The next chapter: delivered. @gostanford x @DoorDash

Bay area schools with the most National Merit Semifinalists & the share of senior class that qualified. South Bay is absolutely dominant, there are many schools where >10% of the class qualified

We built Vera, an internal data agent that gives teams across DoorDash reliable answers to business questions from our data. Frontier models inside general-purpose harnesses like Codex and Claude Code often struggle with the sprawl of enterprise data. Our own harness, with prompting, retrieval, and domain skills tuned to that data, significantly outperformed standard agents with simple data connectors. We benchmark models and reasoning efforts inside the same harness, with an LLM judge scoring tool usage, table selection, and answer correctness. With extended reasoning effort, frontier models from Anthropic and OpenAI achieve similar scores, while differing in token usage:

Today we’re proud to announce Flux, our own cloud platform for running agents across DoorDash. Flux automated 130k engineering tasks in one month and now powers 25k+ code reviews each week. Creating the platform ourselves moved us beyond local, laptop-based agent workloads, while preserving enterprise security guardrails over execution and access to DoorDash systems. Here’s how Flux works and what we learned putting cloud-based agents to work:

EVERYTHING IS AN EDGE CASE @t_xu (Tony Xu), Co-Founder & CEO, @DoorDash, interviewed by @jaltma (Jack Altman) (Uncapped) (PS: This was a special episode for me to transcribe because I worked closely with/for Tony at DoorDash, and this episode perfectly brings out the values - customer-focus, 1% better, etc - that the team lives by on a daily basis, and that show why Tony is one of the most exceptional leaders in technology). Summary: Tony Xu bet DoorDash on the physical world back in 2021, while the team was playing with GPT-2 and before ChatGPT shipped. He split the economy into battles for attention and battles for atoms, and picked atoms, on the logic that an assistant is useless if it cannot do real things. That choice carries a tax. Traffic and weather break the plan daily, roughly 20 small systems sit behind one burrito, and every order involves at least 3 humans. Xu's operating answer is to point AI at measurable customer outcomes and get 1% better every day. 1. The War For Atoms. DoorDash picked its side of the AI economy around 2021, before ChatGPT existed, while the team was playing with GPT-2. Xu divided the world into battles for attention, which is bits, and battles for the physical world, which is atoms. DoorDash committed to atoms and started building the catalog of where every item and every parking spot sits inside a city. In Xu's words, "what's the point of having a personal assistant if it can't actually do real things." 2. Directed Discovery. Every company is burning tokens right now without knowing what it bought, and Xu accepts that waste under one condition. New technology brings a discovery period where you have a new toy and no clear use for it, so the spend has to sit inside problems with measurable customer outcomes. DoorDash points teams at consumers, merchants and Dashers, then gives them as many attempts as possible at those specific outcomes. Merchants now onboard 35% to 50% faster because AI builds the menu, edits the photos and writes the description. 3. The Other Half Of The Workday. Writing code is 25% to 50% of an engineer's day, which caps what coding models can do for a company. The rest is product reviews, design meetings and alignment with business teams, and none of it speeds up because the model got better at code. Xu wants DoorDash AI native in how it operates, not only in how it ships software. Teams that automate the coding half and leave the rest alone will wonder why the gain never reached the numbers. 4. The 35x Distribution. DoorDash sees an average productivity lift near 50%, with individual engineers running 35 times more productive. Xu treats the spread as information, because the outlier tells you the ceiling of what the tools can actually do. The management question is whether that person is simply far out on the curve or running a workflow you can teach everyone. Xu takes the second view and says his job is moving the rest of the class up to the higher plane. 5. Everything Is An Edge Case. In the physical world every day brings traffic and weather, and neither is perfectly predictable. A jam that adds 20 minutes is a real problem for one real customer, and no amount of structured data prevents it. Xu says people on the software side often underestimate this, because controlling information end to end gives a false sense of what is controllable. His prescription is staying expert and proficient at that game rather than expecting to finish it. 6. The Busy Kitchen. Drones and autonomous vehicles are coming, and Xu says they will not touch the part of delivery that takes the most time. Preparation dominates: cooking inside an understaffed kitchen, or picking inventory inside a retail shop. Flying over traffic removes one segment of the delivery and leaves the bottleneck in place, and a drone still cannot walk a store and check out. Anyone pricing an autonomy thesis on delivery speed is working the wrong segment. 7. The Dashing Requirement. Bringing someone a burrito breaks into roughly 20 small systems, and one broken step means a fix downstream. Everyone at DoorDash still does deliveries, a practice the company has kept since day one. You learn the failure modes by getting stuck in the wrong elevator, parking in the wrong alleyway, and finding out that a restaurant makes food at 3 separate stations. Xu says none of this can be known a priori, which makes doing the work a requirement rather than a culture exercise. 8. The Seventy-Five Year Line. In the 1950s Americans spent 70 to 80 cents of every food dollar on groceries. By 2013 groceries were down to 55 cents, and today restaurants take the larger share at 55 cents. Over the same stretch dual income households went from about a quarter to roughly 70%, and in 60 to 70 years of counting there are maybe 2 years where the total number of restaurants failed to beat the year before. People vote with their activity, and the vote says they value food someone else prepared. 9. The Sequencing Question. At DoorDash the distance between a good idea and a great one is usually timing. Xu spent 7 years on restaurants before opening category 2, groceries, and the company now runs restaurant delivery, non-restaurant delivery, international, advertising, and a B2B business that hands merchants the stack DoorDash built for itself. The question asked of every adjacent opportunity is opportunity cost against the current thing. Xu calls the founder version of this the greedy algorithm: keep going while you are on a winning swing. 10. The Advertising Constraint. The hard part of an ads business is holding 2 returns at once, best-in-class for advertisers and best-in-class for the consumer experience. An irrelevant ad is a small tax on the consumer, and the economics of advertising are attractive enough that the tax compounds quietly. DoorDash hit a billion dollars in ad revenue faster than any company in history, and Xu is prouder of the constraint the team held than of the record. After a decade of small trade-offs, unwinding them is close to impossible. 11. Math And Humanity. Every DoorDash order involves at least 3 humans: a merchant, a Dasher and a consumer. Xu's mother worked in a restaurant, which is why he describes merchant work as a life rather than a job, an identity and a household income. Xu says running on metrics alone, and accepting the human cost when the numbers look good, is not good enough in his book. Hiring for both traits comes down to self-selection, because skills are straightforward to assess and values are not. 12. Agency Over Intensity. Thirteen years in, Xu measures the founder experience by agency rather than hours worked per week. People join startups for the feeling of impact and agency, and that feeling is what a scaling company loses first. His stated goal for this period of change is better products for customers and better ways of working inside the company. The payoff he names is people doing the best work of their careers, then leaving with more confidence for whatever comes next.

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. images.nvidia.com/pdf/Open-W…

Because of DashBench, we’re able to quickly discern which model combinations yield the best results and at the best cost. For example, with DashBench we’ve seen Kimi K2.6 + Fable 5 vastly outperform our current Sonnet 4.6 + Opus 4.8 harness at a cheaper cost. Having our own benchmark allows us to build confidence in leveraging the frontier intelligence and open-weight models without compromising on enterprise outcomes.

Just five years ago, finding worthwhile Texas barbecue might have been a challenge in the Bay Area, but now it’s the region’s most popular style. Here are the shining stars of Southern-style barbecue in the Bay Area: sfchronicle.com/projects/bes…

Introducing Open USD: a stablecoin built for the internet economy, designed by the businesses growing it. joinopenstandard.com/blog/in…

"What's for dinner?" is a question AI can now answer. bit.ly/4xSPSNK