This article argues AI is the next cognitive tool, like writing or math, that expands human thinking. Chris Pedregal, founder of Granola, says AI's value is not replacing humans but pulling relevant context when needed. He favors high-frequency, high-quality professional tools, e.g., Granola (an AI note-taking app that enhances notes after meetings, not in real-time, to avoid distraction). He also mentions Anthropic (its Claude model is evaluated by Granola) and OpenAI (its GPT-4 is used by the founder with his kids).
This article explores the views of Granola founder Chris Pedregal on AI as a cognitive tool, arguing that AI is the next evolutionary frontier after writing, mathematical symbols, and data visualization. As an AI note-taking application, Granola extends human cognitive capabilities by transcribing m
Chris Pedregal is the founder and CEO of the AI note-taking app Granola. This issue revolves around a core thesis: AI is the next frontier of "cognitive tools" after writing, mathematical symbols, and data visualization. Its value lies not in replacing humans, but in expanding human cognitive capabilities — taking "externalized memory" and "dynamically generating relevant context" to their extremes. Pedregal uses Granola’s real-world case (meeting transcription + note enhancement) to demonstrate the feasibility of this path, and analyzes the competitive landscape between model providers and application layers, as well as how small teams can create outsized impact in the fast-iterating AI field.
Chris Pedregal argues that every revolution in cognitive tools throughout human history has essentially been about "externalizing what the brain needs to hold"—from writing, mathematical notation, data visualization, to computers, each expansion has extended humanity's "RAM." What makes LLMs unique is their ability to dynamically generate the context most relevant to the present moment**—something no previous tool could achieve.
Historical Trajectory: Pedregal highlights four key milestones:
Pedregal's Core Analogy: Traditional notebooks and pencils are "RAM expansion"—you don't have to keep everything in your head. The breakthrough of LLMs lies in their ability to pull the most relevant context to you at the right moment, dynamically generated. "Imagine if, during a meeting, a computer could instantly retrieve all relevant background information, making you incredibly smart in that moment—that would be an unbelievable unlock."
Falsification Condition: If future LLMs have large context windows but cannot effectively filter out information that is "truly relevant at this moment" (i.e., the "information overload" problem), then the vision of "dynamic context" will be significantly diminished.
Pedregal revealed a key early lesson from Granola: they initially tried to have the AI generate notes in real time during meetings (user types a keyword → presses Tab → AI writes a full sentence), but found this actually distracted users — they couldn't help but look at what the AI wrote, then revise it, only to realize they hadn't been listening to the other person at all. This mistake took six months to correct.
Core Interaction Model:
Core Product Philosophy: "Let users control the tool, not the tool control the user." Every decision revolves around this principle — even the seemingly simple choice of an "editor": most AI note-taking tools generate PDFs or emails that users cannot edit; Granola insists on an editable editor, allowing users to add their own judgments on top of the AI output.
Key Data Point: Granola is a Mac app rather than a web app or meeting bot. Initially it only supported Mac OS 13.4 (about 15% of Mac users at the time), but Pedregal considers this decision a "major blessing" — an app on the computer is more direct and controllable than a browser tab, and the way users interact with the product is therefore more "intimate."
Pedregal proposes an analytical framework to determine whether an AI application will be "eaten" by model providers: two dimensions — frequency of use (high vs. low) and requirement for output quality (high vs. low).
| Quadrant | Frequency of Use | Output Quality Requirement | Fate |
|---|---|---|---|
| Low frequency, low requirement | Low | Low | Eaten by general-purpose assistants (e.g., Claude, ChatGPT) |
| High frequency, high requirement | High | High | The domain of specialized tools (e.g., Granola) — this is the "specialized tool quadrant" |
| Low frequency, high requirement | Low | High | May be covered by general-purpose assistants, but the experience will not be optimal |
| High frequency, low requirement | High | Low | Easily replaced |
Core argument: "This is fundamentally not an intelligence problem, but a UI and user experience optimization problem." If you do something 500 times a day and need it done to the highest standard, a tool specifically optimized for that task will always outperform a general-purpose assistant — because a general-purpose assistant cannot achieve "perfection" in every detail.
Regarding the attitude toward model providers: Pedregal believes "this is the best thing" — intense competition among models benefits application-layer companies like Granola enormously. Granola does not rely on a single model but instead uses "whichever model is the best today," maintaining optimal performance through rapid evaluation and hot-switching.
On views of competition: Granola's biggest competitive concern is not existing competitors, but "startups that haven't launched yet" — those that can stand on the shoulders of the problems already solved by Granola and others, and execute faster.
Pedregal argues that AI application-layer companies can achieve "small teams, big impact", but only if they manage two modes well: the 'exploit mode' (knowing what to build, executing quickly) and the 'explore mode' (not knowing the answer, exploring first). Confusing the two is disastrous.
Specific approach:
Key lesson: Granola spent a full year before launch (only launching in May 2023), at a time when they were "already seven years late in the AI note-taking track." But Pedregal believes that if they had pushed out the real-time note generation version immediately, users would have "learned" that behavior pattern, and Granola would never have been able to pivot to the correct interaction model. "Protect your ability to change direction until you have high confidence in the right direction."
Further extrapolation on the "small team effect": Pedregal believes that in the future, there could be companies with $1 billion in revenue but only 20 employees. He gives a specific example: Granola just made its first customer experience hire. He believes that customer experience departments founded before 2025 and those founded after will be completely different in terms of team size and operating approach – the latter will make heavy use of AI tools, be smaller, and also more efficient.
A sober assessment of "data advantages": Pedregal argues that "data is not a moat" – because today's foundation models need very little data (e.g., 50,000 examples) to optimize for a specific use case, and 50,000 examples are not hard to obtain. What is truly difficult is "data that cannot be obtained" – but such data is rare.
| Position | Guest View | Key Data |
|---|---|---|
| Granola | Bullish (positive position building) | Launched May 2023; Mac app, initially only supported 15% of Mac users; Team <25 people; Users mainly AI practitioners, founders, and investors |
| Anthropic | Neutral (no explicit action) | Mentioned as one of the model providers; Granola will evaluate its new models and consider hot-switching |
| OpenAI | Neutral (no explicit action) | Mentioned as one of the model providers; GPT-4 voice mode is used by Pedregal personally to interact with his children |
| Socratic (Pedregal's previous startup) | Historical reference (not a positive position) | AI education app; Pedregal believes the education sector's "holy grail is one-on-one tutoring", but commercialization incentives do not align with social benefits |
1. "AI is the next cognitive tool frontier after writing, mathematical symbols, and data visualization" (Chris Pedregal) — Support: Every tool revolution has expanded humanity's ability to "externalize memory," and the unique feature of LLMs is that they can dynamically generate the "most relevant context at the moment," which no previous tool could do.
2. "Granola's core interaction model is 'post-meeting enhancement' rather than 'real-time generation' — because real-time generation is distracting and actually prevents you from focusing on the conversation." (Chris Pedregal) — Support: Granola spent six months trying to have AI write notes in real time, but found that users could not resist looking at the AI's text and editing it, resulting in them not listening to the other person. Eventually, they switched back to making the meeting experience like a regular notebook, with all AI polishing done after the meeting.
3. "Whether AI applications will be eaten by model providers depends on two dimensions: frequency of use and requirements for output quality. High frequency + high requirements is the 'professional tool quadrant'; general-purpose assistants can never outperform specialized tools in this quadrant." (Chris Pedregal) — Support: This is a UI problem, not an intelligence problem. A tool specifically optimized for a particular use case far exceeds the general-purpose assistant in terms of detail experience.
4. "Protect your ability to change direction until you have high confidence in the correct direction." (Chris Pedregal) — Support: Granola spent a full year before launch (in the AI note-taking space, it was "seven years late") because if they pushed out the wrong version first, the user behavior patterns learned would make it impossible for the team to pivot.
5. "A company with $1 billion in revenue might only need 20 employees — especially the customer experience department, which will look completely different for companies founded around 2025." (Chris Pedregal) — Support: Granola just made its first customer experience hire and believes that future customer experience teams will be smaller, more reliant on AI tools, and difficult to transition from "old" customer experience teams.
6. The "General-Purpose Assistant vs. Specialized Tool" Framework (Chris Pedregal's own taxonomy) — Two axes: usage frequency (high vs. low) × output quality requirements (high vs. low). Low frequency + low requirements → eaten by general-purpose assistants; high frequency + high requirements → territory of specialized tools. Key point: This is essentially a UI optimization problem, not an intelligence problem.
7. "Data is not a moat — today's foundation model only needs 50,000 examples to optimize for a use case, and 50,000 examples are not hard to obtain." (Chris Pedregal) — Support: Compared to the old machine learning paradigm that required millions of examples, the bar has now been significantly lowered. What is truly hard to obtain is "unobtainable data," but such data is rare.
8. "AI tools should make you more 'human' — they should handle all the tedious, repetitive work, but never outsource your judgment." (Chris Pedregal) — Support: Writing is thinking. If AI writes all the words for you, you lose the opportunity to think. Pedregal hopes the tool will "dynamically bring your personal life and the best external information you believe in before you, helping you think in real time."