LLenny's Podcast
← All episodes
Howie Liu (co-founder and CEO)31 August 2025

How we restructured Airtable’s entire org for AI

7Frameworks
14Insights

Frameworks in this episode

Mindset3 steps

Extract the Chain-of-Thought, Not the Recommendation

Treat every advisor as an LLM: mine their reasoning, not their verdict, because their answer is trained on a different corpus.

Read →
Leadership4 steps

Fast-Thinking / Slow-Thinking Org Split

Split product org into a weekly-shipping AI group and a deliberate-infrastructure group so both speeds coexist.

Read →
Productivity3 steps

Greedy-but-Smart Compute Allocation

Throw hundreds of dollars of inference at high-value problems — the value-to-cost ratio is absurd in your favor.

Read →
Innovation4 steps

Play-First AI Fluency

Build real AI intuition by playing — invent fun side projects, use everything, and share the artifact not the doc.

Read →
Leadership4 steps

The Full-Stack Role Collapse

In the AI era every role needs a minimum baseline in the adjacent two — deep in one, dangerous in the rest.

Read →
Strategy4 steps

The Refounding Test

Ask how you'd rebuild AI-native from scratch — then decide whether your legacy asset helps or you should sell.

Read →
Innovation3 steps

Vibes Before Evals

For a genuinely new AI feature, start with open-ended vibes testing; add evals only once the use-case cluster is clear.

Read →

Insights & moments

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

Myth Buster· 1

Myth Buster1:03:30

For Novel Products, Start With Vibes — Not Evals

Countering the popular 'get good at evals' advice, Howie says for a completely novel product experience you should start with vibes, not evals. Evals require you to define what 'good' looks like in advance, which constrains you before you've discovered the useful cluster of use cases. Only after you've converged on the form factor do evals become useful for programmatically iterating toward improvement.

  • For a novel form factor, throw stuff at the wall ad hoc rather than defining evals up front
  • Evals require pre-defining what 'good' looks like, which can constrain discovery too early
  • First discover empirically which use cases the capability handles well
  • Once the form factor has converged, evals become useful for measured, AB-tested iteration

“you should actually not start with Evals and you should start with Vibes, right?”

Howie Liu · 1:04:00

“EVLs can constrain you too early.”

Lenny · 1:07:30
#evals#ai-development#product#iteration

Hot Take· 3

Hot Take08:00

Every Software Product Has To Be Refounded for AI

Howie argues that unlike the one-time shift from desktop to mobile or on-prem to cloud, AI is evolving so rapidly that every new model release implies novel form factors and UX patterns. That's why CEOs are becoming individual contributors again — to stay continuously relevant, founders have to get back into the details rather than managing from a 10,000-foot view.

  • AI is not a one-time, predictable form-factor change like mobile or cloud
  • Each new model capability implies new UX patterns that must be invented
  • You can't refine product-market fit in this era from a high-level view
  • This is why CEOs are getting back into the weeds, coding and building again

“every certainly every software product in my opinion has to be refounded because like AI is such a paradigm shift.”

Howie Liu · 11:00
#ai-strategy#leadership#product#iceo
Hot Take35:30

If You Can't Refound Your Company for AI, Sell It

Howie says a pre-GenAI company must take a clean-slate approach: ask how you'd execute the same mission if founding from scratch today with a fully AI-native approach. If your legacy assets don't give you a genuine advantage over starting fresh — and you can't honestly say you're better off with them — then you should find a buyer and go start the next incarnation of the mission.

  • Don't just bolt AI features onto a marketing site and call it done
  • Ask how you'd execute the mission if founding a new AI-native company from scratch
  • Honestly assess whether your legacy product is a useful building block or dead weight
  • If you're not genuinely better off with what you have, sell and restart the mission

“then I think you should sell, right? Like you should find a buyer for that company”

Howie Liu · 36:00
#ai-strategy#leadership#founders
Hot Take1:17:00

Treat Advice Like an LLM's Chain of Thought

Howie's meta-learning: don't just accept a recommendation from smart people, because everyone is like their own LLM trained on a different corpus of experience. Instead, inspect the reasoning behind the advice — the 'why' — the way you'd inspect a reasoning model's chain of thought. The recommendation itself may not be one-size-fits-all, but the reasoning is transferable and far more informative.

  • Everyone is like their own LLM, trained on a different corpus of experience
  • Don't blindly follow a recommendation — the giver's priors differ from yours
  • Inspect the 'why' behind advice like a reasoning model's chain of thought
  • The reasoning transfers even when the specific decision doesn't (e.g. Airbnb eliminating PMs)

“we're all our own LMS right and like we all have different training from a different corpus of data”

Howie Liu · 1:18:00

“that like chain of thought like why did you recommend this is actually more informative than the actual like just do this recommendation, right?”

Howie Liu · 1:18:30
#decision-making#leadership#mental-models

Explainer· 3

Explainer18:30

Restructuring Into Fast-Thinking and Slow-Thinking Groups

After several reorgs, Airtable split into a 'fast thinking' group (the AI Platform, shipping jaw-dropping new capabilities near-weekly) and a 'slow thinking' group for deliberate, premeditated bets like infrastructure (HyperDB handling hundreds of millions of records). Howie says the two complement each other: fast execution creates top-of-funnel excitement, while slow thinking lets those adoption seeds grow into durable enterprise deployments.

  • Feature-owned teams think incrementally; mission-owned groups make holistic bets
  • Fast-thinking (AI Platform) ships awesome new capabilities on a near-weekly basis
  • Slow-thinking handles deliberate infrastructure bets like HyperDB, which can't ship in a week
  • Fast creates top-of-funnel excitement; slow converts it into durable, expanding growth

“the fast thinking uh group which officially is called AI platform uh but it really means like we want to just ship a bunch of…”

Howie Liu · 21:30

“you need fast and slow thinking and the common sense uh to operate right like as a human”

Howie Liu · 22:00
#org-design#ai-strategy#reorg#product
Explainer29:30

The Best Way To Deliver AI Value Is Experientially

Howie argues AI value is best conveyed by letting people try the product, not by telling them in a deck. He cites ChatGPT as arguably the most successful PLG product of all time — reaching a claimed 700 million weekly users in under three years — precisely because anyone could just sign up and try it. His personal goal is to shift attention back to builder-led, experiential adoption.

  • Sales-led motions work for some AI products (Palantir, Harvey), but experiential access is more powerful
  • ChatGPT is arguably the most successful PLG product of all time
  • Its frictionless 'just try it' access drove an insane ramp curve in under three years
  • Airtable made its agent the default way of doing everything — the app is an artifact the agent manipulates

“the the best way to get AI value out there is experientially”

Howie Liu · 30:00

“10% of humans on earth use it weekly.”

Lenny · 30:30
#plg#ai-strategy#product#growth
Explainer36:30

Airtable's Building Blocks as a DSL for AI Agents

Howie explains why generating business apps from scratch with an agent is unreliable — bugs, security issues, and context collapse as complexity grows. Instead, Airtable's reliable primitives (real-time CRUD, view types, layout engine, automations) act like a more expressive domain-specific language the agent can assemble, rather than writing SQL, HTML, and JavaScript from scratch. Non-technical users can also fall back to the GUI, unlike opaque code from v0 or Bolt.

  • Agents writing every business app from raw code hit bugs, security issues, and context collapse
  • Airtable primitives act as a more expressive DSL the agent assembles instead of raw code
  • Combines reliable Lego pieces with agentic assembly for the 'best of both worlds'
  • Non-technical users can fall back to the GUI, unlike opaque v0/Lovable/Bolt output

“the Air Table pieces in our Lego kit today can be used by this agent as almost like a more expressive DSL like a domain…”

Howie Liu · 38:00
#ai-agents#vibe-coding#product#airtable

Story· 3

Story04:30

The "Airtable is Dead" Viral Tweet — What Actually Happened

Howie recounts the viral tweet, allegedly from someone at CB Insights, claiming Airtable was underwater and had raised far more than it was worth. He says essentially none of it was true — the numbers were off by a strong multiple. It gained real legs after being discussed on the All-In podcast, which later issued a correction on the figures.

  • The viral tweet's revenue and growth-rate numbers were wrong by a strong multiple
  • The same account had previously posted an unsubstantiated 'Flexport is dead' takedown
  • The All-In podcast amplified it, then issued a follow-up correction weeks later
  • The episode became a proxy for a broader theme about highly-valued decacorns in the market reset

“So very I basically none of it was true.”

Howie Liu · 05:00

“What's that line about how a lie gets around the world some number of times before truth has even has time to get out of…”

Lenny · 06:30
#airtable#social-media#startups#reputation
Story13:30

Being the #1 Most Expensive AI User at Airtable — On Purpose

Howie takes pride in being the single most expensive inference-cost user of Airtable AI — at one point globally across all customers. He describes being 'intentionally wasteful,' spending hundreds of dollars running LLM map-reduce over a year of sales-call transcripts to extract insights. The value, he argues, is like a brilliant chief of staff or a million-dollar consulting engagement, making the cost-to-value ratio absurd.

  • Howie is the #1 inference-cost user of Airtable AI, at one point globally
  • He runs LLM map-reduce over long sales-call transcripts to extract insights
  • Spending hundreds of dollars is trivial versus the strategic value of the insights
  • It's equivalent to work you'd pay a consulting firm millions for

“I take pride in being the number one most expensive in inference cost user of Air Table AI.”

Howie Liu · 14:00

“more people should be like aggressively throwing comput cycles at these very high value problems”

Howie Liu · 16:00
#ai-strategy#inference-cost#productivity
Story48:30

Permission To Cancel a Week of Meetings and Just Play

To get his team to internalize AI, Howie leads by example — sharing links to Replit landing pages or deep-research reports instead of writing docs — and explicitly gives people permission to block out a full day or week, cancel every meeting, and play with any AI product that might be relevant to Airtable. He frames it as the 'ultimate excuse': you can say the CEO told you to do it.

  • Howie shares actual built artifacts (a Replit landing page, a deep-research report) instead of docs
  • He explicitly encourages canceling all meetings for a day or a week to play with AI
  • The point is curiosity-driven play, not box-checking a task
  • Play is framed as both more energizing and a better way to learn

“if you want to cancel all your meetings for like a day or for an entire week and just go play around with every product…”

Howie Liu · 49:00
#leadership#ai-adoption#culture#learning

Takeaway· 4

Takeaway12:00

You Can't Taste the Soup Without Helping Make It

As 'chief taste maker,' Howie says you now can't judge an AI product from the outside — you have to be in the details to understand the solution space. AI isn't something you can evaluate from a screenshot or demo video; you have to play with both the packaged product and the underlying model primitives via API or chat to understand the new ingredients.

  • Being the chief taste maker requires participating in creating the product, not just reviewing it
  • AI must be played with directly — screenshots and recorded videos aren't enough
  • Push the underlying models to their boundaries via API and chat interfaces
  • It's like a chef who gained new ingredients but must get comfortable with them first

“it's actually now also hard to taste the soup without participating in like at least some part of creating the soup, right?”

Howie Liu · 12:00
#ai-strategy#product#leadership
Takeaway16:30

The Barbell Approach: Cut Standing 1-on-1s

Howie cut his default one-on-one roster because standing meetings crowded out timely, insight-driven conversations. He favors a barbell: most meetings should be urgency-driven and seeded with real alpha, while relationship-building happens as a genuine, unstructured lunch or coffee walk every month or two — not a forced weekly ritual.

  • Standing one-on-ones preclude engaging on timely, high-value topics
  • The best meetings are urgency-driven and seeded with a real insight or alpha
  • Relationship time should be high-quality and unforced — a longer lunch or walk
  • Do the free-form catch-ups in person roughly once every month or two

“I actually cut my one-on-one roster u by default”

Howie Liu · 16:30
#leadership#meetings#management
Takeaway40:30

Invent Fun Weekend Projects To Actually Learn AI Tools

Howie deliberately invents little side projects — like generating a comedic short with a scripted deep-research topic, AI dialogue, and HeyGen avatars — to force himself to use AI products in more than a twitch-click way. He says these projects only take about an hour once you're proficient, and they teach you not just the models but the product form factors that models can be placed into.

  • Side projects force real use of AI products, not superficial clicking
  • Once proficient, these projects take only about an hour
  • You learn the difference between understanding a model and understanding its product form factors
  • Part of the value is that it's genuinely fun and unpredictable

“these things only take you like an hour”

Howie Liu · 42:00

“you're not quite sure what you're going to get, you know, it's like a box of chocolates.”

Howie Liu · 43:30
#learning#ai-tools#productivity
Takeaway50:30

PMs, Engineers, and Designers Must All Become Full-Stack

Howie argues that success with AI tools comes down to individual attitude and polymathism — the hybrid unicorn types who cross over between PM, engineering, and design. He says you need a minimum baseline of competence in all three roles, then depth in your specialty. He cites Google's early technical PMs, designers at Apple, and engineers at Stripe (where the DRI isn't always the PM) as multidisciplinary precedents.

  • Success with AI tooling is driven by individual attitude and polymathism, not the role
  • You need a minimum baseline in all three of PM, engineering, and design, then go deep on one
  • A PM should become a hybrid prototyper with design sensibilities
  • Precedents: technical PMs at early Google, design-literate engineers at Stripe where the DRI isn't always the PM

“you need to start looking more like a hybrid PM prototyper who has some good design sensibilities”

Howie Liu · 55:30

“there's just a minimum baseline of like if you're any one of those roles, you need to be like minimally good at the other two.”

Howie Liu · 57:30
#product#hiring#skills#org-design