← Back to list
Colossus (Invest Like the Best / Business Breakdowns)Podcast28 May 2024Source: joincolossus.comHost: Patrick O'Shaughnessy

Howie Liu - Building Airtable - [Invest Like the Best, EP.375]

In plain words

Howie Liu argues that the real bottleneck for AI is not model capability but embedding existing LLMs like GPT-4 into real business workflows, which could unlock trillions in GDP. He is optimistic about platforms like Airtable that allow deep customization. Key holdings: Salesforce (platform model, lets customers customize data), OpenAI (GPT-4 provider), Netflix (Airtable customer building content systems).

AI SummaryAI-generated · may contain errors · verify against the original

This report discusses Airtable co-founder and CEO Howie Liu’s strategy for no-code platforms in the AI era. The core argument is that Airtable, as a platform serving over 300,000 organizations, is integrating AI and LLM technologies while maintaining an intuitive building experience, enabling users

~13 min full read · 8 sections
Deep Analysis

At a Glance

Howie Liu is the co-founder and CEO of Airtable, a company founded in 2012 that now serves over 300,000 organizations, of which approximately 100,000 are paying customers. More than half of the Fortune 500 are paying users. The main theme of this episode: the strategic positioning of no-code platforms in the AI era — embedding LLM capabilities into existing platforms rather than building AI products from scratch. The most impactful judgment of the entire episode: Howie Liu believes that even if LLM capabilities were frozen at today's level (GPT-4/Claude 3 tier), simply embedding them better into real business workflows could still generate trillions of dollars in GDP — the real bottleneck is not model capability, but the speed of enterprise behavior change and process implementation.


Theme 1: Horizontal Platforms vs. Vertical Software — The "Third Bucket" — Both Wide and Deep

Howie Liu categorizes software into three types: wide and shallow (e.g., Office suites), narrow and deep (e.g., Procore/vertical healthcare solutions), and "wide and deep"—the third bucket where Airtable resides.

  • Mechanism Breakdown: The "wide and deep" model is possible because Airtable is not a pre-packaged application but rather provides "Lego blocks"—data models, business logic, and interface/workflow layers—allowing each customer to build their own bespoke deep applications using these blocks. Netflix uses Airtable to build its content production system; Airtable does not dictate how to do it but offers sufficiently good building blocks.
  • Historical Context: This concept originates from Liu's "lightbulb moment" at Salesforce—Salesforce beat Siebel/SAP not because it came with the best pre-built CRM, but because it built a metadata-driven platform that allowed each customer to customize data schemas, object types, and page layouts. However, Salesforce/Jira remain difficult for ordinary users to customize, and Airtable lowers that barrier.
  • Alignment with LLMs: LLMs are also "both wide and deep"—capable of reasoning across broad topics rather than performing a single classification task. Liu argues that cramming an LLM into a narrow application is akin to "confining the beautiful intelligence of humans to repeating a rigid task every day." The platform should let each customer, each business line, and each individual design how to leverage LLM capabilities in their own work scenarios.

Theme 2: The Future of Applications in the AI Era — Fully Customizable, but Requiring "Guidance"

Liu agrees with the assessment that "future applications will be highly customized," but points out that the key bottleneck is not technology, but the "imagination gap" between a blank canvas and a finished product.

  • Insights from Salesforce: The winner in the CRM market was not the company offering the best pre-configured solutions, but the platform that allowed customers to customize on their own. However, customization capability alone is insufficient — if users do not know how to build, the platform is merely "a bucket of Lego bricks."
  • Evolution of Airtable: It has evolved from a pure Lego kit ("a big bucket of red bricks") to offering templates, blueprints, and even semi-finished solutions (e.g., Salesforce's sales CRM/support CRM templates). "Finding the fusion point between platform and solution" has been the biggest product philosophy shift since the company's founding.
  • AI's Dual Role:
  • AI builds applications: The user says, "I work at Nike, and I need a XX process," and AI automatically infers and builds the application based on public information (10-K filings, industry knowledge).
  • AI embedded in workflows: Once the application is built, AI can be inserted into multiple stages of the process — such as automatically generating marketing campaign strategies, storyboards, and concept maps, enabling humans to perform better on that foundation.
  • Falsification condition: If AI cannot significantly reduce the "cognitive friction" from a blank canvas to a usable application, platform-based solutions may lose to vertical "out-of-the-box" solutions.

Theme 3: The "Quality Leap" in LLM Capabilities Matters More Than New Features

Liu argues that what LLMs currently lack most is not new capabilities (such as multimodality), but improvements in output quality, consistency, and ease of use—a leap from a "C student" to an "A+ student."

  • Capability Ladder: Using standardized test accuracy as an analogy—20% accuracy is nearly useless, 60% becomes interesting, 75-80% is equivalent to an average employee, and 95-99% represents top-tier performance. The current model's ability to perform better and more consistently on "the same type of problem" is more critical than adding new capabilities.
  • Prompt Engineering Bottleneck: Currently, a "prompt expert" is needed to extract the model's best performance. If future models can deliver "magical practical value on the first try" (rather than merely entertaining bedtime stories), real and lasting AI adoption will explode.
  • Reasoning Quality Over Multimodality: While multimodality (unified input/output for video, long documents, and audio) is indeed on the horizon, it can already be partially achieved by combining existing specialized models with platform data workflows. Improvements in reasoning quality represent a "profound change."
  • Historical Analogy: AutoGPT (a year ago) demonstrated the prototype of "chain-of-thought reasoning," but it gradually deviated from its original goal. Humans follow a similar process when conducting complex research projects (decomposing problems → searching → reading → synthesizing → drafting → critiquing → revising), only doing it better. Liu quotes a product lead at an AI company: "We may look back in a few years and realize that today we already have all the pieces for more advanced intelligence, but we just haven't learned how to assemble them correctly."

Theme 4: Airtable’s "Crucible Moment" — From Growth Sprint to Organizational Reshaping

Liu recounts Airtable’s 12-year journey through four pivotal turning points, the most recent being 2023: amid dramatic macro shifts, the company proactively conducted two rounds of layoffs in exchange for sustained hiring capacity and organizational agility.

Phase Time Key Event Decision Logic
Product Validation 2012-2015 2.5-3 years building the product, started around the same time as Figma Market timing (browser performance + maturity of bottom-up adoption model) + product details surpassing competitors
Monetization Validation 2015-2016 First $10K client → $500K → $1M → $10M acceleration Proved that pure software can generate genuine willingness to pay; Liu noted, "Software is so abstract yet so powerful"
Scaling Financing 2017-2019 Thrive/Benchmark/Coatue led the first unicorn round Shifted from "independent darling" to mainstream; team expanded from 30-40 people
Organizational Reshaping 2020-2023 COVID shock → demand pulled forward → growth surge → two rounds of layoffs in 2023 After ballooning from 100+ to 1,000+ people, proactively laid off to trade for sustained hiring capacity and agile execution
  • Deep logic behind the layoff decisions:

1. Sustained hiring rights: Maintaining the original cost structure would have required a hiring freeze; after layoffs, the company could instead continue bringing in fresh talent. "It is arrogant to think a company will never need to hire new people."

2. Organizational agility: AI execution requires "small, tight-knit teams working directly with customers" — Amazon’s two-pizza rule and Apple’s Mac team of just 50 people (while the Lisa project with thousands failed) serve as examples.

3. Financial outcome: After the layoffs, the company has achieved positive cash flow, with nearly $1B in cash on hand and still growing.

  • Reader’s note: Here, Liu offers a defensive narrative for his layoff decisions, emphasizing that "this is for the company’s long-term benefit" — from the perspective of a position holder, readers should independently assess the impact of layoffs on employee morale and long-term innovation capacity.

Theme 5: The Role of the Capital Allocator — Not an Excel Game, but an "Executive Producer"

Liu rejects reducing capital allocation to a "numbers game," arguing that the best scale operators remain deeply involved in the details — like a film producer rather than a financial analyst.

  • Counterexample and Positive Example: The largest, best-resourced teams do not always win (Lisa vs. Mac); Elon Musk sleeping on the factory floor, Mark Zuckerberg personally refining the product experience with Oculus engineers — a coexistence of detail orientation and big-picture vision.
  • Liu's Self-Positioning: Not "just filling numbers into different boxes and calling it a day," but an "executive producer" — ensuring the team operates harmoniously, the story direction is correct, and the "feel" is right. Even when the company has reached hundreds of millions in revenue, it is still necessary to maintain a "feel" for micro-level details.
  • On "Drafting" Future Companies: Liu believes the most undervalued "draft picks" are not pure technology companies, but AI-first industry operators — a new generation of media companies, retailers, construction firms, etc., that understand the industry, dare to use AI as a technological lever, and are willing to execute aggressively. The third wave of AI talent flow will evolve from "foundation models → SaaS founders → industry operators."

Mentioned Positions

Position Guest Sentiment Key Data
Salesforce Positive (as a platform model exemplar) Revenue in the tens of billions; metadata-driven platform allows each customer to customize their data schema
Netflix Positive (as an Airtable customer case) Uses Airtable to build content production systems
Slack Neutral (AI features rated "okay") AI can summarize threads, but Liu believes "strategic importance is low"
Microsoft Positive (as a top draft pick in the AI era) Plays a key role in AI
Google Positive (same as above) Same as above
OpenAI Positive (as a model provider) GPT-4 level models; Sora video generation model
Anthropic Positive (as a model provider) Claude 3
Meta (Llama 3) Positive (as an open-source model provider) Llama 3
Atlassian (Jira) Positive (as a platform model exemplar) Flexible and customizable platform
Procore Neutral (as a vertical software case) Construction industry vertical solution
Viva Neutral (as a vertical software case) Healthcare industry vertical solution
Adandy Neutral (as a vertical software case) Dental CRM, a relatively shallow vertical solution
HeyGen Positive (as an AI fast-growing SMB case) Significant growth from small businesses
SpaceX Positive (as an exemplar of detail-oriented operations) Elon Musk sleeping on the factory floor
Tesla Positive (same as above) Same as above
Pixar Positive (as an exemplar of technology and art integration) Fusion of technology + filmmaking + storytelling
ILM/Lucasfilm Positive (same as above) Same as above
Crowdflower (predecessor of Scale AI) Positive (personal experience) Founder Lucas wrote Liu a personal check to extend the startup runway

Judgments Worth Remembering

1. "Even if we froze today's LLM capabilities, trillions of dollars in GDP could still be created" (Howie Liu) — The bottleneck is not model capability, but how to embed existing models into real business processes. The speed of enterprise behavior change and process implementation is the true limiting factor.

2. "Three categories of software: wide and shallow, narrow and deep, wide and deep" (Howie Liu) — Airtable's third bucket (horizontal platform + deep customization) naturally aligns with the "wide and deep" capabilities of LLMs. Stuffing an LLM into a single application is equivalent to "confining human intelligence to repeating a rigid task every day."

3. "Salesforce won not because it came with the best CRM pre-installed, but because it built a metadata-driven platform" (Howie Liu) — Every customer should be able to customize data schemas, object types, and page layouts. Airtable lowers this barrier, enabling non-technical users to do the same.

4. "The future of AI applications: not AI as a single feature, but AI embedded in every step of the process" (Howie Liu) — AI builds applications (from blank to completion) + AI embedded in workflows (auto-generation, review, iteration). The platform should let customers design how to leverage AI themselves, rather than offering pre-packaged "AI features."

5. "The leap in LLM capability from a C student to an A+ student is more critical than adding new features" (Howie Liu) — 20% accuracy is useless, 60% is interesting, 80% equals an average employee, and 95-99% is top-tier. Improvements in reasoning quality and consistency will unlock greater value than multimodality.

6. "Layoffs twice were to enable continuous hiring — believing a company never needs to hire new people is arrogance" (Howie Liu) — Proactive layoffs trade for sustained hiring capability and organizational agility. AI execution requires "two-pizza teams," not thousand-person armies (Apple's Mac team had 50 people vs. the Lisa project's thousands).

7. "Capital allocation is not an Excel game, but an executive producer" (Howie Liu) — The best scale operators (Elon Musk, Mark Zuckerberg) remain deeply involved in details. Detail orientation + fast feedback loops win more often than the largest resource teams.

8. "The third wave of AI talent flow: foundation models → SaaS founders → industry operators" (Howie Liu) — The most undervalued "draft pick" is the AI-first industry operator: a new generation of media companies, retailers, and construction firms that understand the industry, dare to use AI as a technology lever, and are willing to execute aggressively. Pixar is the quintessential fusion of technology and storytelling.