Notes — Howie Liu on the IC CEO, Fast and Slow Thinking Teams, and Airtable’s AI-Native Refounding

Four questions [Adler frame]

Q1 — What is it about? Howie Liu, co-founder and CEO of Airtable, on how he operates as an IC CEO who returns to individual-contributor product detail; how he restructured Airtable’s EPD org into complementary fast and slow thinking groups; and how the AI-native refounding thought experiment functions as a strategic forcing function for any established product company.

Q2 — How is it argued? Through personal examples and Airtable-specific decisions: the EPD reorg, the immersive world generator demo Howie built personally, his use of LLM map-reduce on Slack transcripts, and the AI-native refounding question applied to Airtable’s Lego kit / DSL advantage. Draws explicitly on Kahneman’s fast/slow thinking as a conceptual frame for the org split. Argument is anecdotal and illustrative rather than formal.

Q3 — Is it true? The IC CEO thesis has intuitive force for product-technical companies: taste requires participation. The caveat is that it works primarily for founders with genuine design intuition — it is not a general prescription. The fast/slow thinking org split is empirically grounded in Airtable’s specific context. The vibes-before-evals stance is contrarian but internally consistent with Double Diamond product methodology — premature evals do constrain the solution space before use cases are understood.

Q4 — What of it? The AI-native refounding mental model is a useful forcing function for any established product team. The fast/slow thinking split resolves what is often framed as a false choice (ship fast vs. build durably) by making the split structural and symbiotic. The role-collapsing thesis extends the IC CEO insight beyond the CEO level to every function.

Glossary

IC CEO: A CEO who actively participates in individual-contributor-level product and design decisions — not just as reviewer but as co-creator. Howie’s synonym: chief taste-maker.

Chief taste-maker: To have product taste, you must “taste the soup” by participating in creating it, not merely evaluating finished output. Requires hands-on use of models, prototyping, and building.

AI-native refounding: Mental model and forcing question: “If you were literally founding a new company from scratch with the same mission, how would you execute on that mission using a fully AI-native approach?” Used to identify what to abandon vs. what to rebuild.

Fast thinking group: Airtable’s EPD sub-group (officially “AI platform”) focused on rapid, near-weekly shipping of jaw-dropping AI capabilities. Creates top-of-funnel excitement and maintains relevance.

Slow thinking group: Airtable’s EPD sub-group focused on deliberate, infrastructure-scale work (e.g. HyperDB — multi-hundred-million-record data store). Enables durable growth; converts fast-group adoption seeds into lasting deployments.

Role collapsing: The expectation that PM, engineer, and designer each develop a minimum viable baseline in the other two. Extended to non-product functions: marketing teams, AEs, and SEs all need to collapse roles to reduce dependency chains.

LLM map-reduce: Splitting a large corpus (e.g. all Slack messages or sales call transcripts) into chunks, running an LLM call on each, then aggregating results in a final LLM call. Expensive but strategically valuable.

Vibes-first evals: The stance that open-ended qualitative exploration (“vibes”) should precede formal evaluation criteria. Premature evals encode a definition of success before use cases are understood; vibes allow discovery first.

Lego kit / DSL advantage: Airtable’s existing no-code component library functions as a constrained DSL for AI agents building business apps — more reliable than pure LLM code generation, and a structural moat for the AI-native refounding.

§ IC CEO — chief taste-maker [§1]

Howie’s core claim: you cannot maintain genuine product taste by reviewing finished output alone. You must participate in making it. In the AI era specifically, this means: running models via API, building prototypes personally, using AI tools “literally hourly.” Howie built the immersive world generator demo himself and uses Airtable’s LLM map-reduce on company-wide Slack transcripts daily.

The practical expression: cut the standing one-on-one roster; shift to urgency-driven, alpha-informed meetings; cancel a full day or week of meetings to go build. Lead by example — share prototypes, not just directives.

Connects to Brian Chesky on Airbnb and the Art of Obsessive Craftsmanship (Founder Mode) but extends it to the AI era: the product experience is now in the model’s behaviour, not just the surface UX. You cannot taste that quality without running the models yourself. [§IC CEO concept page: IC CEO]

§ AI-native refounding thought experiment [§2]

The forcing question: “If you were literally founding a new company from scratch with the same mission, how would you execute on that mission using a fully AI-native approach? If you can’t, then you should find a buyer and then if you really care about this mission, go and start the next incarnation of it.” [?]

This is an unusually candid acknowledgement of the innovator’s dilemma applied inward. The clean-slate constraint forces honesty about what existing infrastructure is a moat vs. a liability.

Airtable’s answer: the no-code Lego kit is a reliable DSL for AI agents — the constrained primitives make AI-generated business apps more predictable than pure code generation. This is a structural moat that a new entrant would need years to replicate.

§ Fast and slow thinking teams [§3]

The EPD reorg (most recent iteration): two complementary groups rather than feature-area or pillar-based teams.

Howie draws the analogy explicitly from Kahneman’s Thinking, Fast and Slow — the team can internalise the mental model, not just the org chart label.

The symbiosis: fast thinking creates the narrative and top-of-funnel pull; slow thinking converts those initial adoption seeds into large deployments. The anti-pattern it addresses: applying a single cadence and accountability structure to all product work, which either slows the fast work or destabilises the slow work.

§ Role collapsing [§4]

T-shaped hybrid roles across EPD and beyond. Minimum viable baseline in each adjacent discipline:

• PM: can prototype, reads code, has design sensibility
• Engineer: can think about product and business requirements; some design judgment
• Designer: technically fluent enough to prototype in interaction design tools; understands model behaviour

Howie extends the thesis to non-product functions:

• Marketing: can execute a full campaign independently (copy, creative, performance)
• AEs: SE-fluent; can demo and build with the product without a separate SE dependency
• SEs: product-fluent enough to be the technical voice without a separate PM dependency

The unifying logic: outcome-orientation requires removing internal dependency chains. In a fast-moving AI market, handoffs are latency.

§ Vibes-first evals [§5]

Counterintuitive position: do not start with evals. Evals constrain the solution space before you know what success looks like. For a novel product experience or capability, start with open-ended exploration: “throw stuff against the wall, try different prompts and see how well it does.”

Sequence: vibes → use-case crystallisation → evals. Evals are useful once you have a sense of the cluster of useful use cases — at that point you can start programmatically measuring improvement.

Cross-wiki resonance: Evals covers the technical mechanics and the case for systematic evaluation. Howie’s additive take is temporal: when to introduce evals matters as much as how.

§ LLM map-reduce and personal AI use [§6]

Howie self-reports as #1 most expensive inference user across Airtable’s entire customer base. His daily practice: use Airtable’s LLM map-reduce (chunk → LLM call per chunk → aggregation LLM call) to synthesise company-wide Slack messages and sales call transcripts. Hundreds of dollars per run; compares the output to a very smart chief of staff who has read every sales call transcript in the past year.

The CEO-as-power-user thesis: the best driver of company-wide AI adoption is visible use and shared output. Not directives — examples. Cancel meetings, build prototypes, share deep research links.