Erik Bernhardsson founded Modal, a cloud infrastructure platform for data and AI workloads. He previously served as CTO of Better and worked on data and recommendation systems at Spotify. His open-source projects include Luigi and Annoy. — Erik Bernhardsson — About.

Infographic for "Lessons from Erik Bernhardsson, founder and CEO of Modal".

On Startups and Business

  1. Finding the startup journey: After years of technical work at Spotify, Bernhardsson realized he also valued being part of a startup’s broader journey, not just solving difficult engineering problems. — ML in Production Interview.
  2. Look beyond fashionable industries: He advises evaluating unusual industries on their underlying opportunity: Spotify and ride-hailing looked less obvious before their later success. — Career Advice.
  3. Connect models to commercial outcomes: A churn model matters when it helps decision-makers reduce churn; predictive accuracy alone is not the business outcome. — ML in Production Interview.
  4. Prototype the business value end to end: Begin with a simple end-to-end prototype that tests whether the proposed model can create value, then improve it incrementally. — ML in Production Interview.
  5. Build useful data plumbing early: Looking back on startup work, he often wished he had solved basic data-pipeline and warehouse plumbing earlier, while recognizing the right timing depends on the company. — The Right Track Interview.
  6. Hire problem-solvers, not tool evangelists: For a startup, machine learning should be one tool among many: hire people who start with the business need instead of prescribing ML for every problem. — The Right Track Interview.
  7. Solve code execution before orchestration: Modal grew from his attempt to revisit Luigi-style workflow scheduling; he concluded that a good scheduler first needed a better underlying code-execution layer. — Barrchives Interview.
  8. Use founder experience carefully: Modal’s early product direction drew heavily on Bernhardsson’s prior infrastructure experience, but he later called its limited early customer conversations an unusual approach rather than general startup advice. — The Split Interview.
  9. State platform goals as goals: In a 2025 interview, he described ambitions to expand Modal into enterprise use cases and toward a fuller machine-learning development platform; these were goals at the time, not verified outcomes. — Barrchives Interview.
  10. Buy infrastructure below the scale threshold: His 2026 argument is that most software companies should buy mature non-core infrastructure; operating at exceptional scale can change the build-versus-buy economics. — Software Companies Buying Software.
  11. Pick difficult but tractable problems: He favors startup problems that are challenging enough to create value but bounded enough for a team to make meaningful progress within a few years. — Software Engineering Daily Transcript.
  12. Disqualify unsuitable design partners: Customer discovery should also identify prospects whose edge requirements would pull an MVP away from the intended market. — The Split Interview.
  13. Test AI investment against paid demand: He distinguishes revenue from users who value AI products from the harder question of whether infrastructure returns will justify very large capital commitments. — The Split Interview.
  14. Raise capital with the next round in mind: He warns that taking more funding than needed can raise the bar for a future round, and founders should account for investors’ different incentives when deciding how much to raise. — The Split Interview.

On Machine Learning and Data Science

  1. Treat ML as an enhancer, not always the core: In his 2016 product-stage analysis, machine learning often improved an existing product rather than enabling its basic value proposition. — When Machine Learning Matters.
  2. Build table stakes before sophisticated ML: Early in a product’s life, completing basic features and workflows can have a higher return than adding sophisticated machine learning. — When Machine Learning Matters.
  3. Pair ML insight with durable advantages: A model insight can spread as research is published or rediscovered; he advises combining it with harder-to-copy assets such as data, users or contracts. — When Machine Learning Matters.
  4. Distinguish data-science career paths: He distinguishes business-insight work from production ML engineering; the skills and daily problems of the two paths differ. — Interview with a Data Scientist.
  5. Ground business insights in experiments and users: For business-insight roles, he emphasizes informed experiments and empathy for user behavior over using advanced models for their own sake. — Interview with a Data Scientist.
  6. Keep production ML close to implementation: Production ML practitioners need backend and data-engineering fluency; a research-to-implementation handoff can slow iteration and misalign incentives. — Interview with a Data Scientist.
  7. Learn statistics more deeply: Looking back on his own STEM education, he wished he had studied more statistics; this is personal career advice, not a universal prerequisite. — Career Advice.
  8. Avoid a separate ML implementation queue: He argues that separating model research from implementation costs iteration speed, so the people building models should stay close to shipping them. — ML in Production Interview.
  9. Embed ML practitioners in product teams: He has seen isolated ML teams propose solutions nobody requested; embedding practitioners in product teams ties their work to the team’s priorities. — ML in Production Interview.
  10. Remember how immature data tooling was: Recounting his early Spotify years, he said data teams lacked established patterns and waited hours for simple Hadoop queries; later tools such as Snowflake and dbt improved the experience. — The Right Track Interview.
  11. Make data engineering less about plumbing: His aspiration was for data platforms to remove enough infrastructure work that a separate data-engineering specialty might become less necessary; he presented this as a prediction, not a completed change. — The Data Jobs to Be Done.
  12. Focus data work on business applications: Rather than treating the movement of data as an end in itself, he wants practitioners to start with the business need and build useful applications. — The Data Jobs to Be Done.
  13. Build tools for data teams’ real bottlenecks: When leaving Better in 2021, he identified repetitive infrastructure work, YAML configuration and fragmented workflows as reasons to prototype better tools for data teams. — What’s Erik Up To?.
  14. Specialize where the team has an advantage: In describing custom-model customers, he said teams can concentrate on training models and building applications while a specialist platform handles scaling and infrastructure. — MAD Podcast Interview.

On Developer Productivity and Tooling

  1. Optimize for a fast feedback loop: His developer-experience goal is to make cloud code execution fast enough that iterating on it feels close to working locally. — MAD Podcast Interview.
  2. Account for developer-experience costs: He argues that difficult developer workflows impose real costs on customers and can become a competitive weakness for infrastructure vendors. — Developer Experience as a Competitive Advantage.
  3. Treat future tooling as a forecast: In 2020, he predicted that developer tools available by 2030 would make many contemporary tools feel dated, while acknowledging how hard such forecasts are. — Developer Experience as a Competitive Advantage.
  4. Buy what is not core business logic: His vendor thesis is that companies should consider buying mature tools for non-core capabilities rather than repeatedly building and maintaining them in-house. — Developer Experience as a Competitive Advantage.
  5. Remove infrastructure chores from data work: He criticized the need for data practitioners to master Kubernetes, Docker and Terraform just to ship business logic, and wanted better platform abstractions. — The Right Track Interview.
  6. Become self-sufficient across the stack: He values engineers who can deliver useful work across system boundaries without waiting on multiple other teams, because fewer handoffs speed iteration. — Career Advice.
  7. Understand the developer-tool flywheel: His 2020 thesis is that cheaper software development can unlock more demand for engineers and, in turn, more tools that make engineers productive. — Developer Experience as a Competitive Advantage.
  8. Compound productivity improvements: Looking back over two decades, he argued that many incremental changes in tools and processes added up to orders-of-magnitude gains in what software engineers could build; the exact 10x-per-decade figure is not established. — Reorganization of the Factory.
  9. Blur the local-cloud boundary: He wants cloud execution to feel like a single development environment with local work; that requires low-latency feedback rather than a slow deploy-and-wait loop. — MAD Podcast Interview.
  10. Build abstractions around the workload: To make Modal’s execution layer fit its workloads, his team built its own file system and scheduler rather than relying on Kubernetes and Docker as the user-facing model. — MAD Podcast Interview.
  11. Choose the right abstraction trade-off: A useful abstraction can give up a little low-level control to remove substantial complexity; one that strips away too much capability is a poor trade. — Software Engineering Daily Transcript.
  12. Pool bursty workloads: Modal’s argument for shared compute is that many customers’ variable workloads can be pooled to improve utilization and support usage-based pricing. — Software Engineering Daily Transcript.

On Career and Learning

  1. Keep learning in a role: He suggests changing course when a job stops providing meaningful learning, framing learning as an important career criterion. — Career Advice.
  2. Work with people you can learn from: For him, a valuable colleague is not merely impressive but someone whose experience and habits help him learn. — Career Advice.
  3. Separate learning from wanting credentials: Before committing to a new skill, he asks whether he enjoys the learning process or mainly wants the identity of already having learned it. — Career Advice.
  4. Use side projects when they energize you: He describes after-hours projects as worthwhile when they are enjoyable and teach something, not as a mandatory career obligation. — Career Advice.
  5. Show your work publicly: He recommends publishing useful code and contributing to other projects as ways both to learn and to demonstrate what you can build. — Interview with a Data Scientist.
  6. Practice deeply to build expertise: His career advice emphasizes sustained, intensive practice as a route to skill, while the extreme hours he described are his personal view rather than a universal prescription. — Interview with a Data Scientist.
  7. Use a quantitative background as leverage: He says a physics background and love of mathematics helped him approach technical problems that felt difficult to many programmers. — ML in Production Interview.
  8. Learn by building in an emerging field: Spotify hired him in 2008 to work on recommendations despite limited formal ML experience; because the field’s tooling was immature, his coding background helped him build what he needed. — ML in Production Interview.
  9. Work at an ecosystem layer: His 2026 essay argues that a shared software layer can magnify one engineer’s impact across many companies rather than only one employer; this is an argument about potential leverage. — Software Companies Buying Software.
  10. Value mid-career operating judgment: Reflecting on founding Modal in his late thirties, he says experience with hiring, management and product decisions helped him avoid learning every business skill from scratch. — The Split Interview.

On Hiring and Management

  1. Hire curiosity about business and product: In describing Better’s data-team hiring, he emphasized curiosity about how the business works and an entrepreneurial drive to investigate numbers. — We’re Hiring at Better.
  2. Treat interviews as noisy signals: He argues that a few algorithm questions cannot reliably reveal a candidate’s future performance; interviews are a limited prediction tool. — Interviewing Is a Noisy Prediction Problem.
  3. Avoid pressuring candidates with exploding offers: In his 2016 recruiting essay, he said he did not put deadlines on offers and objected to pressure tactics that force candidates to decide too quickly. — Exploding Offers Are Bullshit.
  4. Consider paying for exceptional talent: When building Better’s technology team, he chose to pay near the high end of the market because he expected stronger output to offset the extra cost. — Headcount Goals and 10x People.
  5. Decentralize product backlogs: He argues that cross-functional teams with embedded engineers can work directly on business needs and iterate faster than a single centralized engineering queue. — Reorganization of the Factory.
  6. Use no-code for the right reason: He supports self-service when it improves iteration, but warns against adopting tools solely to avoid involving engineers in work that would benefit from their participation. — Reorganization of the Factory.
  7. Recognize elastic demand for software: His 2020 economic thesis is that productivity can lower the cost of building software and make previously uneconomic applications worthwhile; he presents growing engineer demand as a forecast. — Reorganization of the Factory.
  8. Avoid falling behind on engineering productivity: He argues that companies that underinvest in tools, engineers and decentralized workflows risk slower iteration relative to peers that improve those capabilities. — Reorganization of the Factory.

Learn more:

  1. ML in Production — Top ML Resources: Interview with Erik Bernhardsson
  2. Erik Bernhardsson — Miscellaneous Unsolicited (and Possibly Biased) Career Advice
  3. Heavybit, The Right Track — The Argument for Less Specialization with Erik Bernhardsson
  4. Amplify Partners, Barrchives — Solving Technical Problems to Help AI Teams Move Up the Stack
  5. Erik Bernhardsson — When Machine Learning Matters
  6. Peadar Coyle — Interview with a Data Scientist: Erik Bernhardsson
  7. Erik Bernhardsson — What’s Erik Up To?
  8. dbt Labs, The Analytics Engineering Roundup — The Data Jobs to Be Done
  9. Modal Labs — Modal Labs Announces Series A Financing Round
  10. Matt Turck, The MAD Podcast — Can AI Infrastructure Work Like Magic?
  11. Erik Bernhardsson — Developer Experience as a Competitive Advantage
  12. Erik Bernhardsson — About
  13. Erik Bernhardsson — Exploding Offers Are Bullshit
  14. Erik Bernhardsson — Interviewing Is a Noisy Prediction Problem
  15. Erik Bernhardsson — We’re Hiring at Better
  16. Erik Bernhardsson — Headcount Goals, Feature Factories, and 10x People
  17. Erik Bernhardsson — Giving More Tools to Software Engineers
  18. Erik Bernhardsson — Software Companies Buying Software
  19. Software Engineering Daily — Modal and Scaling AI Inference Transcript
  20. The Split — Building AI-Native Infrastructure for Developers
  21. No Priors — Erik Bernhardsson on Creating Tools That Make AI Feel Effortless