I’ve spent the past 6 months diving deep into how the human brain learns. I’ve talked to some of the best Jeopardy players in the world, consumed hundreds of hours of learning content, and built two different products to help people learn.

The way I approach learning in my daily life has changed completely.

In what I now know is a selfishly beneficial form of active recall, I wanted to write up my biggest takeaways and lessons.

Common misconceptions

A few of my foundational beliefs about how the brain learns turned out to be completely incorrect.

1. Re-reading is helpful

I fell into the common trap of thinking that going back to re-read something would help me remember and understand it.

However, science shows that passive re-reading has little to no benefit in long term retention (source). It can be even worse than that, creating a “metacognitive illusion” where re-reading makes learners feel like they are mastering a topic just because it feels familiar (source).

The key word here is passive. Unlike passive re-reading, active recall is shown to dramatically increase long term retention. More on active recall below.

2. Memorization doesn’t matter, I can just look it up

I grew up in the era of Google and Wikipedia. I told myself that memorization wasn’t important because I could just look things up.

There are a couple major problems with this line of thinking:

• Science shows that memorization is the fundamental building block to deeper understanding. It’s important to learn the facts and definitions within a field to be able to then connect the dots and increase understanding of the topic (source).

• Even for pure memorization of facts, it can actually save time to memorize them. Using techniques I’ll discuss more in depth below, it can take a combined 5 minutes of studying throughout your whole life to memorize a fact (h/t source). So if you will spend more than 5 minutes looking up a fact throughout your entire life, it’s actually more effective to just memorize it.

3. You can’t reliably remember things, people’s brains work differently

Prior to this year, I thought the concept of “learning styles” i.e., audio, visual, kinesthetic meant that there was large variation in how people remember and learn things.

However, the theory of learning styles has been largely debunked, and we do actually have very clear understanding of how quickly the brain forgets things thanks to the well documented Ebbinghaus Forgetting Curve (which has been around over 100 years).

So, it turns out the way to get something into your head is quite simple: get your brain to actively recall that piece of information before you hit the next drop in the forgetting curve. This is known as spaced repetition, and the idea that you can just choose to put something in your brain forever is nothing short of magical.

How I think about learning now

Effective learning is all about active recall. This is the process of actually retrieving the information out of your brain. As I’ve practiced active recall, I’ve noticed you can actually feel your brain working when it’s pulling information out.

There are many forms of active recall - learning by doing, teaching, writing, the Feynman method, etc. Flashcards are the most direct way to do active recall. I’ve avoided flashcards like the plague for my entire life, but they truly are an extremely efficient way to get something into your head. When paired with spaced repetition they work scarily well, and honestly are even fun. The dopamine hit of getting cards right is real.

I also like Bloom’s Taxonomy as a way of thinking about building up my understanding of a topic. Memorization is the base which allows me to work my way up to deeper understanding, synthesis, and eventually creativity.

How I learn something new - tools and process

I believe that in the AI era, the ability to deeply understand and connect ideas from multiple areas is going to be more important than ever. Having a shallow understanding of a topic is no longer useful; only those who develop deeper insights will succeed.

Bringing this all together, here’s my process for learning a completely new topic.

Understand the building blocks

First, I want to understand the different components that make up the topic. I usually start by asking ChatGPT “What are the key concepts that a student would learn about XX topic?”

This gives me a breakdown of the key areas of the topic. I figure out which basic things I don’t understand, and ask for videos or articles I can read to understand them.

Study my areas of weakness, create active recall artifacts

While I’m reading/watching, I’m consistently thinking about which of these facts or ideas are things I want to remember.

I then upload the entire video/article into my preferred digital flashcard tool to automatically generate flashcards. We created Savvy to help with this (would be a shameless plug but we are winding down work on this), and I’m also testing out Remnote where I can both take notes and create cards. Anki is the tried and true tool in the space but I like being able to auto-create cards from content.

Practice active recall

You have to be consistent or spaced repetition doesn’t work. I try to review my cards every day. I find this practice also encourages my curiosity to continue to further explore the topic, add new cards and expand my knowledge.

To work my way up to deeper understanding and synthesis, I’ll talk to my friends about it and write about it.

That’s it! This isn’t something you want to over-complicate. It’s about simplicity and consistency. It’s not easy - you’ll feel your brain working hard as you do this, but it’s extremely rewarding. Short and very focused sessions are much more effective than hours of skimming. I’ve stopped spending time on things like highlighting, bookmarking, etc. and focused on the key things I actually want to learn.

I’ve loved the feeling of reading a book I really enjoy and weeks later being able to easily recall the key points I care about.

Want to go deeper?

My favorite reading and blog posts on the topic:

Michael Nielsen’s Augmenting Long Term Memory: https://augmentingcognition.com/ltm.html

on Spaced Repetition: https://gwern.net/spaced-repetition

Andy Matchusak on spaced repetition: https://notes.andymatuschak.org/Spaced_repetition_memory_system and ’s interview with him

Dwarkesh Podcast

Andy Matuschak — The reason most learning tools fail

Listen now

3 years ago · 36 likes · 2 comments · Dwarkesh Patel

The ultimate tech showdown

It’s now been more than 2.5 years since ChatGPT launched. One thing I’ve been watching closely is the effect on Google’s search dominance. Google has been the unchallenged leader in search for as long as I’ve been using a computer, so the idea of them getting unseated is fascinating.

At first, I didn’t think ChatGPT was eating into Google search at all. For the first year, ChatGPT didn’t have access to real-time information and seemed primarily used in ways distinct from a Google search—i.e., creative writing, coding, brainstorming, etc. There was talk of their user growth slowing during summer break, and Google’s search traffic continued to grow.

However, ChatGPT added real-time search in late 2023, and it clearly became a daily habit for hundreds of millions of users. Naturally, its users started performing searches traditionally done on Google—travel planning, how-tos, and even basic fact retrieval.

Anecdotally, I can say that I and many others now go to ChatGPT before Google for the majority of our queries.

And yet… Google’s search volume and revenue just continue to grow like crazy. They posted Q1 2025 advertising revenue of $66.89 billion, up 8.5% YoY and up almost 70% since 2022. They revealed they have more than 5 trillion searches per year, approximately 14 billion per day, almost double the 8.5 billion/day estimated in 2022.

ChatGPT has an estimated 1 billion queries/day, so one explanation could be that it’s still small in comparison-just 7% of Google’s traffic. But that still doesn’t explain how Google could be growing 8%+ YoY while a new entrant has 1B+ queries/day.

How can both Google and ChatGPT be growing?

I think the answer is clear: ChatGPT is dramatically increasing the size of the search market. GenAI-powered search has opened up entirely new use cases people previously never would have thought to “Google,” yet are perfectly suited for ChatGPT. No one previously went to Google and typed out their whole life story to get therapy or asked for feedback on important texts and emails before sending them.

On top of that, ChatGPT has transformed searches from single queries into back-and-forth conversations, exponentially increasing the number of queries. The multi-turn, context-aware experience is perfect for complex travel planning, writing, research, or even DIY projects at home. ChatGPT has unlocked entirely new search behaviors, ones that wouldn’t have even been thought of as “search” before.

Thus, the reason Google’s search volume continues to grow is that the overall search market itself is expanding rapidly. It’s a growing pie, allowing everyone to show impressive numbers. Google can also rely on its massive ecosystem (Maps, Android, Gmail, Chrome, etc.) to continue driving search traffic while quickly experimenting with adding AI features.

But, in the “new” segment of the search market—generative AI queries—ChatGPT is the tool of choice. This new set of queries is the battleground where ChatGPT is winning, well on its way to over 1 billion monthly users.

What happens from here?

I don’t think both products will continue to grow indefinitely. Consumer search is a winner-take-all market with strong network effects—users prefer a single place for all their searches, and that demand drives both paid and organic content to optimize for the primary search engine. I think OpenAI has realized the opportunity in front of them and has Google squarely in its sights. They already have consumer search and API products that compete with Google’s two primary revenue streams (Search Ads and Google Cloud).

As they continue to grow, I expect OpenAI to launch a browser, productivity/email services, and eventually even hardware as they take on the behemoth that is Google. The ultimate moonshot product for each company is a universal AI assistant, personalized to each user. Delivering this product well requires controlling the end-to-end experience of hardware, operating systems, communication tools, and search.

In the next five years, I anticipate a tipping point. OpenAI’s growth trajectory and strategic expansions will finally dent Google’s seemingly invincible advertising revenue, which will start to slowly—and then rapidly—decline.

For the past decade-plus, cloud software products have had roughly this architecture:

• A Static back end, very structured set of entities with defined relationships between each entity

• A series of views with a variety of ways to see the back end data, with a specific limited set of configuration options (filter, sort, search, etc.) for each user

• A logic layer that interpolates between the back end and the various views

Basically, they’ve all been variations of the “model, view, controller” framework. (engineers reading, I know this is oversimplified but stick with me)

This has allowed for a very limited amount of customization at the user level - at best a user gets a specific set of saved views with filters and sorting options. But it’s the same set of options available across the entire user base.

Due to this architecture, software has been very limited in the set of problems it can solve. A single, shared user interface has to be built for the average end user to understand and accomplish their daily tasks. This means it has to be designed for a specific set of use cases that are common across the entire customer base, inherently limiting the level of customization possible.

The promise of “low code” or “no code” apps has been around for a while, and while there are some great products out there, none are building truly personalized experiences or have been able to displace the large incumbents. The core of their challenge is they are still tied to the MVC era architecture.

That’s all going to change with the next generation of Generative AI-enabled products.

Interface Principle: Complexity is dictated by Frequency and Precision

To understand the personalized software era, it’s helpful to frame it with a principle I’ve developed around interface design - the complexity of an experience should be based on the user’s frequency of usage and level of precision required.

For example, an airline pilot gets trained for thousands of hours on how to fly a plane. Airplane control needs to be highly precise, so it works best to have a wide array of different buttons and very specific knobs to control the plane. They would be completely ineffective if there were just a single chat window to control the plane. Other examples include music production, professional photography cameras, and surgical equipment.

On the other hand, an employee of a company that once a month needs to find out how to submit an expense report wants the exact opposite. That user wants a simple, familiar interface that has no learning curve and gets them their answer quickly. It’s okay if they are given several responses and can figure out which one is correct. This is where a universal interface like a search or chat works well.

In the MVC era, this resulted in a landscape that looked something like this:

Low-precision tasks were solved primarily via search - a single universal interface that enables searching the company wiki, searching on google, etc. For consumers, examples of low-precision tasks are aggregators such as social media feeds and news feeds.

High-precision tasks with frequent usage were areas of huge success for SaaS products, resulting in the endlessly configurable horizontal SaaS behemoths like Salesforce, ServiceNow, Atlassian, etc. But even these tools ran into their limits, with their power users being siloed into a specific function (Sales for Salesforce, IT for ServiceNow, Engineering for Atlassian).

High-precision tasks with infrequent usage have been a dead zone for SaaS products and have primarily been solved by custom internal tools.

The Personalized Software Era

In the Generative AI world, we are already seeing a major shift for the left side (low precision) of this diagram from Search → Chat. As an aside - the mistake I see a lot of companies making is trying to just slap a chat-focused interface for tasks on the right side (high precision). A user doesn’t want a universal interface for a task that is very specific to their role, there’s still value in interfaces designed for specific tasks.

Generative AI enables a fundamentally different architecture - one that will unlock a whole different way to solve high precision use cases. Here’s how it will look:

• Static back end - a structured set of entities with defined relationships between each entity (for now similar to MVC, but could become dynamic with things like vector stores and composable graphs)

• An interactive view where the user can dynamically describe and iterate on how they want to see the data. Looks something like using natural language to interact with an agent that can create and update a custom interface for that product. Behind the scenes the agent is writing and updating code that is then hosted by the product.

• A constraints layer that specifies the edges/boundaries of what is possible to show the user. The product dictates what elements are available, what types of logic queries are possible, design systems, and more.

These products will have default views for when a user logs in, and will then allow for the creation of a fully customized interface. Customization can be done by each individual user, allowing for user-level personalization. At larger enterprise organizations, there would also be admin-level ability to create default customizations or constraints by role or department.

User-personalized products will be highly disruptive for all high precision categories. It allows for an even deeper level of precision for a single user’s task, or a company’s unique workflows. Imagine:

• A customer support agent building the perfect view with their own hotkeys, 3rd party data integrations, and error log to be able to diagnose and fix customer issues in their area of expertise.

• An engineering manager designing the exact view of their team’s sprints, blockers, roadmap, and deployments, while each engineer on their team can see pull requests, commit messages, deployment issues, and their tasks in the way they prefer.

• A VP of sales creating the perfect view of their quarterly attainment, deals that need attention, and reps that might need help.

All through a simple natural language conversation with the product - no custom development required.

Many categories that have been relegated to internal tooling such as analytics dashboards and customer success tooling can be productized and scaled. The products that have been tried and failed many times such as personal CRMs could be scaled into larger products.

Pete Koomen outlines a similar vision for what AI-native software will look like in AI Horseless-carriages, his concept of users writing system prompts is basically them building the interactive view layer, and his concept of tools is the constraints layer.

I’ve only found a few products that allow for truly custom experiences per user. Lindy, ActivePieces, and other agent builders on on right right track. 0.email look promising, and Retool has pivoted nicely into the AI era.

New architecture, new problems

This new architecture will create a new set of problems, each of which are probably interesting product opportunities in themselves:

Security and privacy - How do you ensure users can’t do SQL injections or malicious attacks in a world where they can literally write code that is hosted on your product?

Forward compatibility - As you iterate on your product, especially making changes to the back end data model, how do avoid breaking every user’s custom built experience?

Analytics and A/B testing - How do you measure the usage of your product if every user has a customized experience? What does an A/B test look like? Cluster based experimentation is the frontier here.

Design systems - How do you enforce a consistent design system to ensure familiarity across your product?

Support - How do you debug customer’s issues when they’ve built a fully custom experience?

Possible Futures

I am personally extremely excited about the next era. The existing SaaS products we have are for the most part not loved by users, charge extremely high prices, and require enormous amounts of administration. There’s a few possible scenarios that could play out:

1. Incumbents adapt, letting users build custom interfaces in Salesforce, ServiceNow, etc. This seems least likely to me, it’s a classic innovator’s dilemma and these players are likely to be disrupted.

2. New startups enter the market and you get next generation the next generation of CRM, ITSM tools that are AI-native and deeply personalized by company and user. This seems most likely to me for enterprise software.

3. The next generation of AI-personalized products are flexible and can be personalized enough that they become fully horizontal. A single product can solve accounting, project management, ticketing, CRM, and more. In this world, the next generation of startups are primarily building infrastructure and constraint tooling. This is likely to happen in certain categories and will be most disruptive.

It’s still early days and the opportunities are huge. If you’re building a user-personalized software product, I’d love to hear about it! I’ll be writing more about AI era software and product strategy, subscribe to get the latest.

Subscribe now

The last major technology shift that fundamentally changed how products were developed was the cloud. This lead to the widespread adoption of Agile and Lean Startup methodologies. These methods were built upon the assumption (which was true at the time) that writing code was the bottleneck for software development. Software engineers were expensive, and no company wanted to waste their resources building the wrong thing.

As a result, significant emphasis was placed on validating, defining, and refining product requirements before writing any code. Organizations grew entire functions dedicated to researching, designing, and validating concepts prior to development. This was a huge improvement from the traditional waterfall style development process and led to the success of tens of thousands of new products.

We’re in the middle of the next technology shift

Generative AI is revolutionizing how code is written. In just the past 6 months, coding assistant tools like Cursor, Windsurf, Lovable, Bolt, and Replit have evolved from being cute ways to help with 10-20% of code to now generating the majority of code for many startups. 1 in 4 companies in the latest YC batch have 95% of their code written by AI. This shift is due to advancements in the capabilities of core models, most notably Claude Sonnet 3.5 and 3.7, explicitly optimized for coding. These models will only continue to improve as we scale training, optimize them more for code, and enhance their "chain of thought" capabilities during runtime.

It’s hard to understate how impactful this will be for the entire world. It’s already unlocked vibe coding, caused major tech companies to stop hiring engineers, and allowed dozens of 1 person teams to get millions in revenue.

A major implication of this shift is that writing code is no longer the bottleneck for building a product. In fact, it’s headed towards becoming one of the easiest parts of product development.

What does this mean? The knock off effects are hard to understate for teams that build software products. All of the assumptions about how to build products need to be revisited.

Let’s say a product team is building a CRM, and they want to add a new analytics feature. Traditionally, teams spent weeks or even months conducting customer research, designing mockups, and gathering user feedback to refine the feature before engineers started writing any code.

In the new world, that process is flipped on its head. Especially if you’re a small startup with a simple codebase, its faster to simply build a basic version of the feature, ship it, and get immediate user feedback. Even if you are an enterprise with a giant codebase, you can quickly ship a clickable prototype of a feature faster than it would take you to design it.

You can skip those weeks or months of research and mockups completely while validating your features faster.

The new product development lifecycle

This new way to build products is much faster and simpler than before, it involves just 4 steps.

1. Prioritize features by impact

2. Ship simple version or clickable prototype

3. Test at scale with users, measure impact

4. Iterate or kill

This change may not seem that dramatic, but it results in an entirely different product development team.

The new product development team

That’s a reduction from roughly 6 different roles (PM, PMM, Design, Engineering, Research, Data) to just 3 (PMM, Engineering, Data). Less roles means less coordination and a more effective team.

This new team will also prioritize a very different set of tools. QA, A/B testing, and analytics tools will become more valuable than ever. User research and static design tools will become irrelevant.

The new bottleneck

So if the primary bottleneck isn’t writing code anymore, then what is it?

User attention.

There will be proliferation of products in every category, so the only thing that matters will be attention. The companies that learn to build an amazing marketing engine around every new feature and product they build will win. These new product development teams will be smaller, faster, more data-driven, and will help build a set of products that users love.

I’m personally extremely excited to see a proliferation of new, cheaper products. I’ll be writing more about the implications of the 100x decrease in writing code, would love to hear what you think.

Subscribe now

Leave a comment

Over the past 4 years, I was on the investor side of the table for 60+ companies when I was a platform manager at Upslope, I’ve been directly involved as an investor in 20+ companies via Kokopelli, and I got to watch many of the 10 companies in my Techstars Boulder 2018 cohort raise rounds.

I took my learnings from these 100+ Seed rounds I’ve been directly involved with to build the process outlined below.

Please note: This post is not advice about how much to raise, whether to do a priced round or note, or how to structure your fundraise. There are plenty of great articles about that from smarter people than me. This post is also not as applicable for founder outside of the US, I unfortunately only have experience in the US.

This is a straight-forward, step-by-step guide on the process of actually building an investor pipeline, working through it, and getting commitments on the round.

1) Build your list of investors

The first step is to build a list of relevant investors to your company. As of this writing, there are over 800 Seed funds.

There are a few great tools to help you build your list:

• Dorm Room Funds’ VCWiz: Each investor is tagged by location they invest in and sectors. Great resource and can also act as your CRM.

• NFX’s Signal: a tool built by NFX that has a large portion of investors, includes check size for many, and it will tell you most likely connection paths based on your Gmail conversations.

• Shai Goldman’s Google Doc: Shai has diligently kept this list up to date over the past few years, and it’s a great resource

• If you’re in Colorado, I made a list of Colorado Early Stage funds.

You should prioritize funds based on location, stage/check size, sector, and regions they invest in. Try to build a list of 50–100 funds that you think are the best match for you.

2) Make a google doc for intros

Next, you should put the list of your highest priority funds you need intros to into a Google Doc. Use Google Docs, not another tool, because you are going to use this doc to get offers for introductions to funds from your network. They will be most comfortable in Google Docs, and you want to make it as easy as possible for them.

The folks in your network also probably have short attention spans, so make sure they focus on the top 10 funds you want intros to. Organize the google doc into an “above the fold” section of your top 10 funds, and below that is all the other funds for your true champion mentors who want to go through them all.

Also, use this doc to add details about relevant portfolio companies, any portfolio companies that are customers of yours, and any other relevant notes specific to this firm. You will use these later on when you are getting the intros.

I created a basic Google Docs template for you to start from ( choose file-> make a copy) https://docs.google.com/spreadsheets/d/1HMbwzRpsTJ7OZb0OHHzBcyER3uaExOxhcYeJHQsTdgI/edit#gid=0

3) Build an intro pipeline

You may have noticed the last 3 columns in the template are for intro offers. The idea here is that you can get up to 3 folks (or more) in your network to offer intros to each fund. After that, you can decide which one to try to get the intro from.

Note: Do not blast this doc out in a mass email. Send individual emails or a mail merge to your mentors and other connections with a link the google doc and asking them to add their name next to anyone they are willing to make intros to.

4) Make a deck and get double-opt-in

While you were building this intro pipeline, you should also have been working on your slide deck for your round. Creating a solid pitch deck is another topic I won’t cover in this post, but it’s definitely worth putting real effort and designer resources into your deck. It’s the most critical fundraising material you’ll have.

Now, go through the intro offer list and choose the person who you think would give the best intro. Send them an email titled “Intro: YOURNAME @ YOUR COMPANY <> FUND NAME.” Include a blurb about your company along with something slightly personalized for that fund, such as specific companies that they’ve invested in or portfolio companies of theirs that are customers of yours (this is where those other columns in the doc come in handy).

• Attached a compressed PDF of your pitch deck so they can open it up directly in Gmail. Use Docsend or some other tracking service if you really want to (I won’t get into that debate).

• The goal here is to make it insanely easy for the person to forward. The less work you make it for them, the faster they will do it. If someone offers you multiple intros, send a separate email for each intro so they can just forward them all and not have to write emails.

As you get introductions and set meetings, update the status of funds so that you aren’t getting more intro offers for funds you’ve already connected with.

You’ll continue to get connected to folks during the process that can make intros for you. Keep going back to step 3 with those people and sending them the doc to make intros.

If you don’t hear back on an intro ask that you send out, follow up with the person who offered the intro once. If you don’t hear back again, try with the next person who offered an intro. Hitting a fund multiple times is fine if you haven’t heard back with a formal pass. Persistence is 🔑.

5) Go to SF, NYC, and maybe LA

While you should definitely start with funds based close to you, you should be looking across the US for sources of capital. If you (like me) are not based in SF, NYC, or LA, I think it is very important that you physically visit at least one if not all of those cities as a part of your fundraising process.

As you are planning trips to these cities, you can create separate tabs in the google doc for each city. I made one for SF and one for NYC and then followed the process of step #3 for my strongest connections in each of those cities.

Make sure you have specific dates booked to be in those cities, hopefully centered around a meeting you have scheduled. This helps create urgency and a reason for VCs to make time while you’re in town. It can often be good to do a 30 min introductory call a week or two before you visit to make sure the in-person visit is a good use of everyone’s time.

6) Pitch, pitch, pitch

You should be able to get intros to at least 20–30 funds. Remember that every interaction you have with them is part of your pitch. Be thoughtful about your emails, phone calls, and scheduling process. Take note of how they treat the process too, you’re finding a partner to help you build the business for the next decade.

Note: Practice your pitch before you have to pitch it live. Use your deck as a guide but be willing and able to go off script. Figure out the most common questions and have answers for them.

Try to enjoy it! You get to tell the story of your business to a bunch of super smart people and have them give you feedback/advice. Don’t take it personally, use it to make your business better.

Make sure you leave time at the end of meetings to ask them specific questions about fund/check size, how they like to be involved, what their ownership target is (so important!), and what their decision-making process involved. Directly ask them how interested they are and what the next steps are.

If you’re doing a priced round, if you get a few firms very interested then pick a specific week where you will accept term sheets. Be clear with them that is when you will be accepting term sheets and drive towards that date.

7) Close the loop

Finally, after you meet with investors on your list, whether its a positive or negative outcome make sure to follow up with the person who made the introduction for you.

You can go back to your handy google doc here, find the person who made the intro and just shoot them a quick note. This goes a long way. These are very valuable connections for many people, and they will want to know what happened so they can better curate intros going forward. Be very respectful of the social capital in play.

If you’re looking for more tips, Jenny Fielding wrote an amazing guide on this topic: https://www.slideshare.net/jefielding/building-an-investor-pipeline-spreadsheet-keep-your-funding-flowing/jefielding/building-an-investor-pipeline-spreadsheet-keep-your-funding-flowing

That’s it! Best of luck, it’s a crazy process, but I truly believe that as of this writing there has never been a better time to raise a Seed round as a founder. Would love to hear any tweaks or tips others have used in their process.