person billion-dollar startup and the B2B SaaS golden age it triggers

Sherwin says people are pricing in the one-person billion-dollar startup but missing its second- and third-order effects. If one person can build that much, it becomes trivially easy to start companies, triggering a startup boom of bespoke vertical software. He predicts tens of thousands of $10M businesses that are set-for-life outcomes for individuals but poor venture-scale returns.

  • Enabling one billion-dollar solo startup requires ~100 small startups building bespoke supporting software
  • Expect a golden age of B2B SaaS and vertical, single-purpose tools
  • Tens of thousands of $10M businesses are life-changing for individuals but weak VC bets
  • The VC and startup ecosystem itself may reshape as venture-scale returns shrink

“in order to enable a one person billion dollar startup there might be like a hundred other small startups building bespoke software”

Sherwin Wu · 26:00

“there might be tens of thousands of $10 million startups and as an individual it's actually pretty great to have a $10 million business”

Sherwin Wu · 27:00
#startups#b2b saas#venture capital#future of work
Hot Take44:00

The models will eat your scaffolding for breakfast

Sherwin argues you can't blindly listen to customers in AI because the field disrupts itself so fast. Quoting FinTool's founder, 'the models will eat your scaffolding for breakfast' - agent frameworks and vector stores that were essential in 2022-2023 got obsoleted as models improved. Today's fashionable scaffolding, skills and file-based context, could be next. His advice: build for where the models are going, not where they are today.

  • Listening to customers can trap you in a local maximum as models leap ahead
  • Vector stores and agent frameworks were scaffolding that better models made obsolete
  • Today's skills/file-based context management may be the next thing to get eaten
  • This is a version of the bitter lesson applied to building with AI
  • Build for where the models are going, not where they are today

“the models will eat your scaffolding for breakfast.”

Sherwin Wu · 45:00

“make sure you're building for where the models are going and not where they are today.”

Sherwin Wu · 49:00
#scaffolding#bitter lesson#building with ai#product strategy
Hot Take53:30

Business process automation is the underrated giant outside of coding

Sherwin is bullish on business process automation precisely because Silicon Valley ignores it. Software engineering is open-ended, non-repeatable knowledge work; most of the economy runs on repeatable, high-determinism business processes with standard operating procedures. Applying AI to those integrated, data-driven workflows is a massive, underdiscussed opportunity.

  • Software engineering is open-ended and non-repeatable; most jobs are repeatable processes
  • Business processes have SOPs where the goal is determinism, not ingenuity
  • Automating data-integrated, repeatable workflows is a huge opportunity
  • It's underrated because it's outside the Silicon Valley wheelhouse and X discourse

“it's bigger than you would think it is based off of how how people talk about it or don't talk about it on X or…”

Sherwin Wu · 57:00
#business process automation#enterprise ai#opportunity#workflows

Explainer· 5

Explainer04:00

The numbers behind AI coding at OpenAI: 95% of engineers, 100% of PRs

Sherwin shares concrete internal metrics on how deeply Codex is embedded in OpenAI's engineering. Nearly all engineers use it daily, every PR is reviewed by it, and the code that hits production is close to 100% AI-generated first. He also notes a widening productivity gap: heavy Codex users open far more PRs than light users.

  • 95% of engineers use Codex on a daily basis
  • 100% of PRs are reviewed by Codex daily before merge
  • Close to 100% of code is generated by AI first, then reviewed
  • Engineers who use Codex more open 70% more PRs, and the gap keeps widening

“So 95% of engineers um use codeex. Um 100% of our PRs are reviewed by codeex daily as well.”

Sherwin Wu · 04:00

“So uh they're actually opening 70% more PRs uh and uh than than the engineers who aren't using codecs as much.”

Sherwin Wu · 04:30
#codex#openai#engineering productivity#ai coding
Explainer15:00

How Codex turned code review from the worst job into a 2-minute task

Sherwin explains that the tasks engineers hand to AI first are the ones they hate most, which is why work is more fun now. Code review is a prime example: Codex reviews 100% of PRs and collapses review time. For small PRs, teams increasingly trust Codex as the second pair of eyes rather than requiring a human reviewer.

  • The most-hated, most-boring tasks get handed to AI first
  • Codex reviews 100% of PRs; review drops from 10-15 minutes to 2-3 minutes
  • For small PRs, teams often skip human review and trust Codex
  • CI, lint fixes, and deployment steps are heavily automated via Codex

“Yeah, I mean one thing is Codex reviews 100% of all of our PRs at this point.”

Sherwin Wu · 15:00

“it makes you know code reviews go from a you know I don't know 10 15 minute task to sometimes even just like a two…”

Sherwin Wu · 16:30
#code review#codex#automation#developer workflow
Explainer19:30

Why AI lets managers run far larger teams than 6-8 reports

Sherwin argues the manager role has changed less than the IC role so far, but the trend is clear. AI tools that surface organizational context, do research, and write deep-research performance reviews will let people managers operate at much higher leverage, breaking past the old best practice of six to eight direct reports.

  • The manager role has changed less than the IC role, but trends point one way
  • ChatGPT hooked to GitHub, Notion, and Google Docs speeds up performance reviews
  • Managers will manage far more than the current 6-8 direct-report norm
  • The same leverage applies to non-engineering functions like support and operations

“I think managers will be able to manage much larger teams in this world kind of like how you know like software engineers are managing…”

Sherwin Wu · 22:00
#management#org design#ai tools#leverage
Explainer37:30

Why most enterprise AI deployments have negative ROI (and the fix)

Sherwin isn't surprised many AI deployments are net-negative: Silicon Valley forgets it lives in a bubble, and most workers are basic users, not power users. The failure pattern is top-down mandates divorced from real work. The winning setup pairs top-down buy-in with bottoms-up adoption, often via a dedicated internal tiger team that evangelizes and builds best practices.

  • Many enterprise AI deployments are likely negative ROI
  • Silicon Valley lives in a bubble; most employees are basic, not power, users
  • Winning deployments combine top-down buy-in with bottoms-up adoption
  • Staff a dedicated 'tiger team' of excited, technical-adjacent people to spread best practices
  • These evangelists are often not software engineers, but Excel-wizard operations types

“I think we in Silicon Valley just forget that we live in a bubble.”

Sherwin Wu · 38:30

“find or maybe even staff a full-time team internally that is this kind of tiger team internally”

Sherwin Wu · 41:30
#ai adoption#enterprise#roi#change management
Explainer50:00

Multi-hour coherent tasks and audio are what's next

Asked where the models are heading in the next 6-18 months, Sherwin points to two things. First, task length: the METR benchmark shows models trending toward multi-hour coherent software tasks, which will reshape the products built around them. Second, audio, which he calls hugely underrated because most of the world's business runs on talking, not text.

  • The METR benchmark tracks how long a task models can do 50%/80% of the time
  • Products today optimize for ~10-minute interactive tasks; multi-hour is coming
  • In 12-18 months, models may run 6-hour dispatched tasks coherently
  • Native speech-to-speech audio models are a hugely underrated enterprise domain

“in the next 12 to 18 months we could see models that could do multi-hour long tasks very very coherently.”

Sherwin Wu · 51:30

“A lot of the world's business is done via audio.”

Sherwin Wu · 52:30
#model roadmap#metr benchmark#audio#long-running agents

Story· 2

Story12:30

The OpenAI team maintaining a 100% Codex-written codebase with no escape hatch

Sherwin describes an internal experiment where a team maintains a fully Codex-written codebase and deliberately removes the 'escape hatch' of hand-coding when the agent gets stuck. The forced constraint surfaces the real bottleneck: most agent failures come down to missing context, so the fix is encoding tribal knowledge into docs, comments, and MD/skills files.

  • An internal team maintains a 100% Codex-written codebase and can't fall back to hand-coding
  • Most agent failures are a context problem, not a model problem
  • The fix is encoding tribal knowledge into the repo via comments, structure, and MD/skills files
  • OpenAI plans to publish a blog post of best practices from the experiment

“there's a team that that's actually doing an experiment right now uh within OpenAI where they are basically maintaining a 100% codeex written codebase.”

Sherwin Wu · 13:00
#codex#context engineering#agents#best practices
Story1:16:00

What surprised an Open Door pricing engineer about home values

Before OpenAI, Sherwin built the model that told Open Door how much to pay for houses. He shares the variables that surprised him: high-voltage power lines near a home hurt price significantly, floor plans are hugely important but almost impossible to quantify in code, and curb appeal matters more than expected - there's reportedly a Zillow finding that front-door replacement is the highest-ROI home improvement.

  • High-voltage power lines near a house meaningfully lower its price
  • Floor plans matter a lot but are extremely hard to quantify in code
  • Ops teams could 'feel' a bad floor plan that data couldn't capture
  • Curb appeal is underrated; front-door replacement is reportedly the highest-ROI improvement

“power lines and like uh high voltage power lines like are super super uh actually impact your price quite a lot.”

Sherwin Wu · 1:16:30

“there's a Zillow book on on this where the front door replacement tends to be the highest ROI uh for homes.”

Sherwin Wu · 1:17:30
#open door#real estate#pricing models#career

Q&A· 1

Q&A57:30

Will OpenAI squash your startup? Sherwin's honest answer

Asked how startups avoid being crushed by OpenAI, Sherwin says don't overthink it: the opportunity space is so big that no startup he's seen has failed because a big lab squashed it. They fail because they built something customers didn't love. He frames OpenAI as an ecosystem-platform company whose charter commits it to spreading AI's benefits, so every model ships in the API and competitors aren't blocked.

  • No startup Sherwin has seen failed because a big lab squashed it; they failed on product-market fit
  • The opportunity space is so big VCs now fund directly competitive companies
  • OpenAI views itself as an ecosystem-platform company, neutral and open
  • Every model released in a product also ships in the API; competitors aren't blocked
  • This traces back to the charter mission of spreading AGI's benefits to all humanity

“Every startup that I've seen that has kind of fizzled out is not because open AI or you know big lab or Google or something…”

Sherwin Wu · 58:00

“I can't overstate how big of an opportunity there is right now.”

Sherwin Wu · 58:30
#openai platform#startups#api#competition

Tool· 1

Tool1:11:30

Sherwin Wu's book recommendations: antimemetics, Breakneck, Apple in China

In the lightning round, Sherwin recommends one fiction and two non-fiction books. The fiction pick is a sci-fi/horror title about a government agency fighting things that make you forget them, which he devoured in two days. The non-fiction picks are Dan Wang's Breakneck and Patrick McGee's book on Apple and China, both shaping his thinking on US-China relations.

  • Fiction: 'There Is No Antimemetics Division' - sci-fi about fighting things that make you forget them
  • Non-fiction: Dan Wang's 'Breakneck' on the lawyerly US vs the engineering society of China
  • Non-fiction: Patrick McGee's book on Apple and China, full of inside information
  • Both non-fiction picks reflect his year of reading on US-China relations

“the lawyerly US is the lawyerly society. China is the engineering society uh and their pros and cons to each.”

Sherwin Wu · 1:12:30
#books#recommendations#lightning round#us-china

Takeaway· 1

Takeaway31:30

Spend 50% of your time on your top 10%: the surgeon model of management

Sherwin's core management philosophy is to spend more than half his time empowering his top ~10% of performers, and it matters even more in an AI world where high-agency people pull ahead. He borrows the Mythical Man-Month's surgeon metaphor: treat your best engineers like a surgeon, with the manager looking around corners to unblock them before they even ask.

  • Spend more than 50% of your time with your top ~10% performers
  • AI supercharges high-agency people, widening the productivity spread on a team
  • The Mythical Man-Month's surgeon model: everyone supports the one person doing the work
  • A manager's job is to look around corners and remove organizational blockers

“it's like more than 50% of your time with your top performers with maybe your top like 10% uh performers”

Sherwin Wu · 32:30

“software engineering might end up moving into a world where that software engineers are like surgeons or like in a surgery room there's like one…”

Sherwin Wu · 33:00
#management#top performers#surgeon metaphor#leadership