Ivan Bercovich

Software started as something you delegated to someone else because it was too hard to build yourself. Adam Smith 101. Mining companies don’t build their own bulldozers. Why? Because it would be cost prohibitive to build a handful of unique bulldozers, so implicitly all bulldozer buyers agree to fund an oligopoly of bulldozer manufacturers. Same with planes. Software started the same way. Every company needs bookkeeping. Outsourcing it to Intuit is win-win. But then software started eating the world. Rather than outsourcing arbitrary functions, vertical SaaS became the “system of record” for specific industries. At first glance this is not that different from bookkeeping. But it is. If every company in an industry converges to identical processes, there’s less differentiation. DoorDash substitutizes restaurants. The same happens for vertical B2B. It’s worse than that. The best run companies in an industry are subsidizing their competitors by providing better feature ideas and data to the SaaS companies that serve everyone with the same product. If software were cheaper and individualized, successful companies would want their own. Just like United would prefer having exclusive supersonic planes if they could (and in fact have a big order with Boom). The future of one-size-fits-all software is predicated on whether it is better than custom software. Think about this. How many people love their B2B software? Not many. Fair or not, employees often feel burdened by the workflows forced upon them by a piece of software someone else decided to buy. For vertical SaaS companies to defend their positioning, they will have to deliver supersonic airplanes.

The moats will be at the intersection of software with X. Software x Autonomy, Software x Logistics, Software x Atoms, Software x Money, Software x Military, and so on. The value will come from deeply integrated software, rather than from selling the same reusable software to everyone in an industry.

Five years ago, 0.4% of the population was actively employed as a software developer. What percentage of the population will be producing software more or less full time, five years from now? Think about it. What would a financial analyst do all day, but ask Claude to write disposable programs to analyze data? It won’t be everyone, but it will be a lot of people. I think 4%, 10x more, is a reasonable, if conservative, estimate. So we will have ten times more people writing software, and each of them will be 10 times more productive than those writing software five years ago. So 100x more software. How will that software be distributed? Today, most software is written by professionals hired by mostly software companies to write their software. When everyone can do so, it will be more evenly distributed across industries. Every kid coming out of high school will at some point contemplate just starting a vibecoded/agentic business. Why go to college anyway? Where will these apps run? What percentage will be free? Here are some predictions: - 40% of net new software will be disposable programs (less than one month of use) - 20% will be for free or nearly free applications for personal or internal company use - 20% will go to automating business processes in previously hard-to-automate industries (this will be the most valuable) - 20% (this is already 4x) will go to pure-play big software companies Point is, having 100x more software doesn’t mean pure-play software companies are 100x larger. It just means software becomes more abundant and more diffused across the economy.

A big challenge I have when trying to help someone grasp an aspect of AI is that people tend to think of AI as some homogeneous thing. But what is actually happening is that we are speedrunning through many paradigms, which last just a few months each. Understanding the paradigm is essential in anticipating second and third order effects. For a while in 2023, ChatGPT, a 175B parameter LLM, was "AI" to most people. We learned that GPT 3.5 was an all-in-one NLP tool capable of translation, sentiment analysis, summary, and also capable of writing silly songs, poetry, and jokes. Most people never used it, and of those that did, most only used it a few times. So when GPT-4 came out, a model 10x as large, and it was much better at deep reasoning, it barely changed people's perception (Google Trends shows interest peaked during launch and didn't reach the same level again for 2 years). By the time people started to appreciate GPT-4, those of us who were paying attention were experimenting with tool use (I wrote a misunderstood Slack bot years before Clawdbot). While the average "expert" was talking about prompting, the frontier circles were already talking about agents. Agents were a complete paradigm shift. We went from supplemental intelligence for our own reasoning to having AIs that operate on their own. Even as the paradigm changed, models weren't good enough, so again they were dismissed as "stochastic parrots." I'm not talking about developing-world farmers, either. It took until 2026 for the geniuses at hedge funds to question the prospects of AI in a world of infinite software written by agents. Now, everyone seems to have vibecoded an app (the actual number remains minuscule). You might think that finally, you get it. But the paradigm keeps evolving. At the frontier, we are thinking about 1000-agents swarms that will write so much code it would be impossible for a human to look at any of it. The same is true for second, third, and Nth order consequences. By the time you finally grok that if agents can code, software will be more abundant, you only moved one step up a long ladder. And it's important everyone is trying to climb this ladder, quickly, because AI will have a massive impact on the economy and society as we know it, and it's going to happen very fast, within a decade. So everyone should be thinking about the implications of having AIs that are 100x better than the best AIs today, at everything, and cheaper, and faster. And datacenters that consume as much energy as 100 Hoover Dams. And a consumer economy where a meaningful fraction of purchase decisions are never approved by a human. And so on and so forth ad infinitum. Decades, or centuries worth of economic shifts within our lifespans. The final destination might be incredible, but it's highly unlikely that this amount of change goes easy on our social norms, institutions, legal frameworks, geopolitics, etc. So, please, while this is something only humans can do, THINK.

There's not time like the time before the AGI takes over and turns Yosemite into solar panels. Go out and enjoy nature this summer!

@Mihonarium @NeelNanda5 @OwainEvans_UK It’s an elegant experiment that defines new semantics for us to keep advancing our empirical understanding of AI. These are some of the most valuable papers.

@NeelNanda5 @OwainEvans_UK I’m so confused about what is the development here. Isn’t that just obviously what LLMs do? Like, the default way they think when without chain of thought?

Out of context reasoning is one of the most fascinating developments in the science of how LLMs work. This primer by @OwainEvans_UK, one of the main discoverers of the phenomena, is a great introduction

Vibe Revenge: If someone rubs you the wrong way, vibecode their startup over the weekend and offer it to their customers for free. If a colleague crosses you, automate their job over the weekend and share the news in a team email.

@krishnanrohit Did you try building a graph of email relationships first purely based on who emailed whom and use that (or similar static analysis) as aid?

We live in a world of Big Software. Relatively simple products are capturing an enormous amount of upside. The first word processors were built in the 60s. You can select among hundreds of awesome free word processors for your computer, in addition to the free one included by default. And yet, enterprises pay Google a lot of money for a word processor in a browser. Same for email. These things were awesome innovations at some point, but they haven’t changed very much. Yes, they are on the cloud, but storage and CPUs are cheap. Why do businesses pay so much for this? Simply because they’ve gotten used to the status quo. That peace of mind costs money, and so you have to pay $20/user per month for practically any tool your employees need. The argument that businesses wouldn’t pay if it wasn’t worth their time is incorrect. With most SaaS purchases, we are talking about a middle manager who is not doing some deep analysis. Finance departments are accustomed to a certain portion of P&L going to software. This reality hasn’t been questioned in a long time, and in the meantime, software costs have become a bigger component of operating a business. And unlike consumer products, you rarely have employees raving about how much they love their SaaS software. So, as people wake up to the possibilities of what their own employees can do with AI, the status quo will vanish, expectations will go up, pricing power will decrease, and the market will evolve much faster than it has over the past 10 years.

If you want to build a large software company, you should build it for agents, not people. There’ll be more agents, moving faster and spending more money than people. People will build their own apps for their own needs and to express their creativity. Agents will need resources like infrastructure, payment rails, and so on. That’s where the venture-scale returns will be.

I've been doing this a lot. Sometimes it's too much work to explain in plain English what I want. So I have the artifact that's being generated include affordances so I can make manual adjustments. This is for an interactive bookshelf I am building for my personal website.

@crispheaney I thought you vibed more like a rhombicosidodecahedron

You the avg of your 5 closest friends & aint no squares in my circle

@martinvars Spain seems to have this colonial idea that Spanish is a universal language like English. I remember a Spanish family in 2001 going to a McDonald’s in Prague and insisting on talking Spanish to the cashiers who were clearly confused.

Estoy en Argentina Week en NYC escuchando a numerosos líderes de Argentina, empresarios, gobernadores. Casi todos hablan un inglés impecable como el del embajador Alec Oxenford y el canciller Pablo Quirno. Ahora la pregunta. ¿Por que en España los líderes, la elite, habla un inglés tan rudimentario?

Human relative advantage can still be dystopian. Consider a human with relative advantage of dexterity who doesn’t have plumbing training. Same human now wears AR glasses with earphones. A cheap AI tells the human exactly what to do. The human becomes a meat robot. All physical labor trends towards being an Amazon warehouse worker.

A big determinant of AI's job impact is driven by the lack of compute, especially for agentic work, which takes a lot of it. That makes AI expensive. So companies will only want to burn compute on high-value tasks (eg coding), because, in other jobs, humans remain much cheaper.

@justinskycak Isn’t it just mimesis all the way down?

It is better to begin building a messy version of the life you want than to settle for a nicely packaged, off-the-shelf version of someone else's.

@HalderAlaka It definitely obliterates flow.

@neversupervised That kind of iteration can exist, but one complaint I've heard is that it feels worse as you're running multiple threads in parallel (or expected to). Must be harder to enjoy the craft when you're context switching or feel that you have to sacrifice perfection for throughput.

1/ Some data science and engineering peers are anxious about aggressive pushes to adopt Claude Code at their workplaces. They feel it's taking away creative work and replacing it with tedium: designing tests, compiling test datasets, evaluating outputs For now, this surprises me

You have to think of the game differently. If you can make a SaaS a month, you can also expose yourself more to luck and see if anything sticks. Yeah it’s unlikely it will stick with zero effort but the equation is very different than when building a rudimentary SaaS tool 5x-10x more time.

How founders feel when vibe coding their 15th SaaS with zero users:

I've always had this perception that a mentor is someone who listens and, through Socratic dialogue, helps you arrive at the answer you already knew was right. But when I've engaged with mentors in that capacity, it hasn't worked for me. I'm already introspective to a fault, and I don't reach out very often, so when I do, I'm often confused and need a stronger push. In spite of not being famous for my charm, over the years I've mentored a lot of people. I wouldn't be surprised if the number of mentoring conversations I've had puts me at the 95th percentile. And my mentoring tends to be very confrontational. It's not unusual for someone to be upset or down after a conversation with me (although that's still the minority of instances). I'm most helpful when someone is looking to strategize a change and needs a brainstorming partner. But if I think someone's current approach is bad, I will say so. And then I will text them every day until they do something about it. Sometimes I reach out to people I haven't talked to in a long time and call them out for a tiny thing they said in a group chat. That drives people crazy, especially if there is some truth to it. One time, I decided someone had to get divorced and basically planned (and funded) the whole process. I'm probably leaving a lot of mentoring value on the table by being this way. I don't recommend it. But I do think that on average, mentors could be more forceful. A lot of young mentees need momentum in any direction to get going. Driving that momentum is better than optimizing for the perfect decision. And if someone reaches out to you, you shouldn't hold back. Whenever I had to have a difficult management conversation, I would start by telling people that while it definitely sucks for them, it's not fun for the manager either. Most managers hate the conflict so much that the feedback comes out languid, which makes it hard to course correct. Being tough with someone you care about is really hard, and is often the best thing you can do for them.

The SaaSpocalypse is one of those things that is easier to understand from first principles than most other changes, and yet people fail to make the leap. It’s basic microeconomics. Software is free. When the internet became a thing, did newspapers become more or less valuable? It’s not that hard to figure out the trend. The specifics and the timeline might be off. The markets might be overreacting. Vibecoders don’t need to copy all of Salesforce. That’s where that side of the debate is wrong. They just need to do the following: •Provide a credible threat that they could do it, to negotiate •Replace individual features as standalone internal products •Buy a vibe-coded competitor that is simpler but cheaper •Reject upsells and bundling because there are so many more options •Use Claude to migrate out of any platform into any other This is a non exhaustive list.