Uncapped with Jack AltmanTony Xu on AI Consumption, Atoms vs. Bits, and the 1% Better Mentality | Ep. 55
CHAPTERS
- 0:00 – 1:48
AI budgets, token burn, and measuring real outcomes
Jack opens with the industry’s rising AI spend and the worry that heavy token consumption doesn’t necessarily translate into better unit economics or growth. Tony frames the purpose of technology as lowering costs via solved customer problems, while acknowledging an unavoidable discovery phase with inefficiency.
- •AI spend is becoming a major line item; outcomes can be unclear
- •Technology’s goal: solve problems and make things cheaper over time
- •Compute constraints and cost profiles can lag behind desired outcomes
- •Early-stage exploration with new tech is inherently inefficient
- •Success metric should be tied to customer-impacting results
- 1:48 – 4:23
A practical framework: start from customer jobs, then experiment with guardrails
Tony explains how DoorDash tries to avoid undirected experimentation by anchoring AI work on specific customer outcomes. Teams get multiple “shots on goal,” but the exploration is intentionally bounded by measurable goals across consumers, merchants, and dashers.
- •Begin with customer jobs-to-be-done, not model capability
- •Accept some inefficiency, but keep experimentation contained
- •Direct teams toward measurable customer outcomes
- •Balance exploration with intentionality (not pure ‘YOLO’)
- •Apply across marketplace participants: consumers, merchants, dashers
- 4:23 – 5:24
Where AI already delivers: faster merchant onboarding, safer deliveries, better ops
Tony gives concrete examples of AI use that DoorDash can measure today, especially around merchant onboarding and operational risk. These are immediate, trackable improvements that benefit end customers through better selection, reliability, and trust.
- •Merchant onboarding 35–50% faster via AI-generated catalogs/menus
- •AI-assisted photo editing and better store/restaurant descriptions
- •Fraud and safety incident detection with preventative actions
- •Preference for use cases with near-term, measurable impact
- •Operational improvements translate into customer value
- 5:24 – 7:59
AI-native isn’t just coding: redesigning workflows for real productivity
They discuss why coding assistance alone doesn’t guarantee large productivity gains: much of engineering time is meetings, reviews, alignment, and coordination. DoorDash looks at productivity distributions and tries to replicate the behaviors and workflows that create outlier performance across the org.
- •Only part of engineering time is writing code; the rest is process overhead
- •Need AI-native operations, not just AI-assisted development
- •Productivity gains vary widely; some individuals become extreme outliers
- •Studying the distribution reveals what’s possible and what to replicate
- •Leadership challenge: create an environment that lifts the whole team
- 7:59 – 11:43
Atoms vs. bits: DoorDash’s ‘war for atoms’ and why assistants need real-world execution
Tony and Jack contrast attention-based ‘bits’ businesses with DoorDash’s physical logistics (‘atoms’). Tony argues that personal assistants become truly useful only when they can cause real-world actions, and DoorDash’s strength is executing in the messy physical world at city scale.
- •DoorDash focuses on physical-world logistics, not just attention and pixels
- •AIs/assistants need partners that can execute real-world actions
- •DoorDash’s long-term strategy: be best-in-class at moving ‘atoms’
- •The physical world is full of edge cases: traffic, weather, variability
- •Opportunity for deep partnerships with ‘bits’ companies building agents
- 11:43 – 13:18
DoorDash as an applied AI company: recruiting and the next AI frontiers
Tony describes DoorDash as ‘applied AI’ since day one (math/ML/LLMs), but notes recruiting pressure when the market is hottest for frontier AI work. He also highlights interest beyond LLMs, including autonomous delivery, where learning-based approaches are improving physical decision-making.
- •DoorDash positions itself as applied AI solving real-world problems
- •Recruiting is competitive for ML and applied research talent
- •LLMs matter, but physical-world AI/robotics also becomes important
- •Autonomous delivery efforts began in 2019
- •Shift from rules/maps toward more learning-driven autonomy
- 13:18 – 16:43
LLMs inside the DoorDash app: from complex marketplace to personal city agent
As DoorDash expands from restaurants to many retail categories and countries, the app becomes harder to navigate. Tony argues LLMs can simplify discovery and decision-making, turning DoorDash into a personal agent that helps users transact with businesses across their city.
- •DoorDash expanded into grocery, convenience, alcohol, retail, reservations/deals
- •More categories increase product complexity and user friction
- •LLMs can improve search, research, and personalized recommendations
- •DoorDash aims to be a ‘personal agent’ for city commerce
- •Leverage large-scale order history to tailor suggestions and choices
- 16:43 – 18:19
Was the premium delivery burrito obvious? Early skepticism and learned willingness to pay
Jack asks whether consumers’ willingness to pay delivery premiums was clear in 2013–2015. Tony recalls skepticism even for the Taco Bell launch, but demand proved that time savings and convenience can justify surprisingly high delivered prices.
- •Not obvious early on that customers would pay large delivery premiums
- •2015 Taco Bell partnership was a key test; Tony was skeptical
- •Value proposition: time savings and convenience outweigh price differences
- •Consumer behavior evolved to internalize time/value tradeoffs
- •Delivery demand validated even for low-cost, ubiquitous food
- 18:19 – 20:56
The 70-year shift: why food trends keep favoring prepared meals
Tony cites long-run data showing a steady shift from grocery spend toward restaurant/prepared food, continuing past DoorDash’s founding. He ties the trend to structural forces like dual-income households and the persistent growth in restaurant count.
- •1950s: 70–80% of food spend on groceries; 2013 closer to 55/45
- •Today: roughly 55% restaurants vs 45% groceries (trend continues)
- •Dual-income households rose from ~25% to ~70% over decades
- •Restaurant counts increase almost every year despite business churn
- •Behavior (spend/time) reveals preference for food prepared by others
- 20:56 – 22:20
Consumer psychology and affordability: small frequent purchases, big emotional utility
Jack notes the tension between affordability pressures (housing, etc.) and rising spend on food convenience. Tony argues food is a high-frequency need (20–25 times/week) and also a reliable way for people to ‘feel good,’ making it resilient even in tougher macro environments.
- •Food decisions recur 20–25 times per week—unlike major purchases like homes
- •Time constraints make constant cooking difficult even for avid cooks
- •Affordability pressure is real, but food delivers frequent perceived value
- •Convenience purchases can shift from indulgence to emotional necessity
- •DoorDash focus: reduce costs while increasing value simultaneously
- 22:20 – 26:15
Beyond food: logistics in messy ‘atoms’ markets and the inventory/substitution problem
Tony explains why DoorDash must track broad consumer trends as it grows beyond food, especially affordability. He contrasts the controllability of software with the edge-case-heavy physical world, then dives into grocery and retail challenges like inaccurate shelves and costly substitutions/returns.
- •As a large consumer business, DoorDash must watch trends beyond food
- •Affordability spans categories: housing, transport, healthcare, education, banking
- •Physical-world systems are unpredictable; everything is an edge case
- •Grocers often don’t know exact shelf inventory due to constant movement
- •Substitutions, refunds, and returns drive hidden costs that hurt affordability
- 26:15 – 28:50
Sequencing and focus: the ‘greedy algorithm,’ seven years on restaurants, then groceries
They discuss resisting tempting adjacencies by focusing on sequencing—what to do now vs later. Tony describes DoorDash’s early underestimation of the market and why the company stayed on restaurant delivery for seven years before expanding into groceries and additional lines of business.
- •Many ideas are ‘good’; the real question is timing and opportunity cost
- •DoorDash underestimated scale early (YC revenue estimate ~ $100M)
- •Staying on a ‘winning swing’ can be rational (greedy algorithm)
- •Seven years focused on restaurants before moving to groceries
- •Expansion later included international, ads, and B2B/first-party merchant tools
- 28:50 – 31:45
Building a billion-dollar ads business without degrading the consumer experience
Tony distinguishes between ‘building ads’ and building an ads marketplace that preserves relevance and user trust. He credits the ads team for reaching $1B in revenue quickly while enforcing dual goals: strong advertiser ROI and a high-quality consumer experience.
- •Ads are easy to add; hard to do without harming UX
- •Core tension: irrelevant ads create long-term product ‘tax’
- •DoorDash prioritized relevance and constraint-driven execution
- •Fastest to $1B ad revenue (per Tony), but pride is in how it was achieved
- •Dual objective: best-in-class returns for advertisers and consumers
- 31:45 – 38:34
The 1% better mentality: execution culture, math + humanity, and hiring for values
Tony describes DoorDash’s operating philosophy as relentless incremental improvement under constraints, while remembering every order involves real people. He emphasizes the blend of rigorous measurement with empathy for merchants, dashers, and consumers—and the importance of self-selection when hiring.
- •Execution requires objective measurement plus respect for human impact
- •Every order involves at least three humans: merchant, dasher, consumer
- •‘1% better every day’ across selection, affordability, reliability, support
- •Values are harder to test than skills; look for motivation and empathy
- •Self-selection matters: some thrive in messy real-world problem spaces
- 38:34 – 47:27
Growth levers at scale and the future: speed, autonomy, drones—and preserving agency
Tony outlines two parallel growth motions: improve the core continuously and build new lines that start inefficiently. They discuss speed as a perpetual customer desire, why prep time dominates delivery time, and why everyone at DoorDash still dashes to understand edge cases; Tony closes on excitement about new products and new ways of working that sustain startup-like agency.
- •Sustain growth by strengthening the core while creating ‘the new’
- •Customers always want faster delivery, but prep time is often the bottleneck
- •Drones/AVs will matter, yet end-to-end orchestration is the real challenge
- •DoorDash requires employees to do deliveries to learn physical-world realities
- •Tony’s excitement: rapid change enabling better products and better internal velocity; keep agency high in year 13+
