It’s a narrow question, but still an interesting one to think through. There were three elements at play with mobile:

1. Mobile enabled a dramatic expansion in internet minutes that were up for grabs. Rather than the zero-sum game of fighting for the limited time people spent in front of a desktop computer, suddenly there were all these new minutes up for grabs -- when someone was commuting, waiting in line, on the toilet, etc. This is why the consumer mobile wave was biased towards a particular set of players: those that offered more dopamine per minute by connecting to the internet than staying in the analog world.

2. Mobile introduced new ingredients with which a consumer founder could build (multiple cameras, GPS chip, accelerometer), creating greenfield opportunities for startups.

3. The AppStore dramatically reduced the friction of installing a new application, making it easier for people to download apps to their heart's content.

Now let's do a thought experiment:

If we didn't have (2), and we had these powerful computers in our pockets *without* any cameras/GPS/etc, would mobile have still been a huge catalyst for consumer startups? Definitely. True, we wouldn't have had Uber, Instagram, Snap, or TikTok, but apps like FB, WhatsApp, Twitter, LinkedIn, Pinterest, Reddit, Airbnb, YouTube, etc. would still be the dominant players they are today.

What if we didn't have (3)? I think consumer internet would have been fine -- the motivation was so strong. And the friction would have gotten ironed out over time.

But for (1), what if we held the # of minutes a person could spend online constant pre-mobile vs post? Instead of the almost seven hours we now spend on our devices, you’d have companies fighting for ~2hrs of daily internet time. This would definitely have tempered the magnitude of the mobile consumer internet wave.

Implications on Consumer AI

If we take this thought experiment back to AI consumer apps, there are broadly two categories of AI consumer apps: There are going to be those that will orient towards consumer attention, and those that will use AI as more of an enabling technology for a utility (think Google Maps or Uber in the mobile era). The mobile era saw an explosion of startups going after both categories. How about the consumer AI era?

Let’s start with the first category: AI apps that orient towards consumer attention. Here, the question comes down to how much AI is able to expand the # of minutes we spend connected digitally, and then what % of existing digital minutes AI can cannibalize.

For example:

• You can engage with an AI friend or romantic partner instead of a real life friends / romantic partners.

• You can talk to your AI / AI friend / AI therapist / AI cooking coach / etc while you are cleaning the house, prepping dinner, walking your dog, commuting, instead of listening to music / podcast / radio, watching TV, etc. (Related post: A framework for consumer attention. Voice AI is a “liquid”.)

• The education use cases are obvious for younger people — AI will be a constant companion through school work.

• You might have more leisure time in a world where AI means you need to work less. etc.

• (As an aside: Yes of course, there are COUNTLESS use cases where you will use AI during the day to accomplish work, but I’m focused here on the strictly consumer use cases.)

In a way, the success of the consumer internet (including gaming) has created the conditions for one of the biggest AI consumer opportunities: loneliness.

Still, this isn’t like the mobile era where the dopamine maximizing drive of consumers made changing behavior easy. An AI friend has to compete for time against the dopamine optimized torrent of TikTok et al (which will all be taking advantage of LLMs in their own way to further optimize engagement), whereas when Instagram was ascendent, it was mostly competing with the unoptimized analog world. Earning a minute today is a way harder battle to win than it was in the early days of mobile.

Don’t get me wrong: AI will win a lot of consumer minutes (maybe too many, if you think about the importance of human-to-human connection), but it won’t be the same magnitude as the minute land-grab of the mobile era… until we have new AI-native devices or neural implants that will once again dramatically expand the amount of time we connect digitally. So by this measure, the consumer AI wave should be smaller than the mobile wave.

But consumer minutes is not the only battleground. There will be countless consumer AI companies that emerge that are more akin to Uber or Google Maps: companies that take advantage of AI to make consumer’s lives easier or give consumers new superpowers (think companies like Manus* or Lovable). In other words, consumer AI companies that are more about (2) than (1). I think this will be more of the battleground for consumer AI startups, at least in terms of quantity of companies.

It won’t be easy, and as I’ll write about next, we’re early in the wave. But get ready. We’re in for a lot of change.

If you are building on the frontier of consumer AI, I'd love to hear from you. My email remains Sarah at benchmark dot com.

*A Benchmark portfolio company.

As I mentioned in my last post, An AI Metamorphosis, I love Borislav’s perspective on AI because I feel like he has fully immersed himself in it, he thinks deeply about the craft of coding, and he has a similar philosophical perspective on things. This is the second half of our discussion, which contains three sections:

1. Why Rekki’s CTO, Borislav, still codes two days a week without any help from an LLM.

2. The theory of mind, and reviewing code generated by an LLM.

3. What the future holds as more and more code is written by LLMs.

Coding two days a week without the help from an LLM.

One of the things that I hadn't heard before is how Borislav insists on coding two days per week in a blank terminal, with no AI help.

"I have days where I don't code with any AI.... I open a blank terminal and then I just code. I don't do any syntax highlighting or autocomplete or anything, just code... Because you start to forget... If you just use language models all the time, so much code gets written. You start losing touch with what is a lot of code, because [ideally] you want zero code. The more you use an AI, the more you forget. It gives you a bunch of code — you would never have written code like that, but.. the code is in front of you, in just a few seconds, and it works. You accept it and feel kind of ok because you don’t consider it ‘my’ code, ‘I’ didn't write this code. You have no empathy towards it. You just read it and accept it like some alien code."

There are two things embedded here: first, is this "slow boiling frog" element that Borislav elaborates on below, where you lose touch with what great code looks like. Second is how an engineer can get abstracted from the prior pride of authorship of the code they create. They didn’t create the code and don’t feel the same level of ownership of it, so now, it's less about the code itself, and more "does it get the job done"? But as we all know, the test "does it get the job done" can be short term gain, long term pain.

Back to Borislav:

"On one of the AI-free days, as you are writing a bunch of boilerplate yourself, you start hating it. Why, why am I writing this? This code is just nonsense. So your nonsense filter stays sharp. The next day, when the LLM spits out 500 lines of code, I tell it 'that's too complex, I don't want it’, I can still feel the spidey sense. Otherwise, it's slow boiling the frog. When you're pushed little by little, your tolerance changes so much. And you forget what is good or not."

This comment made me think of how so many engineering teams now report on what percentage of their code is written by AI, but this can be a bit of a vanity metric (and long time readers know how much I hate vanity metrics!). I get the desire to show how much they are embracing AI by reporting on this, but if you optimize for this metric, you won’t be motivated to edit the extreme amounts of code the AI creates, creating challenges in the future. Something to watch out for. Back to Borislav:

"The key is you have to grow your capacity to make the system do what you want it to do, for it to be a way to express yourself. How easy is it for you to tell the computers what to do? For example in 1980 you just poke an address and a pixel appears, the distance between the code, the wires and the display is in the order of a few hundred lines of machine and micro code, now, there are easily 10 million lines of code. It is easier to add new abstractions and new indirections than to fix anything, and in the last 15-20 years we have been doing that, and this is what language models learned from. Why do we even add layers of indirection and abstractions? Because we are humans, and this is our way to manage complexity. The language model is a stochastic ghost of a human, so it also abstracts. When there are 10 billion lines, I won’t even try to comprehend what is going on, I will have to settle for a summary of the code. It's just going to be too much if you don't try to take control of it... You want to conserve complexity, not let accidental complexity bubble up. But it's going to be really hard to distinguish accidental complexity from actual complexity when you are 100k lines of alien code deep."

The theory of mind, and reviewing code generated by an LLM.

One of the challenges with editing LLM-generated code Borislav talks about is the "theory of mind" -- basically, how humans have an intuition for how other humans think through a problem, but that that intuition doesn't map to how a LLM thinks.

Borislav gives the classic example of being in the room with Sally. Sally leaves her keys on the table and then leaves the room. You take those keys and put them in a box on the table. When Sally comes back to the room, she knows to look in the box because she can pretend to be you in her mind, and know that that is probably where you would hide the keys.

Back to code, "when I read somebody else's code, there are all kinds of reasons for this code's existence, for example historical. At Booking.com, there was a column in a table that was called country code, and inside of it, it had currency. You read code that seems like nonsense, but you know there is a human reason for it. You ask yourself ‘why did they do it like that?’ But when you have this stochastic machine that produces tokens, this is just not true. You don't have a theory of mind for the symbols. Imagine it produces a comment ‘this is a clean version of the function’, you read it and read the function and ask ‘clean for who?’. You can't reason about it in the same way. You can't think about it in the same way. You can’t understand the reason for their existence. It did it because vectors aligned like that, that’s it."

And so you end up "disconnecting" from the code. Instead of refactoring it to a way that makes sense in your brain, many engineers instead act more like a product owner of the code -- "you don't want to read the code, you just want it to work."

"That's a completely valid option, I think, and many people do it and that's fine. And they just have to recognize what exactly we are doing. But the other option is to treat it as your code. And you have to ask yourself, okay, is this what I would have written? And that's what your critique should be, what your bar should be. You must still have to have a strong sense of responsibility and ownership."

I put that part in bold because it really resonated. I wonder if it will stand the test of time?

FWIW, I asked Borislav what percentage of his code is written by an LLM, he wasn't sure but guessed 70%.

What the future holds as more and more code is written by LLMs.

"Imagine your life depending on something and it's like 20 million tokens of code. This thing drives your car or whatever it does. The fact that we tested a black box doesn't really help us, we have to really think carefully about what we are doing. What is this doing? How are we building on top of these layers, because the layers are also going to increase.

“If a language model can just write the program for all kinds of architectures, maybe we don't even have a compatibility layer anymore. And now humans cannot even read or write the programs that the language models can. And that's like a very disturbing thought because imagine, well, maybe the next instruction set is actually designed by just AI. A compiler takes our code and compiles it for this thing, which we don't even know how it works. It's just ten times faster than our thing. Now we can't even step through it, it's just a complete black box. So we move the black box deeper and deeper into the silicon, and then we completely lose touch with how this is working."

One of the risks Borislav highlights here, is that basically our ability to build programs with true complexity is going to get harder and harder, because just accepting what code an LLM produces will introduce "fake" complexity and force us to completely "dissociate" from the code itself, making it harder to build real complexity over time.

Part of the journey Borislav has gone on is learning when to use an LLM.

"You must think of the pressure of the tokens. Imagine you ask it to do sentiment analysis, and you expect a single Positive or Negative from it. Now, that's too much pressure, you must allow it to get to the answer. For example, make a plan of how it would do the analysis and maybe instead of asking it to do positive or negative, ask it to give you a score for each, so if it samples ‘negative’ it can still recover. I don’t think people realize how much is lost when a token collapses. Think of how much is lost when a word is written. Imagine you write the word ‘positive’ on a piece of paper, and then 1 year later, you read the piece of paper and you have to add one more word, you think for a second and write ‘attitude’, then 2 years later you see ‘positive attitude’ what do you write next? All the high dimensional nuance is lost after the collapse.

Another good way to think of it is imagine a dark room with a document and a note, the note says ‘do sentiment analysis on the document, answer only positive/negative’, so you pick a random person from the street, drop them in the room, you see the result, you don’t like it, you change the instructions on the note, and then you pick another random person. Each document will get a random person, your instructions better be good.

I really advise anybody just to watch Karpathy's lectures. It's like 10 hours. I watched them like four times. Watch them, pause them, and it will help you to kind of have a mental model. Because it's a machine that we have never experienced. And you try to think of it either as a human or a computer, and it's neither."

Thank you Borislav for your insights!

For too many company leaders, I’ve noticed they think this means having your engineers “tab” their way through their coding tasks with a tool like Cursor, for their employees to be able to ask ChatGPT questions of their internal data, and for them to adopt AI vertical tools.

But these things, while important, are all on the surface. To become AI-native, it takes, what one company I’m lucky to work with describes as a full “metamorphosis.”

What follows is an interview I did with Borislav Nikolov, the CTO of Rekki. The transcript for our conversation was 50+ pages, so I did my best to edit for conciseness and clarity. Still, I consider Borislav as AI-native as you can get so in order not to cut some of the nuance out, I am going to publish this in two posts.

As background, Rekki has a marketplace for restaurants and their wholesale suppliers. They deal with thousands of restaurants making thousands of orders every day. This creates a stream of nitty-gritty engineering tasks to glue things together like payments, risk management, and complicated business logic. Every time the operations team had a new need or a tweak to the system, they’d create a task for the engineering team. Until one day Borislav realized: what if I can empower everyone on the team to “solve their own problems?”

I’m not sure what the company of the future looks like once AI fully permeates, but I think we all have something to learn from how Borislav approached this question.

In this post, we get into:

1. The “ah ha” moment of realizing that every single person in the company would soon be able to code, and how this would let him push tasks back to the domain experts in the company and have everyone “solve their own problems”.

2. How this changed his job as CTO from leading a team that executes on tasks to one that creates the internal infrastructure almost like a PaaS provider, where the whole company is full of developers he “can’t trust”.

3. His realization that if an employee can’t do something themselves, it is his fault for not giving them the right primitives.

4. How he pushed the behavior change internally. This started by first, having everyone in the company deeply understand large language models, and then challenging the conditioning everyone had of assuming they couldn’t solve their own problems because they couldn’t code.

In the second post, we’ll explore:

1. Why he still codes two days a week without any help from an LLM.

2. The theory of mind, and reviewing code generated by an LLM.

3. What the future holds as more and more code is written by LLMs.

The “Ah Ha” moment.

“As they say, necessity is the mother of invention.”

Shortly after GPT 3.5 came out, Borislav, was staring down a massive backlog of engineering requests from the company’s operations team. The backlog was daunting, both in its count, and also because each task itself, while potentially trivial from an engineering perspective, required real time to ramp up on the problem before any code could even be written (let alone tested). Then it hit him: literally everyone in Rekki could now write SQL.

“So I was like, look, why can't I just enable the people that are actually asking for the tasks to do the task? If a restaurant orders three times very close together, that's probably a fraud thing. So instead of me making a system that has to check the data, has to have some state, etc., they can just do that themselves… This was the first real moment where I realized that everybody would soon be able to code and solve the tasks in whatever domain they are in.”

This was something that wasn’t possible with other no-code tools he’d tried, because inevitably the no-code platforms were not expressive enough.

“Previously what was always missing and why the low code and no code solution would fall flat was because at some point, it's just not expressive enough. Somebody has to write the code. There is this law of irreducible complexity.” But he realized that if he could give his team the right infrastructure for their code to run somewhere safely, with things like privacy considerations, security, and notifications if there is a problem, he could empower everyone in the company to code and solve their own problems.

Rather than divide the company into developers and non-developers, if everyone could write SQL, Borislav realized that the jump to building automations on a service like Zapier or Make.com was finally possible.

“Now, if we could just make this SQL into API primitive, someone can create the rest of the workflow at Zapier or Make.com, and of course the testing they can do herself by placing orders with a testing restaurant.”

What this meant for Borislav’s role internally, and the infrastructure he set up.

So, his thought was "how can I make it possible for everyone to do their job, without me defining what they can and cannot do? What primitives are needed? How can I make it reliable and as safe as possible?”

By primitives he meant things like load/store/search, turn a SQL query into an API, or SQL query into a trigger that pushes to a webhook when there is new data, monitoring, alerting, etc.

If an employee gets stuck not being able to do something they want, Borislav considers it his fault for not giving them the right primitives.

“I almost think of myself as an internal PaaS provider, the whole company is full of developers that I cannot trust, kind of how AWS engineers think about REKKI's engineers, they have to run our code, but they don't trust any of their users, but at the same time they have to empower us giving us the right primitives to express ourselves and do what we want to do with as little friction as possible… In the future I hope I can empower everyone to use Cursor and just deploy code, but we have a long way to go to make it safe and secure.”

Depending on the company, the primitives will be different. But to Borislav, “the important part is to think of the whole company as people who can truly build anything they put their mind to, and the only limit is you, your attitude, and your infrastructure.”

“You will be surprised how much you can do with a workflow with SQL to API, ChatGPT classifier and few API calls.”

Now, rather than a company divided into developers and non-developers, the company has developers, semi-developers, and a small (~10%) non-developer make-up. Yes, the developers are still needed to unblock or enable the semi-developers to do their jobs, but that work is amortized because the developers are giving the semi-developers the “fishing rod, not the fish”.

How to make this organizational behavior change

I asked Borislav what advice he’d give other companies to make this transition from pre-AI to AI-native.

“If you put a company in front of me and I can ask them one question, it would just be ‘do you understand this technology? Can you use this machine?’” In Borislav’s experience, most people are quick to point out the AI’s limitations – hallucinations, producing bad code. But they haven’t taken the time to truly understand what the machine is and how it works.

Borislav himself studied Andrej Karpathy’s videos and others on LLMs for close to 50 hours until he could back-propagate by hand on pen and paper, and deeply understood “what a plus means and what a multiplication means and what the collapse of the vector space into a token means.” As he said, “it’s a machine we’ve never experienced.” To use it, first understand it. Once he could understand “the way the transformer programs itself to emit the token” and therefore its limitations, he understood how to take full advantage of a LLM to code, and transformed Rekki’s culture to be AI-first.

“You have to understand that by talking to it you're programming it. That's what you're doing.”

At Rekki, Borislav had the entire team watch Karaphy’s Zero to Hero series, as well as a bunch of videos from other teachers like Justin Johnson, Fei-Fei Li, and 3Blue1Brown. They did Learning Fridays where they’d watch lectures together and discuss the content.

Then, he started pushing all the requests back to the teams that had the domain expertise. His default changed from doing the tasks for people, to teaching them how to do it themselves, and making sure that they had the infrastructure that empowered them.

At first, this was hard.

“You have to believe in the capability of your non-tech people. They are conditioned to think they cannot code, to think that only engineers can not solve their technical problems. So the very first step is for you to believe in them. Then the second step is having the infrastructure to support them, and the third is to teach them to solve their problems themselves.”

For example, someone on the operations team needed to get a bunch of data. Borislav could have written a scraper for him that gets the data, parses it, and analyzes it. Instead, he told the employee to download Cursor, and just “type the request in.” “Now, he has 20 scrapers that he just uses nonstop.”

“So again, first believe that non-coders can code, then figure out where the code lives, because that is a massive problem, e.g. how do they deploy and monitor it, make.com, zapier, bolt, their own laptops etc.? And then the real issues with security, PII and GDPR, anonymization, etc. How many outages are you willing to tolerate because ChatGPT wrote a stupid infinitely recursive query?”

The Bigger Picture: AI-Native is a Mindset Shift, Not a Feature List

What Borislav demonstrates so clearly is that becoming AI-native isn't about layering AI features onto existing processes. It requires a fundamental restructuring of how we think about work, capability, and organizational design. This is the "metamorphosis" he references.

Three critical insights emerge from Rekki's transformation:

1. The democratization of technical capability is real and imminent. When everyone can code, the traditional bottlenecks between technical and non-technical teams dissolve. Rekki transformed from an engineering backlog to hundreds of automation workflows built by the people who understand the problems best.

2. Leadership roles fundamentally change. Technical leaders shift from task executors to platform builders. They create secure, reliable primitives that enable domain experts to safely solve their own problems. The parallels to cloud computing's transformation of infrastructure are striking: CTOs become an internal PaaS provider to the whole company rather than gatekeepers.

3. Organizational psychology is the biggest barrier. Both engineers who are used to being the exclusive "problem solvers" and non-technical staff who've been conditioned to believe they "can't code" must undergo significant mental shifts. The limitation isn't the technology - it's our deeply ingrained beliefs about who can do what.

The question for founders isn't whether this transformation happens, but how quickly your organizations adapt. The companies that drive this change aren't just adding AI features - they're fundamentally rethinking what it means to be an organization in the AI era.

More soon.

Subscribe now

After eight years at Benchmark, I’m shifting to a Venture Partner role. In this new role, I'll continue to make new investments on behalf of Benchmark, serve on my existing boards, and contribute to Benchmark and our portfolio. What's changing is my focus and time allocation.

When I joined Benchmark, I was drawn to its distinctive model in venture capital—a small equal partnership of no more than six general partners, all generalists, deeply committed to partnering with the most ambitious founders of our generation. With no junior people or platform team, the structure demands total dedication from each partner. It's a model that has proven remarkably successful, and also one that demands a certain approach to venture.

AI isn’t just another technology wave – it’s a transformational force that will reshape every corner of our lives. As I've watched the space evolve, I've increasingly realized that for me to best support founders building in AI, I need more space. More space to go AI-native by fully immersing myself in AI tools at the edge, more space to reflect on the last ten years and try to understand what the next ten might bring, and maybe even (don't hold me to this!) more space to think through it all by writing.

After reflecting on this with my partners, it became clear that transitioning to a Venture Partner role would give me the flexibility I need while maintaining my connection to Benchmark's core mission. I remain fully committed to Benchmark and our craft of deep partnership with the founders we back, just now with more room for curiosity and exploration (and yes… vibe coding).

For the extraordinary founders I currently work with: my commitment remains unchanged. I’m grateful for our partnership.

For founders building at the edge of AI: I'll now have more bandwidth to engage deeply with your vision. Whether you're just beginning your founder journey or scaling a team that's already found product-market fit, my email remains sarah at benchmark com.

Sarah

There are so many tidbits he shared on his mindset which he describes as “AI-Forward”. I am going to include a bunch of quotes on his thinking (slightly edited for brevity). And then he shared a few fun examples on how he uses AI day-to-day.

James’s Mindset and Approach

"I just assume anything I'm doing, like almost every decision I'm making is better with AI. At this point, it's been built into almost every single thing. If I'm trying to name my team. I'm trying to write my LinkedIn post. I'm trying to write a strategy. I'm trying to think of anything I do. I go to Claude, my favorite, or my default tab is Perplexity."

"I'd say if you look at my podcast time, it's probably gone from three to five hours a week to probably one hour a week because for the other three to four hours advanced voice mode has taken over.... I think that my biggest realization is again, AI knows more about everything than we know about anything. And so therefore, why wouldn't I be curious? And so therefore if you have a choice of talking to the most knowledgeable person who has ever lived, why wouldn't I do that on just about everything?"

"I think that you and I grew up with this idea that 'Effort equals output'. I'm like, oh cool, you must have put so much time into this Sarah. Oh, what an amazing dossier you wrote. What an amazing briefing. Now we're moving to this world where the connection between effort and output is completely disappearing." To this point, James mentioned the example Ethan Mollick gives in his book Co-Intelligence on writing recommendation letters for his students. It used to be a signal when a Professor would take the time to write a recommendation letter for a student, because that effort would create scarcity. The professor isn't going to say yes to everyone who asks. What happens to the value of a recommendation letter when AI can do it in three minutes?

"I keep on reminding myself, like Ethan Mollick says, AI is the worst it will ever be today."

Use Case Examples

The three big themes for James are about 1) brainstorming and writing partner, 2) synthesizing information for him, and 3) scouring the web and other sources for information that is relevant to him (vs him having to go out and get it himself).

Three examples he gives:

Brainstorming and writing partner with Claude:

"Anything, LinkedIn posts, LinkedIn long form articles, and internal stuff, like documents internally, phrasing's a lot. I find there's a lot of like, what should I call this team? What should I call this? Is there a really pithy phrase? The key there is that you ask for 25 ideas. And then say, I only like two and four. Give me 25 more. And you say, I only like one, two, and three. You just keep going.”

"Each average idea for any human is probably better than each average idea for an AI. But the AI can give you 200 of them in 10 seconds and the human can give you like 13 in like two hours. So Claude can give you 200 taglines, 150 are absolute garbage. But 50 are pretty good and 10 are magic."

He often prompts Claude after it's written something, "Let's remove the jargon and make it more like me."

Bringing information to you.

James has started experimenting with OpenAI’s Operator. "It's not that each individual task is amazing. My mind was blown because I'm watching TV with with my kids, and in parallel, I'm preparing a dossier on you, and I'm finding some candidates for a search, and I'm researching our dishwasher. The average person doesn't have an EA or a Chief of Staff to give work to, but now they do."

To show how simple it was for James to get value out of Operator, here is the prompt he used:

"I'm preparing to meet Sarah Tavel from Benchmark on Monday. Can you please create a full Google document dossier style that tells me as much as you can about her. Please start with broader information and then narrow down to anything that she's written about AI on LinkedIn, X, her blog, or other public sources in the past 12 months."

You can watch Operator in action here. And the output here (with some formatting from James). Definitely not perfect, but as James reminds us, “worst it will ever be.”

Daily email newsletter synthesizer, built with Relay

One of the core use cases James uses AI for is synthesizing content. A great example of this is for newsletters. Like all of us, he is subscribed to a lot of email newsletters (make this one more!), and often doesn’t read everything or know what to focus on. So he built a tool using Relay that ingests the newsletters in his Gmail, and then synthesizes the content so he gets an easy-to-glance summary every morning, and he can then dive in on the posts he’s curious about.

This use case is a classic example to me of something I’m noticing the “AI Savants” I’ve spoken to all have in common — they are willing to do the extra upfront effort to get the downstream benefits.

James kindly put together 10 slides to show how he does it here: Getting a Daily Digest of the So Whats from a Series of Newsletters using relay.app.

Here’s an example of what the output is. Check it out!!

Well… thanks Google. The more I find myself taking advantage of AI, the more I realize I’ve been unlearning this trained self-sufficiency. What we all have now is an expert available to us to ask any question or even delegate many tasks. Yes, sometimes we need to give it some back-up documents, or clearer instructions, and sometimes it has no idea what it’s talking about (not that it will tell you). But if you can move from a mindset of self-sufficiency by default, to asking the LLM expert at your fingertips by default, a world starts to open.

This isn’t about moving queries off of Google into ChatGPT. It’s also a shift in the kind of problems we can tackle, and how we’d approach them:

• Google Era: I’m stuck on a bug. Let me search and read a bunch of StackOverflow threads to debug this.

• AI Era: Pull request to an AI developer who has access to your entire codebase.

• Google Era: I wonder what diet changes or supplements I should be taking given my latest blood test results. [Does a lot of searching.]

• AI Era: Uploads latest blood test result into a ChatGPT Project, “You are a medical analysis expert specializing in interpreting blood work results and creating personalized health recommendations….” [full prompt instructions here].

• Google Era: I wonder how good our team is at communicating together. [Bookmarks a lot of articles on team communication to read later.]

• AI Era: "Tell us something we don't know, or wouldn't want to know about ourselves" [uploads file of team Slack conversations]. (-idea and instructions from Tom Lawrence)

• Google Era: My wife just did an embryo transfer. Let me search for articles on what she should be eating to maximize the changes of a healthy pregnancy.

• AI Era: "You are the world’s foremost expert on maternal and pediatric nutrition like Dr. Lily Nichols. You’ve studied the nutrition studies and lore from every culture in the world that best promote maternal and baby health and brain development. Help me put together a week by week menu — breakfasts, lunches, dinner, snacks — for my wife from embryo transfer to end of fourth trimester.

  My wife is a pescatarian but will do bone broths, butter, dairy, etc. She just doesn’t like eating meat, though maybe that may change with pregnancy. We have a worldly palate.” (-instructions from Aike Ho)

Are the experts perfect for every question? No, of course not. (I like how Ethan Mollick describes the “jagged frontier.”) Is this the worse they will ever be? Yes.

This isn’t just about an AI doing the work we once did of finding/synthesizing information. One thing that hits me is how much more personalized an AI result can be than a Google search will ever be. The paradigm is just so different. When I do a Google search, we all have some expectation of objectivity in the results. We critique Google when two people do the same search (e.g., on a news event) and get different results. And of course the means by which Google personalizes results has been more implicit - inferring from my searches, geo, etc. So personalization gets pushed to the edges to sites that are also stuck in a Google / search paradigm. Take TripAdvisor as an example. I have to figure out the best hotel for me by looking for clues in reviews that are relevant to me personally. Whereas with an AI, I can just tell it about my family and vacation preferences. I can store those preferences in a project, add my own reviews on hotels or trips we’ve taken, and get personalized recommendations.

This transition won't happen automatically. At least for everyone older than GenZ, self-sufficiency is deeply ingrained. Unlearning it will take conscious effort. But once you start, you realize how different the future will be. Have fun.

My first “leverage AI” conversation was with Nathaniel Emodi, the cofounder and CEO of Highlight. Inspired by the one and only Patrick O'Shaughnessy, Nat has developed a morning routine leveraging Claude Projects that effectively functions as an AI-powered executive assistant to set his day on the right path. As he wrote:

Personally, as a busy founder I've been using Claude's Projects feature to help organize my day. I have a project called "Morning Goals" that helps unpack a long, often rambling voice memo and organize it into to-dos for myself and others; ideas for different projects; structured questions and notes for meetings coming up that day; and drafts messages for follow-up with different people.

I thought this was a great “starter” use case to write about because it’s one of those ideas that’s tangible, and once you start doing it, you start to come up with more and more ideas for how to leverage Claude or ChatGPT Projects.

For those of you unfamiliar with Claude Projects, Anthropic describes it here.

Nat’s project is “Morning Goals”.

He provided the following instructions, which Nat encourages you both to use yourself and share your feedback on if you have ideas for making it more useful:

**Morning Goals**

As a busy founder and CEO, balancing family life and the fast-paced demands of running a startup, help me organize my thoughts and drive daily productivity.

**Daily Structure**

Every morning, follow these steps:

1. Date & Overview:

Note the current date and time in [YYYY-MM-DD] format.

Record my overall sentiment: [Positive/Neutral/Negative + a brief reason].

Identify two primary focus tags reflecting key themes for the day.

2. Action Planning:

Self: Create a checklist of prioritized action items for myself.

Team: List action items delegated to others, prioritized by importance.

Carryover: Include follow-ups from previous days, noting their original date and any incomplete tasks.

3. Open Questions:

Maintain a running list of open questions I’m exploring, adding context from project documents where needed.

Identify potential people or resources to consult for insights.

4. Key Ideas & Insights:

Summarize key themes and focus areas, adding necessary context.

Propose actionable ideas, follow-up tasks, or challenges to assumptions from the perspective of a [consumer marketplace founder with 20+ years experience.]

5. Draft Communications:

Draft succinct and professional messages for different recipients, tailored to their role or relationship to the project.

Include recipient details, purpose/subject, channel (e.g., email, DM), and keep the tone casual but professional.

6. Knowledge Connections:

Identify connections between current and past notes, referencing prior entries with their date.

Highlight potential strategic implications based on these insights.

7. Outreach Opportunities:

Propose potential outreach strategies, sharing the context of why it’s relevant.

Draft message templates and outline next steps for engagement.

In addition to the instructions, Nat uploaded a handful of documents to Claude’s Project Knowledge, like some Highlight strategy documents, product launch plans, revenue goals.

After dropping his kids off at school, Nat opens Claude and then does a stream-of-consciousness verbal brain dump. Claude transforms this unstructured thinking into actionable output:

• Synthesizes key insights from the stream of consciousness

• List of to-do’s

• Drafts emails based on mentioned communication needs

• Structures everything within the context of company priorities

The result? When Nat arrives at his desk ready to begin the day, he has a structured agenda that allows him to execute on his morning thoughts without the typical context-switching tax. As he described it, it’s like having an EA who not only listens to your morning thoughts but transforms them into an optimized work plan.

It’s a great simple example of how AI can bridge the gap between our natural thinking patterns and a structured productivity systems. Most productivity tools require us to adapt to their structure – here, the AI adapts to us.

Of course, this isn’t perfect yet. The output is static - you can’t check to-do’s off, or keep a running list of all the action items over past entries. And there are still times when it makes mistakes (e.g., when it surfaces prior mentioned dates, they are often wrong). But it’s clear this is just the beginning. As the models progress, Nat notes the experience will evolve from having "an experienced EA to having a Harvard MBA at your fingertips."

If you're looking to start experimenting with practical AI applications, this is an accessible entry point with immediate returns.

As always, if you're building tools in this space or have innovative approaches to AI-powered productivity, I'd love to hear from you. And for those already using similar workflows, Nat (and I!) would welcome community iteration on making it even more effective.

Note: This is part of a series exploring how the best operators are leveraging AI technologies. Stay tuned for more insights from upcoming conversations with AI innovators. And yes, a lot of this blog post was written by Claude. Except that last sentence. And this one.

His approach to technology makes me reflect on where we are with AI today. We're all as individuals at a similar inflection point - a moment where we can choose to stick with familiar methods or push ourselves to embrace new tools that could fundamentally change how we work.

We're already seeing a divide emerging. Some coders have become AI savants, seamlessly integrating these tools into their daily workflows. Some stick to their old ways. Those that are leaning in hard on AI are not just using AI; they're reimagining what's possible in their roles - whether they're desk workers, founders, or CEOs.

I'll admit it: I feel like I'm just scratching the surface. While I've experimented with AI tools, I know there's so much more to do. It reminds me of watching my dad stick to his yellow legal pads - there's comfort in familiar tools and methods, but also opportunity costs we might not fully appreciate until later.

So I thought it would be fun to find a way to learn from these AI pioneers. I want to understand how they're leveraging AI to enhance their productivity and creativity. What daily workflows have they transformed? What unexpected use cases have they discovered? How has it changed their approach to problem-solving?

This will be the first in (I hope!) a series of posts exploring these questions. I want to try speaking with non-coders across different roles and industries who are pushing the boundaries of what's possible with AI. My goal is to share their insights and experiences so we can all better understand how to meaningfully integrate AI into our work lives. At least that’s my goal: forgive me if this ends up being a better idea (to me at least) conceptually than in practice!!

Please consider this post an open invitation: If you're someone who has found innovative ways to use AI in your work, or if you know someone who has, I'd love to hear from you.

And as a first step: I got to enjoy the magic of dictating some rough ideas to Claude for this post, and watching it write 95% of it for me (+ create the title image).

Happy New Year!

It seems inevitable that as the underlying foundation models become more powerful, the LLM players will seek to justify the enormous investment that has gone into training their models by moving "up the stack", and evolve from an API or chat interface, to async agents.

I can't help but think back to Twitter's early incarnation as a platform, which then gradually competed with its platform developers. Right now, OpenAI/Anthropic/et al have an API consumable by almost anyone, but it's not hard to imagine a world in which they begin to compete with some of their API developers. I'd guess the async coding agents are most vulnerable to this potential in the near term given the seemingly unbounded economic value of owning this use case, and the already existing product market fit LLMs have found with coding.

But this will extend beyond coding. The most AI-optimists amongst us (Leopold Aschenbrenner does an excellent job articulating this view) believe that as the underlying LLMs get more powerful, they will get to a place where they can power “drop in” async AI workers that will act like super intelligent remote employees with code-creating superpowers. In this view, the AI workers obviate the need for most AI application software.

As an example, imagine a future enterprise buying decision: Why buy a specialized AI application that lets you automate internal IT ticket response, when the foundation model companies offer an AI agent that if you point it in the right direction with a job spec, will read your knowledge base, build its own integrations to connect to your existing systems of record (e.g., Jira), and then handle all the internal requests automatically?  

Some might laugh at this scenario, but I’d suggest that if you are a B2B founder building an AI-native application, you NEED to do the thought experiment of assuming it over the next 3-5 years as you consider the strategy for your company. Not just because of the risk of this scenario happening, but because any progress down the path of the scenario will meaningfully increase competition for your company (as I describe in #3 below). So how do you future proof your B2B AI application layer company? 

The best answer I have, in the face of this fast-changing future, is three-fold:

1. A network effect. If you've got one, run like hell to get the flywheel to tip. And email or DM me :). By the way, the last investment I led was because of a cold email from a founder who read one of my posts, so I promise you: it works.

2. Capture some proprietary data or hard to access data, either that you’ve accrued as you grow, or that you have access to through some other means. This forms a moat.

3. Execute like hell and land grab in an overlooked vertical. The foundation model companies will inevitably focus on the big markets (e.g., coding, as discussed). But outside of that, it’s hard to imagine the foundation models ever develop a GTM and packaged offering to go after the smaller (but still large!) verticals, which do require more care and packaging for a less sophisticated customer. So if you are going after these other verticals, assume it will be more symmetrical warfare with other focused startups. The difference is that as the underlying LLMs continue to improve, it will become a lot easier for other startups to compete. Imagine what a “LLM Wrapper” startup can accomplish now vs two years from now. So you have to assume more startup competition and more homegrown competition. For example, eventually, it might just take one employee to decide to train an Anthropic agent to compete with the offering you took years to get right. Being obsessively customer focused is always critical. If anything, that obsession will lead you to finding more workflows that you automate faster — which means you’ll add more value out of the box than anything else. That might just be enough to hang your hat on.

Taking a step back, as LLMs progress, they are able to handle more complicated tasks. If today LLMs can handle tasks that would have taken a human thirty minutes to complete, as LLMs progress, they'll be able to handle increasingly complicated tasks that would have taken a human more time. In the next decade, they should be able to handle tasks that would take years for a human to do. Therefore, as the LLMs become more and more sophisticated, the economic value that they will be able to unlock becomes greater and greater.

For example, annually, it is estimated that we spend $1T on software engineers globally. When people talk about GitHub Copilot, you hear people throw around numbers like 10-20% productivity improvements (of course, GitHub claims higher). That translates to $100-200B of value annually were it to be fully deployed (of which GitHub would capture some percentage).

As LLMs progress and are able to go beyond code completion ("copilot") to code authoring ("autopilot"), there is almost no limit in value creation as it would dramatically expand the market – a potential multi-trillion dollar opportunity if someone emerges a dominant player. And that's just coding. We've all experienced the productivity-improving benefits of LLMs (or been on the receiving end of an automated customer support response). The potential value creation and capture with AI is beyond our existing mental models.

The challenge is the amount of capital required to train each successively more sophisticated LLM increases by an order of magnitude, and once a model is leapfrogged by another, the pricing power of the older model quickly falls to zero. There are now more GPT3.5 equivalents for a developer to choose from than would make sense for them to test. Not surprisingly, when GPT3.5 launched in November 2022 it was head and shoulders ahead of any competitive model and cost $0.0200 for 1000 tokens. It's $0.0005 now – 2.5% of its original pricing in just 1.5 years. I can’t remember another technology that has commoditized as quickly as LLMs. It’s a dynamic that makes it almost impossible to rationalize any ROI at this stage in the game because any investment in a LLM is almost instantly depreciated by the next version. But you can’t really skip a step. You need to go through countless worthless versions to get to the ultimate (the idealized “AGI”).

So you have a bit of a perfect storm:

1. The economic value you are able to unlock as models become more sophisticated should increase significantly with each upgrade of the model. The economic value of AGI is constrained only by our imaginations. 

2. Pricing leverage comes from being a step function ahead of the competition, at least along some dimension. If you fall behind, the value of your model to external customers gets rapidly commoditized (of course, there is still value for your internal use cases).

3. MSFT, GOOG, and META have core businesses that produce fire hydrants of cash, Anthropic has found love with GOOG and AMZN, and OpenAI should continue to be able to raise money from sovereigns that have their own (more physical) fire hydrants of cash. 

The net result is that in the short term, until an efficient frontier is reached on the marginal value of continuing to invest in infrastructure with the existing transformer architecture, or we run out of electricity, or a group pulls ahead with an untouchable lead thanks to some smart algorithmic work, investment in this space by these giants should continue to increase dramatically, and costs necessarily precede revenue. The prize is theoretically so large, and if a clear winner emerges, their market opportunity so uncapped, you have to keep increasing your bet.

We all are massive beneficiaries of this battle playing out. The extreme pace of investment in infrastructure / training / etc, combined with the urgency that only comes from intense competition, is giving us all the gift of an insane pace of innovation with models that are able to handle increasingly complicated tasks at bargain basement prices. Applications that might not be possible today, let alone economic (such as most voice and video applications), will be profitable before we know it. Giddy up!

Yes, it's certainly true that as the foundation models progress from 3 to 4 to 5, etc., we will mark time in retrospect with these milestones and how each step-function improvement unlocked increasingly complicated tasks that can be automated by LLMs. What feels different here is that it’s also true that single research papers will unlock new opportunities.

To take two recent examples:

What preceded ElevenLabs? https://arxiv.org/abs/2305.07243

What preceded Krea.ai? https://latent-consistency-models.github.io

The combination of both broad-based (foundation model upgrades) and narrow (research breakthroughs) step-function changes will continue to unlock brand new AI opportunities.

As Ben wrote in his comment on my last post:

One concept I like is that while the raw capacity of something like an LLM is increasing continuously over time, there's a hard threshold at which it crosses from being [not at all useful] to [useful] for a given application. Until we get true human-level AI-generated audio, ElevenLabs is impossible...but the second we do, it's a 10x improvement. Feels like part of the reason it's harder to spot these opportunities in advance.

So if you are a founder worried you’ve missed the window, don’t. It’s a land grab right now, but a single research paper can mean the difference between [not at all useful] to [useful] and therefore a new opportunity unlock. Obvious in hindsight, but tricky timing to predict. It's going to be an exciting (and wild) few years.

Amazon, Netflix, eBay, Uber, Airbnb - when each launched, they seemingly targeted small markets, and yet had insatiable demand for their offerings. The secret, obvious in retrospect, was that they provided a 10x better experience than the incumbent substitute, and provided that service cheaper by using technology to have a structural cost advantage over the incumbents. 

In each case, almost everyone underestimated the potential opportunity of these companies. The “miss by a mile” mistake is that when you use technology to create a step-function improvement in an experience and are able to provide that service at a markedly lower cost, the demand and potential use cases for that product explodes. 

Which brings me back to AI. 

What if, instead of having to find an artist on a freelancer website, go through alllll the effort of selecting someone who has the style you like, has great reviews, actually responds to you, and then iterating with that person to get to the design you want over days if not weeks, you could quite simply describe what you want with words, and in seconds, an image is generated at a small fraction the cost? That’s MidJourney. 

Or what if, rather than having to find, hire, and manage human translators, you could instantly have your documents or application translated as quickly as words are written in whatever languages you want, at a price impossibly cheaper than hiring a human. That’s DeepL (a Benchmark company). 

HeyGen opens up a market that few would have bothered with given the cost and friction - hiring an actor for your company. The hours and hours of effort of hiring an actor, staging the production and post-production of a shoot…. Compressed into five minutes, with limitless iteration on scripts and other control mechanisms moving forward. 

Companies like ElevenLabs provides a 10x better experience than hiring voice actors (instant audio, none of the overhead of hiring, none of the cost of recording). And because it's built with AI, it provides that experience significantly cheaper than hiring humans, thus dramatically expanding the market opportunity for voice acting.

What’s critical in each of these cases is that these are not explicitly productivity improvements to existing employees for existing workflows. Instead, they take markets that were constrained by all the effort, friction, and cost of hiring and working with people, and unlock them. Exactly the type of market that gets overlooked by an incumbent.

If the last couple waves of startups felt like 10x improvements, AI provides what feels like a 100x better experience than the incumbent substitute (humans!) by compressing what is almost always the significant effort of hiring and managing another person, into a near instant experience that will only get better over time. To do this at a small fraction of the cost of hiring/managing that human dramatically opens up limitless use cases and therefore dramatically expands the market. If people underestimated the size of Uber's market initially, we're all underestimating the size of many AI startups' opportunity.

The fun question to ask is which categories are next?As I wrote about before, it’s easy to go down this list of Service sector value added to the Gross Domestic Product (GDP) of the United States of America in 2021, by industry, and start to imagine where the startups might emerge within each of these high level categories… And as always, if you are building one of these companies, I’d love to meet you. Sarah at Benchmark dot com.

As I write this, I am once again aware of the 3rd rail I am touching – the fear that AI will replace humans over time. With every new technology, there is this fear. And time and time again, we see that while the new technology does indeed take on the work that humans used to do, it also creates new work for humans that lets them take advantage of what is most special about being human - our originality and creativity. As much enthusiasm as I know I have for AI, the reality is that we are still very early with this technology, and there is still near infinite work that only humans can do. AI gives humans even more space and bandwidth to achieve those pursuits.

I am grateful for all the conversations I've had with founders tackling "sell work, not software" opportunities.

Some follow-up reflections from those conversations. As always, if you are working on something in this space, I'd love to hear from you (sarah at benchmark dot com).

Reminder: Selling work opens up markets that weren't attractive for software businesses

One of the core hypotheses behind "selling work, not software" is that when you sell a 95% productivity improvement (vs squishier productivity improvement you sell against with software), you are able to charge *substantially* more for your service than you would have otherwise been able to if you sold software on a per seat basis. This can mean that the same end markets could be 10-50x larger than your software-model intuition would tell you. As I mentioned in my previous post, I’d guess a good test of the viability of a market opportunity to sell AI-built “work” is, crudely, whether there already exists a focused, outsourced group internationally to support it. 

It’s not about selling "virtual employees". It's about unbundling the specific tasks or outcomes that employees have traditionally performed, and selling those.

I've seen a bunch of companies that talk about selling "virtual employees". My gut is this is a limiting mental model. First, the reality is that LLM technology, while seemingly magic, is still early. It still performs best when given a limited scope vs the more general variety of tasks an "employee" typically handles. Second, how do you measure and compare the best employees? It's squishy. I prefer to think about what the skills or “jobs to be done" are, and what is the success metrics for each job. As you scale, you add more "jobs to be done" modules, which gives you more advantages with scale. This may eventually bundle back up into a reconstituted "employee", but by the winner, not the entrants.

It's ok to have a human in the loop - having it be *your* human in the loop (vs the customer's) creates clarity around the selling work model.

The reality is that LLMs are still early, and especially in cases where preciseness is important (e.g., legal use cases), and that preciseness requires a human in the loop to extend the surface area of what's possible. I think embrace this. But when embracing this looks like having a customer’s employees in the loop, it pulls you back into a world of selling software. On the other hand, if it is your own employees or contractors that are in the loop, you not only maintain more clarity on your business model of “selling work”, you also can ensure that those humans don’t just perform QA but also act as power users to give explicit feedback into your system to help it get better faster. Your job is to make progress towards automating them away.

Do these look like VC or PE opportunities?

I believe there will be both. The companies that align with the VC model are going to be ones that have the potential to escape competition -- to be dominant in their market. I've written about this here. My hypothesis is that the winner in a market is going to be able to keep on adding more "jobs to be done" automations, so eventually they are able to have the largest breadth of automations vs any competitor, giving them the economies of scale classic to software and the potential to escape competition. But as I said recently to one founder, to do this, you need to run like the wind. This is a skills land grab and there is going to be a lot of competition in the most obvious categories.

An "earned secret" even more important for selling work. Hard to see the opportunity if you aren't from the vertical.

I'd guess that the founding teams going after "work" are going to look different than cliche "college drop-out" profile. They are going to be teams that have lived and breathed a problem in a vertical, and see the wedge in the form of the first task to go after.

Enter Large Language Models (LLMs). The first tranche of products and startups leveraging LLMs has kept within the mental model of selling software to achieve step-function improvements in end-user productivity. The "Copilot for [x]" trend reflects this mental model. While there are fantastic startups innovating to improve employee productivity, LLMs create an opportunity for startups to look beyond this way of thinking and discover surface area that previously was out of bounds for selling software given the required GTM and pricing limitations of software. To do this, rather than sell software to improve an end-user's productivity, founders should consider what it would look like to sell the work itself. 

Subscribe now

Selling work opens up new vertical opportunities that wouldn’t have otherwise supported a software company. Take EvenUp as an example (who I have no doubt will dominate their vertical). If you are a personal injury lawyer, a work product you create on behalf of a plaintiff is called a demand package. Essentially the demand package is a summary of the case, the medical costs of the injury (including lost wages), and then a recommendation on the settlement value from the defendant’s insurance company. Law firms have stretched lawyers, paralegals, or outsourced groups writing these documents. 

If you were still in the mindset of selling software, you could imagine a software offering for personal injury law firms, sold on a per-seat basis, that uses AI to help people in the firm create a demand package (imagine a builder where you drop in the medical records during one step of the process). But then EvenUp would have been stuck in the paradigm of selling software – selling a 10% productivity improvement instead of 95%. Instead, EvenUp had the foresight to sell the work product itself: the demand package. 

When you sell work, the sales cycle is different, it’s priced relative to the cost of a human performing the work instead of as a productivity improver, and the competition for a similar product (besides the company’s own human capital) is essentially outsourced groups internationally. An AI-driven product with the consistency and SLAs it can achieve should be vastly superior to an outsourced offering – a 10x and cheaper opportunity. Indeed, I’d guess a good test of the viability of a market opportunity to sell AI-built “work” is, crudely, whether there already exists a focused, outsourced group internationally to support it. 

For example, just to spark some ideas, here are some of the BPO services for an outsourcing provider:

And here are some examples of Legal Process Outsourcing options:

Any of these, I imagine, are vulnerable to automation leveraging AI.

As I write this, I’m aware of the 3^rd rail I am touching – the fear that AI will replace humans over time. Here, I look to EvenUp as an example. When the lawyers and paralegals are freed up from putting together demand packages, the customer (the plaintiff) benefits from better demand packages, and the people in the firm benefit from being able to shift their time to less mechanical things like client services, getting more customers, or of course the finishing touches on the demand package. 

If you see any other examples of this, I’m all ears. And if you are building a company leveraging LLMs to sell work, I’d love to hear from you. My X DMs are open, and my Benchmark email address is sarah at benchmark dot com.

Also, it’s not enough to sell work — you must escape competition. To this end, I’d also suggest my prior post How to escape competition -- Building enduring application-level value with LLMs, so you don’t fall into the trap of providing a service that gets all the margin squeezed away.

Mitochondria … act as the power plants of the cell.— Nature.com

I believe all employees in a startup are on a spectrum between two poles:

On one side, there are employees who think of the company like a job. They come in, and they do their job. Some excel at doing their job. But ultimately, it’s a job. They want to make sure they’re fairly compensated and appreciated for their work, have a manager who invests in their growth, and like the work they are doing. As long as they believe those things are in balance, they’re stable and stay. In summary, they start from the perspective of their own optimization, are rational actors, and the value they add to the company, while valuable, scales linearly.

Then there is the other side of the spectrum. This group has a different DNA – they act more like founders than employees. They pour their passion into the company because they believe in its mission, and it is how they operate. They add value to the company beyond their job description and responsibilities. They ask and do what is best for the company. They work hard and late, because for them, the company isn’t a job. Most of all, they best embody the company’s values, and because they do, their value is not linear: they energize and power startup teams through good times and bad. I think of this class of people as the mitochondria in startups.

When capital markets are loose like we all experienced in 2021 and growth is easy, it can be difficult to distinguish between employees who are there for a job, and mitochondria. But when growth becomes more of a grind, when each incremental hire is evaluated more carefully, now is the time to make sure you are attracting and retaining mitochondria.

At the early stages, this rare group of individuals is the core of the company. As your startup scales, they are your leaders. This unique DNA should be cherished, recruited, and as much as possible, the mindset encouraged as much as possible in other employees. This starts with three things:

Awareness

Awareness is always the first step. Who are the mitochondria in your company? Make that list. Invest in your relationships with these people, and in the people themselves. Stay informed about who falls into in this group and what they are thinking.

Value interviews

The values of a company should be a blueprint for how to be successful in that company; the canon for how the team behaves and makes decisions. As I mentioned earlier, mitochondria aren’t just good at what they do — they best embody a company’s values. It follows that in order to maximize your chances of hiring mitochondria, you can’t just interview for competence. You need to interview for values.

A values interview helps you determine if the candidate is going to be a match with the core values of the company. These interviews should be done by the founding team, or as you scale, by mitochondria that are already working at the company (takes one to know one).

To be clear, interviewing for values doesn’t guarantee you’ll hire mitochondria. But if you don’t interview for values nor communicate the importance of your values during the interview process, your hiring process is suboptimal at best. Just like actual mitochondria, mitochondria employees won’t scale directly to total employees if you don’t make efforts to replicate them.

Be ready to break “bands” / “levels”

At some point, a company formalizes its compensation structure — what it offers to new and existing employees based on the job and experience level.

The biggest mistake I’ve seen startups make is using the same comp scale for these two classes of employees. Comp levels are a rational structure to overlay on an irrational commitment and impact. Companies that don’t break bands will undervalue the energy-giving impact the mitochondrial employees have, which will ultimately challenge their ability to motivate and retain them.

I realize this suggestion might raise concerns of creating inequality throughout the team. To be clear: I’m talking about a very small percentage of your employee base, and for them, their outsized impact should be self-evident.

That said, this is a mindset you want to encourage in all employees. One way to do this is emphasize values or impact outside of your job description as part of the performance review process. For example, at Pinterest, we had a 2x2 where one axis was “Impact”, and the other was “Values”. According to my former Greylock partner John Lilly, at Mozilla they talked about impact across project and company. I suspect either way works — the importance is emphasizing something bigger than the individual, particularly in the most critical stages as a company scales from a few dozen people, to several hundred. As startups scale to companies, it becomes harder and harder to make this part of the culture.

In asking “Who do we want as cofounder?”, Victor captured our imagination in three ways:

• Radical understanding of the founder’s journey. Having been turned down by every investor in Brazil, Victor took a mere $100 and turned it into a transformative gaming company that generates hundreds of millions in annual revenue.

• A go-to resource for some of the best founders of our generation, from Dylan of Figma, Pedro and Henrique of Brex, Alex at Scale, and Howie of Airtable. Founders feel comfortable and supported in bringing Victor their biggest challenges and being vulnerable with him.

• Refreshing, low-ego curiosity that allows him to traverse with alacrity areas as wide ranging as AI gaming to fintech. We realized we often found ourselves asking for his perspective to sharpen our own.

Our enthusiasm to work with Victor comes from the bright signal of working together through the challenges of building a company, when success is far from obvious, and deeper values emerge. After investing and working through the transformation of Wildlife with Victor, we saw first hand the depth and magnitude of his capacity as a partner, thought leader, and human. That perspective gives us complete conviction that he will be the partner of choice for the next decades of iconic entrepreneurs.

At Benchmark, we strive to offer a radically different level of service and commitment to the founders we serve. Done this way, venture does not scale. To do this and continue to perform at the level we aspire to, each person we add to our partnership must raise the bar and make us as a partnership better. We have the highest conviction Victor will not only excel at the multi-disciplined craft of early-stage venture investing but also push all of us in our thinking. We can’t wait for you to spend time with him.

Peter, Eric, Sarah, Chetan, and Miles

You can follow Victor on Twitter at @victorlazarte.

We ended up shipping something that was, honestly, weak, and the metrics showed it almost immediately. The growth team knew that they had to re-do it, but as sometimes happens, redoing it got deprioritized and Pinterest went through a period of hyper growth with an underperforming NUX.

A couple months later, I was curious to remind myself of Twitter’s NUX and I remember signing out and going through the new user flow. Want to know something funny? They’d updated it… to something similar to Pinterest’s. The NUX that we knew was under performing the previous NUX. The NUX we knew we desperately had to redo.

And you can imagine exactly how it happened. Just like we looked at Instagram’s assuming it was God’s gift to NUXs, Twitter looked at ours, assuming we knew what we were doing. Narrator: we didn’t. :)

This mistake in logic happens all the time. A company is successful and it’s easy to assume that all the details about that company are part of its success. You might aspire to constrain yourself to the number of employees Instagram had when they got acquired, or build a fully distributed team like Gitlab from the start, but unless you are able to understand from first principles if that decision is the right decision for you, be careful taking it at face value.

Yes everyone felt that the quality of Twitter’s algorithmic feed was decaying, but for me at least, to actually watch my average view count per tweet decline over the last few months to a fraction of my following count is all the demotivating confirmation I needed that something has changed for the worse.

As with any social system, your creators have to invest their time and energy into the network by creating content, which in turn creates content for your viewers (and liquidity for your advertisers). But if you no longer believe in the ROI of that time, why keep investing? And nothing about Twitter’s new leadership makes you feel like the delicate balance one needs to strike in an algorithmic feed is going to be found once again.

That feeling of demotivation and decay creates the kindling for a new social product. Of course, there have been several startups betting on Twitter’s demise (BlueSky, Mastadon, Post News, etc). But I’ve been skeptical of these offerings, as talented as the founders are and well executed the products, because they didn’t introduce a new atomic unit.

Enter Threads. Same atomic unit, and yet it’s fascinating to me how alive it already feels by porting Instagram’s graph, an algorithmic feed that works, and of course leveraging Meta’s infrastructure and unlimited budget. If all other social products have to tackle the cold start problem, Instagram gave Threads a running start.

Exhibit A with Reels and suffice to say, these guys aren’t going to give up, which spells obvious trouble for Twitter.

To be clear, this doesn’t feel like an either/or situation that is going to be determined in the short term. Threads *is* a different graph than Twitter. And anyone who wants to can cross post easily. But certainly, my quick take is that much like Instagram has been chipping away at TikTok’s atomic unit and user time with Reels, Twitter’s decay just got accelerated by Threads. Time to start cross posting. :)

Surprised, I asked how he created the list.

His answer? “They’re the ones who reached out to me.”

I laughed. “You’ve got to take control of your process and figure out who you want to partner with. Right now, by relying on inbound, you’re sorting for the largest firms because they’re the ones that have the resources to have junior people cold outreaching.” We then brainstormed a better list.

Reflecting on it, I realized there is a funny quality with some things where what you think is correlated with the qualities you want, may actually be an anti-correlation. For example, let’s talk plumbers…. 🙂 Your water heater breaks and you *think* you want the plumber who can come right away. But for anything more complicated than a clogged drain/toilet, you don’t really want the plumber who can come tomorrow. What you want is a high-quality, ethical plumber. Turns out, those plumbers are probably booked until next week (trust me on this!). As I learned the hard and expensive way, the extra week of suffering is worth it.

I can’t tell you how often I see this pattern. Indeed I see it with my own companies when we kick off a process to do a follow-on. It’s so much easier just to look at the people who are reaching out to you, and indeed it’s a great place to start. But the decision of who you partner with is so important (particularly at the early stages), you owe it to yourself not just to look at the list of people who reached out, but to ask yourself, "who do I want?”, and then find a way to them. 

[Fell off the weekly-post horse when the SVB stuff went down, but trying to get back on. Will need to feel out pacing…]

It seems clear that over the next 5+ years, any work for which a human is a critical input into the work product is vulnerable to substitution with a software product built leveraging large language models (LLMs) and some of the related technology. 

Copywriting was the first visible category of work that startups leveraging LLMs went after, with Jasper and Copy.ai being the first two startups going after the opportunity, and wow did they grow. There are countless other examples like this now. It’s very clear that if you are able to build a product that uses LLMs to automate a work product that has previously required hiring someone for that job, the demand will be there.

But that doesn't mean it's been all sunshine and rainbows for these companies. The critique that has haunted both companies (and many more) is their defensibility. This question takes two forms:

1. If anyone with access to ChatGPT or any of OpenAI’s APIs could essentially get to the same output, you are always vulnerable to customers moving their business to whomever offers the work product at a cheaper price. As an example, both Jasper and Copy.ai got caught on their backfoot when OpenAI released ChatGPT, which basically democratized access to the service they were offering, spiking churn for both companies.

2. Won't the incumbents just add this? We've see Notion, Hubspot, Canva, Microsoft, etc., quickly announce GPT-driven features in their products. It becomes the classic race of either the startup figures distribution or the incumbent figures out innovation. Here, the "incumbents" aren't sleepy companies but innovative tech companies, and the innovation necessary, while not trivial, has mostly already been done by OpenAI so it's just a matter of fitting it cohesively into the product.

I personally think these critiques underestimate the value of focus, and don't give the teams enough credit to execute. The reality is that very few software companies have ever had a technical moat — it’s always been about focus and execution. I do however think that the second critique speaks a bit to how we’re in a “skeuomorphic” generation of LLM-based applications where the first obvious applications use the surface level functionality of what’s possible using a new technology. So - how can Jasper/Copy.ai prove the naysayers wrong? Or more generally, how do you build enduring value if you are a startup looking to leverage LLMs to create a new application?

I think this will take a few different variations that increase the odds of escaping competition:

1. Narrowness in initial focus. One of the areas I’m most excited by right now is companies pursuing vertical application opportunities. In a world with multiple competitors seemingly focused on the big horizontal opportunities, and still rapidly evolving underlying technology, a focused competitor can win and then can expand from that position of strength. These will often involve tuning a model to a specific use case, hooking into if not replacing existing workflows (often times leveraging other ML techniques to do so), and therefore means that there will be more to the execution than a simple API call to a foundation model. 

2. Feedback loops. I’ve written about feedback loops earlier so won’t belabor the importance here. Certainly if you build an application that can leverage user engagement to improve the accuracy of your model, there will be advantages to scale and thus the ingredient to escape competition. Having a human in the loop that acts as a bit of a power user to provide feedback in the beginning is another mechanism companies use that is also effective in providing advantages to scale, provided you are able to leverage that feedback to fine-tune.

3. Accruing data asset. I’ll confess I’m most taken by companies where a positive externality of users leveraging their LLM-driven application is the creation of a new, useful data asset that wouldn’t have been possible before at scale. In a way, this externalizes the moat outside of what’s possible with LLMs themselves and creates an even more differentiated offering that can escape competition at scale. 

There is no question to me that we are at the beginning of a strong current that is going to transform the landscape of work. If you are building something here, don't be shy. My DMs are open.