In response, Devin shares a plan for implementation:

First, Devin got stuck because he didn’t know how to authenticate. I told him to create his own workspace through our sign up flow and that guidance worked.

Two hours later I got this:

Holy crap! The PR went up, someone on the team approved it, and the fix was merged. This was admittedly a simple bug to fix, but Devin still did it perfectly. And without Devin, we would not have prioritized the bug anytime soon.

But here’s where it gets interesting. You know that clarifying question Devin asked above about how to authenticate? I told Devin to create his own account, and Devin automatically saved this as under his “Knowledge repository”. Devin learns on his own — I don’t anticipate having to provide that same guidance again.

This is a glimpse of how engineering will evolve. Right now, Devin needs my permission. He waits for instructions, then goes to work. But as these systems improve, we’ll trust them more. Eventually, they won’t just wait around for requests; they’ll be proactive. A new bug ticket shows up in Linear? Devin will start implementing the fix before I say anything. A new way to improve performance that Devin discovered in another codebase? Devin will just implement that too. Codebases will be self-optimizing.

Already, Devin can handle different kinds of input. I showed him a screenshot on Slack and asked for a copy change in the UI. It found the relevant code and fixed it. It’s not hard to imagine the next steps: upload a video of a bug (that was sent by a customer) instead of describing it in text, and Devin fixes it. Or maybe Devin records a short demo video of each change so we can see it in action without manually testing. These agents will be multi-modal on input and output.

Just this week I discovered Devin, made videos with Sora, and upgraded to ChatGPT Pro. Each experience was profound. I’m so excited about what’s next.

But can you picture Steve Jobs laying out Apple’s entire product roadmap for everyone to see? Or Elon Musk at Tesla doing the same? You might argue, “Well, those are consumer brands.” But look at top B2B companies like Stripe and Shopify: they also don’t broadcast what they plan to build next. Indeed they almost do the opposite: they wait until the time is right, and then they drop everything at once in seasonal releases.

If public roadmaps were such a great idea, wouldn’t we see at least one of these leaders doing it?

When you publish a roadmap, you give away your story. Instead of revealing new features when they’re ready and explaining why they matter as part of a bigger narrative, you end up with a static list of upcoming stuff. It’s like showing the script of a movie before it’s filmed. You lose the tension, the drama, the room to improvise. A public roadmap boxes you in.

The narrative matters, but so does the ability to surprise and delight. Look at OpenAI’s 12 Days of OpenAI campaign. It’s Day 4 now, and each day they promise something big. People are excited and full of anticipation. Nobody knows exactly what’s coming, but the predictions, debates, and gossip — all of that is free, powerful marketing.

Public roadmaps commit you to plans too soon. Good product development isn’t linear. You try something, learn something else, and pivot. But if you’ve promised certain features by certain dates, you can’t change direction without looking foolish. Suddenly you’re at war with your own words.

A public roadmap can push you toward the obvious requests. After all, you asked users what they wanted, and they told you. But the best products often come from what customers didn’t even know they needed. OpenAI didn’t ask users if they wanted a chatbot. They just made it.

The pace slows down too. Internally, you can set ambitious deadlines, fail to meet them, and still land in a good place. But once you’ve told the world a feature will launch on a specific date, you’re forced to pick safe timelines. You give yourself more margin for error and hedge, and that means shipping slower.

By publishing your roadmap, you hand competitors a guide to your future plans. Why should they do their own thinking if you’ll do it for them?

So don’t give in to the pressure. Resist the idea of a public roadmap. The best companies don’t need to promise what’s next. They’d rather show you when the time comes. You should too.

AI Sparkles

AI Sparkles are tiny bits of AI-driven polish — small enhancements that don’t change the product’s core but make some part of it easier, faster, or more delightful. They’re subtle — maybe a smart autocomplete here, a better suggestion there. No single AI Sparkle will blow you away, but they add up, and in aggregate they can make a real difference to the user experience.

Three examples:

1. In Linear, if I have Slack connected, I can highlight a message and pick “New issue from Linear.” The issue title and body get filled out automatically.

2. In Arc, tab names and downloads are automatically renamed and tidied up.

   

3. In Zapier, instead of manually picking triggers and actions, I can just describe my workflow and Zapier figures it out.

Embedded Assistants

If AI Sparkles are like little touches of delight, Embedded Assistants are more like having a helper standing right next to you. Think of them as ChatGPT baked directly into a product, but with two differences: they can access your data, and they can actually do things in the product.

Two examples:

1. In Notion, I can ask questions and get answers from everything in my workspace — no need to search around.

   

2. In Shopify, there’s Sidekick. I can say “discount all my products 10%” or “How many orders did I get this week?” and it just does it or tells me.

   

AI Employees

Now, what’s next? Imagine that instead of just having an AI helper on the side, you could hire AI Employees inside your favorite product. You’d go to the “Teams” page in Notion, Linear, or Figma, click “Invite member,” and choose from a lineup of specialized agents. Each costs a monthly fee, each with its own specialty.

The distinguishing characteristic of AI Employees is that they are proactive. They also get the same access and permissions as a human employee, so they can start contributing right away.

For example, in Notion imagine being able to invite a “Product Manager” Agent for $75/mo. You can give it access to a page or the whole workspace. It’ll add comments with clarifying questions about workflows, turn complicated paragraphs into tables and diagrams, and do competitor research by browsing the Internet if it notices that the section is empty. If I have a specific ask, I can just tag it with a question or a request to do something for me.

In Figma, maybe you add a “Copywriter” Agent for $25/month. This one would specialize in writing better product and marketing copy, and ensuring it works with the designs already made. It might rewrite some text directly, leave comments with suggestions, or bring in examples from other products.

I don’t think we’ll see a singular super-intelligent AI that’s great for everyone and everything. More likely, some products will have a lineup of AI Employees, and whoever’s in charge will pick which ones to bring on board based on priorities, taste, budget, and other criteria.

Predictions

1. AI Sparkles will become standard and lose the “AI” label. They’ll just be part of what good software does.

2. Embedded Assistants will be embraced by companies with products that are complex — platforms, compound products, etc. The more complexity there is the more valuable an assistant is that can collapse complexity.

3. Pricing will change. Right now, you can add AI features for $5-10/month per user. That works for Sparkles and Assistants, but AI Employees will cost a lot more.2 When it becomes clear that AI Employees can replace expensive human labor, it’ll be easier to justify premium prices.

I've always wondered about this. Large companies seem to be quite good at reliably turning money into effective solutions when it comes to marketing, sales, recruiting, and many other problems. But somehow building great products is an exception — money and resources don’t seem to help that much. What makes it so difficult? And why are large companies particularly bad at it?

Products have objectively little surface area of significance. A product will only ever have one home page and one onboarding experience, for example. As a company grows, the ratio of important surface area per employee goes to 0. When you have something important that an increasing number of people care about, pressure accumulates. This battle to dictate how important real estate is used happens within product and design, but as a company grows demands increasingly come from other functions like sales and marketing. And those requests often aren't in the interests of users at all.1 Users don't really care about your startup's ability to grow, for instance.

The dynamic creates an environment where the typical path for company leaders is to meet in the middle, to compromise, and to delegate. These are also the traits that lead to a deteriorating product. The best product CEOs and product leaders — Steve Jobs, Elon Musk, Brian Chesky — are the opposite. They’re in the details, opinionated, and top-down exercise judgement. Great products are never designed by committee. 2

Product decisions, especially early ones, matter so much because some subset have high path dependence. If you run a disappointing marketing campaign with an agency, you can fire the agency and work with another. But some product decisions are like early branches of a tree. The latest iPhone is remarkably similar to the original one introduced seventeen years ago. The core Google search experience has barely changed in over two decades. It's often only apparent after the fact what decisions have high path dependence, and so you need to have a high-caliber thoughtful product culture from the start and not lose it, or you might not be able to course correct later.

Features have an innate tendency to get flushed out over time. Even if a feature isn't a hit, the excuse to "Make it a bit better and it might work" comes so naturally and is rarely questioned. The accumulated bloat that results and the reality that mediocre features are rarely deprecated is how great products lose the original simplicity that made them compelling in the first place. Everything you add to a product has a complexity cost and given the human incentives at play, subtraction is difficult. Products evolve to become easier to use only in cultures that encourage ruthless deprecation and re-imagining things that aren't working. In contrast to Apple, which quickly deprecated the Touch Bar on MacBooks and Force Touch on iPhones after going all-in, Google somehow believes users want to have all of Google Messages, Google Chat, Google Meet, Google Voice, Google Duo, and Google Allo.

As organizations mature more decisions tend to be data-driven. While analytics is a helpful tool, analytics should never be the thing that's most important. The problem with obsessive data-driven cultures is the inevitable end state of landing in a local maximum that blinds you from bigger opportunities. Any feature that's been A/B tested to death 100s of times will naturally outperform a reimagined approach because the latter hasn't undergone any product evolution. This is why Meta, being so obsessed with optimization on their big blue Facebook app, missed the novel UI paradigms that Snap and Instagram pioneered.

A data-obsessive approach is also how you end up with something that lacks any semblance of cohesion and character. The big blue Facebook app, shaped through A/B tests, performed better and better on metrics until suddenly it didn't. And at that point it became too rigid to substantially change things.

Getting the details right is what separates many of the best products. It's why the startup employee stack looks shockingly similar with Stripe, Notion and Linear almost always making an appearance. Are there no alternatives to these products? Of course there are — there are 100s of note-taking and issue tracking apps — but Notion and Linear won the hearts of users, not by having more features, but by being more enjoyable to use and by nailing more of the details.

In practice, it's difficult to sustain a product culture that is obsessed with details because arguments for new functionality always sound more compelling. Whereas a new feature means new stuff for marketing and sales to pitch, what's the argument for making something a tiny bit faster or a bit more fun to use? The right interpretation is that those improvements compound in ways that will eventually help a company set itself apart on a dimension that is indeed *more difficult* to replicate that a feature. For example, imagine how difficult it would be for Jira to become as fast as Linear, or for Adyen to become more developer friendly than Stripe.

So why aren't there more exceptional products? Product decisions are particularly difficult to delegate, business functions have emerging anti-user incentives, path dependence is high, analytics can be a trap, bloat is difficult to avoid, and prioritizing product details is unpopular. The default trajectory for a growing business is product decline. Startups have to actively fight against the many forces that will over time make a product mediocre.

In startups, a lot of the work you do as a founder or early employee is in those last two buckets. When people ask me for career advice I often tell them that a great reason to join something early-stage is precisely because you get to do work that isn’t solution-certain. Let me explain with some examples.

Solution-certain work is implementing well-understood solutions. If you join a big tech company like Google or Facebook, most projects have this structure. More senior people will have already done the work of outlining the solution in detail and your goal is to implement it. While you'll have the opportunity to refine your craft, you are robbed of the opportunity to explore the full solution space. For example, you might be tasked to design or implement a new page in an onboarding flow using existing components.

Solution-uncertain work happens when the problem is clear and you have to figure out the best solution. This work is more difficult, more creative, and requires more persistence because there are more unknowns. You probably won't get work like this at big tech companies until you've been there for many years, but you might get it at some late-stage startups. An example of a project like this is to improve customer onboarding. The solution can take many shapes — make it faster, make it simpler, provide more guidance, etc. You have to figure out which path to take to best solve a given problem.

Problem-uncertain work happens when the problem itself is uncertain and your objective is to first define it rigorously and ensure you're asking the right questions. Since perfectly executing a project that solves a non-existent problem is worse than making progress towards solving an important problem, this is the most important and highest-leverage work. Continuing with the onboarding example, perhaps the state of the onboarding experience is indeed poor, but poor onboarding is not the best articulation of the root problem. Instead, maybe the root problem should be reframed as bad retention, and you should therefore consider spending your time on other initiatives like acquiring different users in the first place, improving your positioning, or changing your pricing.

If you want to start a startup, get comfortable with work that requires you to think about problems rigorously and be open to solutions that take many shapes. Indeed, one of the first decisions you'll make when you start a company — what idea to work on — is the perfect example of a situation where understanding and defining the problem really matters. The reason most startup teams fail is not because their solution is sub-par but because they're solving an insufficiently important and misunderstood problem.

Most interview questions are more indicative of interview prep than skill, they don’t let outstanding candidates shine, and they take too long to answer. So what questions should you ask if you care about fast and accurate assessment?

1. What's a complex project you've worked on that you're particularly proud of?

This is my favorite question to assess if someone is exceptional at their craft. After the initial answer, I like to go deep and ask lots of follow-up questions. The deeper I can go, the stronger the candidate.

Don't be afraid to really dig deep and ask Why an uncomfortable number of times. Exceptional candidates will understand the chosen topic so well that they’ll enjoy navigating the details with you. Also, try to understand the candidate's contribution vs. that of the team. If everything is framed as "we did x, y, z", and the candidate won't speak about their own substantive contributions that's a negative sign.

2. What's your perspective on work life balance?

I’m looking for people that have an optimistic relationship to their craft. People with genuine passion for a field have the most important advantage of all because all the time they spend honing their skills outside of work compounds in ways that it doesn’t for someone that has an adversarial relationship to work.

By asking the question in a neutral way you can usually tell if someone thinks about work and life as a strict tradeoff or if work and life are more complementary and mutually reinforcing. Also, passion is the most important pre-requisite for hard work, and this question tends to elicit it in a way that asking a question like “What’s the hardest you’ve ever worked?” does not.

3. What kind of culture do you want to be part of? Besides the things everyone wants like smart and kind coworkers, what uniquely matters to you?

You have to reject some super talented people if they are bad for your culture. I’m often asked about culture at the end of interviews and I’ve noticed it’s useful to try and ask the candidate first so that you get a less biased answer.

There are many flavors of great culture. Some prioritize speed at all costs. Others prioritize craft and nailing the details. Some are intense and top-down. Others are more creative and collaborative. There is no best culture. But there are cultural attributes that bring out the best and unlock the most potential in a person. If there is a cultural trait at your startup that makes you unique and that you want to cultivate, you should ask about it directly.

4. If you were to join us, how would you want the first days and weeks to go?

People that get things done tend to think about solutions in sooner and shorter time intervals. Today instead of tomorrow. Days instead of weeks. Weeks instead of months.

You can get a sense for someone’s views about speed by asking them how they think about onboarding. Do they want to jump on a sales call or put up a PR on the first day? Great sign. Do they expect to spend weeks in formal training? Not the right fit for a startup where learning happens through osmosis, failure, and getting your hands dirty.

The worst hire I’ve ever made wasn’t a good fit because all initiatives, no matters how mission-critical, would never be kicked off before next quarter. The person believed that as a startup we were constrained by the same bureaucratic process that stifles large companies. But the most amazing thing about startups is that you can have a brilliant idea in the morning and start executing it immediately.

5. Who are some people that you admire or look up to in your field?

Great people are intensely curious about the field they operate it and almost all genuinely curious people make an effort to find people they look up to — heroes, experts, mentors, etc. While easy to lie about curiosity when asked directly, I like this question because it surfaces curiosity in a way that lets really great candidates stand out.

6. What questions do you have for me?

Always leave sufficient time for candidates to ask questions. So many things you do at a startup come down to making sure you are asking the right questions, so being able to see if a candidate can figure out what the most important questions are for them to assess the role/startup will tell you a lot about how they tick.

This is also the part of the interview where a candidate can best showcase their curiosity, creativity, and generativeness. If an interview was okay until now but then they ask about a brilliant tactic I hadn’t considered, that’s the best way to win me back over.

1. Scientific progress

   1. How much will continued human-led scientific progress in non-AI fields matter in the long run?

   2. How should ambitious people decide between pursuing deep domain expertise vs trying to push the AI frontier forward? For e.g., when all cancers are cured, will we look back and be more indebted to individuals who pursued AI research or oncology research post 2022?

   3. Do we need new technical breakthroughs to get to AGI?

   4. Will the step change from GPT4 to GPT5 will be as large as the one from GPT3 to GPT4? When does progress start to asymptote?

2. Training data

   1. The first LLMs required large amounts of training data to work. Will we reach a point soon where more human-generated training data won’t be needed?

   2. What role does synthetic data play in the long run?

   3. Lack of training data is the main constraint holding back faster progress in robotics. How does this get solved?

3. Regulation & piracy

   1. There are many things LLMs don’t let you do like generate videos with copyrighted characters or porn. Are we about to enter an LLM era of piracy that will be like the torrenting era of the mid 2000s - early 2010s?

   2. If we don’t enter a piracy era, how will these use cases be supported given the obvious demand?

   3. If we do enter a piracy era, how long will it last and will it be followed by a legitimization period much like what happened with Netflix/iTunes? Will incumbents adapt or get disrupted?

4. Business

   1. Which company will have the assistant with the most users in the long run? ChatGPT is the early leader but it doesn’t have access to most of my data — email, calendar, Slack, Notion, etc. And while ChatGPT has a plugin ecosystem, will big tech companies let another company own the relationship to end users?

   2. Will each tech giant have their own assistant and plugin ecosystem, and then leverage their large user-bases to force smaller B2B companies to build plugins? How much room is there for “all-in-one assistants & plugin ecosystems”? Does Glean have a shot at being very successful for enterprise use cases vs the tech giants?

   3. Many smaller B2B startups like Notion have or are building their own assistants. Will this trend continue or is it short-lived because narrow assistants all suffer from insufficient context (for e.g. Notion’s assistant not knowing about Slack conversations)?

   4. Google is paying Reddit $60M/yr for training data. Will UGC platform like Reddit make an increasing amount of revenue from data sales? If so, what changes? Will Reddit focus more on creating product experiences that result in higher quality data vs. more traffic for advertisers?

5. Startups

   1. We got the iPhone in 2007 and the most successful mobile-enabled startups in 2009 (Uber, WhatsApp), 2010 (Instagram), 2011 (Snap), 2013 (Robinhood, Doordash), and 2016 (TikTok). So the average very successful mobile-enabled startup started ~4.5 yrs after the iPhone release. ChatGPT launched late 2022, so we’re only ~1.25yrs into the new AI paradigm. Are we so early into the AI era that perhaps none of the long-term AI winners have been founded yet? Or is there something structurally different about this wave that makes being early more important?

   2. Will value from AI mostly be captured by large incumbents or by startups?

   3. Many AI pioneers are Canadian (Geoffrey Hinton, Ilya Sutskever, etc.). Will we see any massively successful Canadian AI startups?

   4. Will human and synthetic data live side-by-side in UGC platforms? Or will we see products like “YouTube for generative content” as generative data makes up an increasing fraction data on the internet?

   5. What happens to SEO? What tricks will marketers come up with to get included in assistant responses? Who’s building the Ahrefs / Semrush for the generative data world?

   6. Will any organization build a “Common Crawl for robotics” or will these dataset be guarded by private companies?

6. Competition

   1. How will the relationship between Microsoft and OpenAI evolve? Microsoft owns a huge fraction of OpenAI, but the stipulation that OpenAI does not share post AGI technology with Microsoft must motivate Microsoft to build independent expertise.

   2. How much room is there for new foundation model companies? How screwed are Pika and Runway in a world where OpenAI keeps investing billions in Sora? Should new startups be building foundation models?

   3. To what extent will profit be competed away for foundational model companies? What gives these companies pricing power in the long run? Will it be more important that models are smart or that they have access to unique and time-sensitive data sets like X, Reddit, and the NYT?

   4. What’s the future of open-source models? So far it looks like the most capable models won’t be open-source (too much compute $ needed) and the most knowledgeable models won’t be open-source (no access to expensive proprietary datasets). But maybe those limitations are fine for the majority of B2B and B2C use cases?

   5. What happens with Perplexity? They have a much better product than Google’s and Microsoft’s assistants, but is that enough to be massively successful?

7. Google

   1. What happens to google.com in the long run? Does an assistant fully replace search, will there be toggles to switch between both paradigms, will the interface depend on the search input, or something else?

   2. How much of an advantage is having access to YouTube’s video corpus?

   3. How does Google keep increasing revenue overall and ad revenue specifically in a world where assistants are monetizing directly without ads?

   4. Will the bet in TPU architectures pay off or will it be a crutch in the long run?

   5. How will Google handle the backlash of choosing particular LLM responses in their models (vs being able to deflect blame to an algorithm that merely ranks external links)?

   6. Does Sundar get fired because he failed to get Google to an early lead despite having every conceivable advantage?

8. Arts & creativity

   1. We’re not far away from a hyper-personalizable future where anyone for e.g. can generate their own ending to Game of Thrones. But in a world where this capability exists, will people be excited by it? So much of what makes media & entertainment great is the shared global context — i.e knowing that after a season finale drops you can talk about it with anyone IRL, in subreddits, etc.

   2. AI both democratizes creation (anyone with a phone will be have the power to make Hollywood-quality movies) and destroys the economics of it for some fraction of the population (some copywriters, illustrators, etc.). On the whole, will the impact on creative jobs be more positive or negative?

9. Robotics

   1. Will robotics have an inflection point “ChatGPT moment” or will progress in robotics be more gradual given training data limitations?

   2. How large is Tesla’s advantage given their experience with self-driving cars?

   3. How large is Figure’s advantages given their relationship to OpenAI?

10. Consumer hardware

    1. Is the intersection of LLMs and AR/VR actually interesting or will future generations of Apple Vision Pro / Quest be successful primarily due to other reasons?

    2. Does continued LLM progress create conditions for a consumer device to emerge that’s more popular than the smartphone?

11. Philosophy

    1. How will our understanding of consciousness evolve as LLMs continue to get better?

    2. Will consciousness of some sort emerge in more advanced LLMs?

    3. Are we in the early stages of a crisis wherein we realize that we’re not as uniquely smart and self-aware as we thought?

    4. LLMs have already passed the Turing Test without much fanfare. Will the same be happen when we pass AGI thresholds?

The status quo

While it is straightforward to set up a Shopify store and sell a physical product on the Internet, entrepreneurs that sell and deliver a service on the Internet have to stitch together business apps to support payments, file-sharing, eSignatures, intake forms, and other such functions. Since these apps don’t work seamlessly with one another, the client experience is fragmented and messy. Clients are bombarded with email notifications from different senders, can’t access a unified dashboard, and can’t even update their own information. If you've worked with an accountant before, these issues might sound familiar.

The solution

Businesses that understand this problem know that having a portal — a product where clients can log in and access everything in one place — means higher customer satisfaction, improved retention, new growth vectors, and more efficiency. Those with the resources hire engineers to design and build a product that solves this issue. But creating custom software is no easy task — you need to build authentication, permissions, notifications, security infrastructure, and that’s just the foundation. All of this is expensive and takes a long time.

At Portal, we're creating building blocks (messaging, payments, file-sharing, etc.) that businesses can configure and use to design their own portal. And crucially, since all our infrastructure is white-label, companies can completely customize their portal and have direct relationships with clients.

The future

Service companies of the future will look more like product companies today. Clients will be able to sign up online, go through customized onboarding flows, complete todos, send messages, schedule meetings, make payments, and more. Anything that can be automated will be automated, freeing time to focus on work that matters.

Not only can we help more companies get started but we can also empower SMBs and startups to compete against large incumbents. By democratizing access to easy-to-use composable infrastructure, we can level the playing field.

If this sounds interesting to you, we are hiring for most roles. Please consider applying or email me if you have any questions.

Triggered by Facebook's rebrand to Meta, last month much of the tech world started talking and opining about the "metaverse". This tweet by Shaan Puri struck a nerve with many:

The metaverse is the moment in time where our digital life is worth more to us than our physical life. This is not an overnight change. Or an invention by some steve jobs type. It's a gradual change that's been happening for 20 yrs

There hasn't been a year in the last decade where my digital life to physical life ratio has decreased. The trend of our digital lives becoming relatively more important is both unquestionable and unstoppable. And it's both exciting and terrifying to think about what AR/VR will be like after 50 years of compounding progress. If we aren't already more robot than human, it seems like a good bet that by 2070 we will.

This digital shift has come about so suddenly that I worry about long term effects on our brains. To protect myself, moving forward I'm going to make conscious effort to disconnect completely from the digital. Inspired by this excellent post by Erik Torenberg, this upcoming year I'm committed to doing at least one device-free, silent, solitude day or weekend each quarter.

Large companies have many advantages over startups – better distribution, bigger marketing budgets, optimized sales processes, etc. But large companies can never match the customer support of a startup.

Startups should aim to create something so good that people tell their friends about it. Everyone knows that you can accomplish that goal with an excellent product. But I've found that providing really great customer support is surprisingly effective. More reviews on our G2 page mention customer support than any one feature.

A single A+ customer support experience can turn a customer into an advocate. And since advocates drive word-of-mouth growth, extraordinary customer support should be a priority for every startup.

I've handled a few hundred customer support tickets now and these are some learnings:

1. Always introduce yourself as a founder. Users always appreciate speaking with a founder directly and will be more friendly and forgiving.

2. Commit to doing the work, if it makes sense. If you get a feature request that's aligned with the roadmap and quick to get done, communicate you'll do it  and offer a clear ETA.

3. Give public credit. If you release a feature that was requested by someone, follow up personally and give credit publicly (on your changelog, Slack community, etc.)

4. Be ultra responsive.  It's much better to respond to a support ticket in 1 minute vs 15 minutes. It's much more impressive to respond on a Sat night than a weekday. Respond so quickly that people wonder how you managed to type so quickly.

This mistake can happen in many ways, but the most common is to hire a sales person because you can't sell yourself.  When this happens, the issue is almost always that you haven't made something people really want.

About 9 months ago, I nearly made this mistake a third time. We attended office hours with our YC group partner Jared Friedman, and Jared said something that has stuck with me since:

"The early work of finding PMF is not parallelizable."

He's right. To increase the chance of finding PMF, you need see the problem you're tackling in its entirety and be able to connect all the dots. Here's what happens when you hire too early:

• More abstraction layers. When you put people in between yourself and the user, information that reaches you gets diluted. Specific anecdotes get lost in summary notes. Emotions get lost in bullet points. Nuggets of information that initially seem irrelevant get skipped altogether. All of this is bad – you want to hear raw insights, be able to follow up on the spot, and build direct rapport with users so that you can follow up personally. It's crucial to hear the precise language people use to describe their problems because you will want to use that same language when you speak to your users.

• Wasted time. In the early stages, you should be building product and talking to customers. Both (1) the process of hiring people (writing job descriptions, reaching out to candidates, scheduling, interviewing, offer negotiations) and (2) managing people (training, 1/1 meetings, etc.) are distractions that you will likely underestimate.

• More excuses. As soon as someone else is responsible for a job, when things go wrong, it's no longer clear whether the person is bad at their job or your product is bad.¹ Because founders are so attached to their creations, your inclination will be to believe the former even if there is evidence to the contrary.

In addition to the above points which directly hinder progress towards PMF, there are other reasons to initially grow your team slowly:

• You will hire better people if you're further along. A startup with PMF and a growth chart that goes up and to the right is much more compelling to candidates.

• You maintain flexibility. The larger the team, the more momentum you have. It becomes much harder to pivot or make adjustments when the team is large. Most people don't handle uncertainty well.

¹ When you do hire people, this is why for some roles the standard advice is to hire in pairs. If you bring on two account executives, for example, you're more likely to understand how sellable your product is.

Image synthesis

Tools like Figma have sidebars filled with property controls (sizing, positioning, alignment, fonts, colors, etc.). To complete any complex task, you have to perform a long sequence of steps. But with natural language input and powerful image synthesis, that need not be the case. Imagine being able to open a Figma file and entering commands like this:

Recreate the stripe.com/payments page and replace all logos with the selected image.

For the selected frame that shows a web view, create an equivalent mobile page. Show me 5 variants.

Generate an illustration that matches the selected text. Ensure it's consistent with illustrations A and B.

These capabilities wouldn't just provide 10% time savings. They can turn a 100 hour project into a 10 minute one. And of course a shift in capabilities of this magnitude opens the door for startups that reimagine how creative tools should work.

Video synthesis

Powerful image synthesis will precede powerful video synthesis because videos are more complex. But eventually similar capabilities will exist for video. Writers will be able to write stories and see them come to life in video. Initially the technique will only work with simple stories (maybe children books will be the first application!) but eventually we will have movie-quality video generated by text instruction. Imagine the possibilities:

Creating characters
You'll be able to describe a character in text – how they look, how they speak, their accent, etc – and see that character come to life in video. You will be able to upload a picture of your own face and make that the main character of your movie.

Generating speech
You'll be able to input text dialogue and watch your character say those words – with mouth movements that perfectly map the intonations of words.

High-abstraction controls
Imagine a video editor where instead of adjusting the brightness, you can adjust the personality of characters or the weather in the background.

Infinite paths
As the cost of video synthesis approaches 0, it will be much easier to create choose-your-own-adventure games and videos.

The future

We shape our tools, and thereafter our tools shape us.
- John Culkin

GPT3 and similar models are powerful enabling technology. So far, most products built on top are doing text synthesis – for example GitHub Copilot and Copy.ai. There are going to be many more soon and some will attempt to build the next creative suite.

These new creative tools will lead to new a branch of art. If you are interested in this space, I highly recommend this blog post to see examples of what is already possible. I'll leave you with this quote from the blog post that summarizes what creativity might look like in this new world.

And despite the fact that the model does most of the work in actually generating the image, I still feel creative – I feel like an artist – when working with these models. There’s a real element of creativity to figuring out what to prompt the model for. The natural language input is a total open sandbox, and if you can wield words to the model’s liking, you can create almost anything.
- Charlie Snell

And yet I frequently come across large companies with websites that are terrible. A great marketing website alone is not a reason to invest in a company, but to me a bad one is a reason to pass. It's a canary in the coal mine – a guarantee that a company doesn't understand how to prioritize things and, in all likelihood, means that other parts of the company are also incompetently run.

The honeymoon test. Few ideas are exciting after the first 2 weeks. In that initial period you're in a honeymoon phase – the positive parts of an idea consume you and you leave no space for doubt. As you dedicate more time to research, you talk to potential customers, research competitors, and figure out what execution looks like. Most of the time, excitement fades away. The takeaway: give yourself at least 2 weeks before fully committing to an idea and going into build mode.

The IPO test. Imagine a world where your idea achieves its full potential. Would you be proud of the impact you've had on the world? I made the mistake once of pivoting to an idea that I didn't really love.¹ If you're going to embark on a journey for 10+ years, you should pick something that you will be proud of. The startup journey has many highs and lows, and working on something you're proud of is the best way to keep your head up during the lows.

¹ The company was called Lightout. We installed digital screens on Uber cars to help drivers make money and help advertisers run hyperlocal geofenced campaigns with a self-serve advertising platform. I loved the first part of the idea but hated the second. At our peak, we had about 250 cars in the Bay Area, and it really was exhilarating to see our product in the wild. But deep down, a part of me always felt that the idea taken to its extreme – cities filled with digital billboards – was not one I would be proud of.

I've been a believer in Tesla since 2014 and have been asked hundreds of times to defend my bullish position. My justification has always been this:

1. Talent disproportionally matters in the car industry.

2. Tesla can hire a significant fraction of the most talented people in the functions that matter.

3. Given enough time, Tesla will create superior products - better factories, better robots, better batteries, better powertrains, better AI capabilities, etc.

Let's break down the first 2 points.

Talent matters in the industry. Talent matters more if you're working on difficult problems. Tesla, of course, is the exemplary example - battery design, chip design, ML, and the dozens of other technical problems they are tackling are hard.

There are industries where talent isn't sufficient – for example, it is near impossible to compete with an established social network because of existing network effects. But in the case of the auto industry, I didn't think in 2014 that the head start that competitors had were insurmountable. Counterintuitively, acquiring capital, building factories, and figuring out supply chains are easier problems than overcoming social network effects.

Tesla can hire a significant fraction of the most talented people in the industry. In 2015, Mark Fields (the CEO of Ford at the time) came to my college to give a talk about Ford and the car industry. I remember thinking that the guy was completely full of shit and lacked a compelling vision about the future of Ford. And my second thought was this:

If you're a person that a car company really wants to hire - an AI PhD grad at Stanford, a Staff engineer at Google, etc - why would you possibly join a company like Ford over Tesla?

My guess is that a tiny minority of the top 0.1% critical talent pool would choose Ford over Tesla (and of course the same is true for GM, Chrysler, etc.).

Predictions

I predict Anduril and Boom Supersonic will similarly absorb the majority of top talent and go on to do extraordinarily well. For Anduril, the best people interested in defense will choose Anduril over incumbents like Northrop Grumman and Raytheon. And for Boom, the best people interested in building planes will choose Boom over incumbents like Airbus and Boeing.

After this happened, I realized that this was an instance of something I've now seen many times. People have a natural desire to allocate time on subjects they are most familiar with. While for most jobs this is fine (if you're a lawyer you should mainly be doing legal work!), founders have to be careful. Since the only objective for early stage startups is to find PMF, almost all of your time should go towards building and selling, regardless of your background or interests.

The founders that have to be most careful are the ones that have no experience building or selling – most MBA grads, bankers, consultants, etc. For example, ex-consultants are likely to waste time creating the perfect pitch deck and ex-bankers are likely to waste time on revenue forecasts.

Technical founders and ex-sales founders have an advantage because they will naturally focus on half of what matters. But since you need both halves to succeed, don't forget to check your blindspots. In more cases than not, technical founders will underestimate sales and sales founders underestimate what a great product will do.

¹ There are exceptions in highly regulated industries. But even in those cases, you are better off hiring someone or working with a consultant.