Deep Dive with Cursor's Founder: The Death of Code and the Rise of 'Taste'

The company behind AI coding tool Cursor, AnySphere, recently reached a $9 billion valuation, making it one of the fastest-growing startups in history — achieving $100 million in Annual Recurring Revenue (ARR) just 20 months after launch. In the latest episode of How to Build the Future, Cursor co-founder and CEO Michael Truell shared his radical vision for the future of programming: code itself will eventually be replaced by something better, and the one truly irreplaceable skill for human engineers is taste.

The Ultimate Goal: Not Augmenting Code, But Replacing It

Truell opens with an ambitious claim: the company's goal isn't to make existing programming faster — it's to "replace programming with something better." He and his three co-founders are all seasoned programmers who love coding because it lets them "build things fast." But the reality is that even straightforward features require editing millions of lines of arcane, formal programming language.

He predicts that within the next five to ten years, software development will evolve into a higher-level, more intentional form — where developers define how software should behave and appear, rather than manually writing every for loop and if statement. Cursor's strategy is to always be "the best way to code with AI" at any given moment, and then continuously push that process away from traditional programming until it becomes something entirely different.

Why "Vibe Coding" Doesn't Work in Professional Development

When asked whether we've already reached an era where you just describe and AI generates, Truell gives a measured answer. He acknowledges that in small codebases and small teams, things have already changed — people are "rising above the code" and letting agents handle changes directly. But in the professional world, so-called "vibe coding" (programming without reading or understanding the code) "doesn't really work."

The reason is the "nth-order effects" of complex systems: when you're dealing with millions of lines of code and a project built by dozens or hundreds of people over many years, you can't avoid thinking about the code. He reveals that Cursor users let AI write an average of 40% to 50% of their lines of code — but the process still involves "reading every line of AI output."

He frames current AI-assisted coding in two modes: one where you delegate a task to an agent ("go do this for me"), and another where AI watches your screen and occasionally takes over the keyboard via Tab autocomplete. The game for the next six months to a year is to raise the utility of both modes by an order of magnitude. He predicts that once these mature, roughly 25–30% of professional development work can be handled end-to-end by AI without line-by-line review.

---

Annual Recurring Revenue (ARR) is the core metric in the SaaS industry for measuring the scale of subscription-based businesses. It projects current subscription contracts into an annualized expected revenue figure, excluding one-time payments — making it a better indicator of business sustainability and momentum. For investors, ARR growth rate is a key signal of a company's trajectory. Reaching $100M ARR in 20 months means Cursor's revenue curve is nearly vertical — an extreme rarity in B2B SaaS history. By comparison, Salesforce took roughly five years to reach the same milestone.

"Nth-order effects" in software engineering refers to how a change in one part of a system propagates through dependency chains, triggering cascading reactions across multiple layers. In large codebases, modifying a single module can cause test failures, interface incompatibilities, and performance regressions — consequences that are often impossible to fully anticipate before making the change. This is the fundamental reason why legacy "spaghetti code" is so hard to refactor, and why it can't simply be handed off to an AI for automatic handling.

---

The Real Bottlenecks to Superhuman Coding Agents

Truell calmly outlines the technical obstacles preventing coding agents from reaching human-level performance — a section that's especially illuminating for understanding the current ceiling of AI-assisted programming.

Context windows and continual learning: One million lines of code equals roughly 100 million tokens. The challenge isn't just whether a model can physically ingest this content, but whether it can "truly attend to" that context in a cost-effective way. He emphasizes that this isn't only a codebase problem — it's a continual learning problem. Models need to know an organization's history, what has been tried before, and who the team members are. He candidly states that the industry "doesn't have a truly good solution for this yet," and acknowledges that training institutions broadly lack high-quality long-context data.

Long-horizon tasks: He references a widely circulated chart showing that the maximum time AI can sustain progress on a single task has grown from "a few seconds" to "about an hour" over the past year or two.

Multimodality and tool use: Software engineers need to run code, view outputs, and sift through Datadog logs. "If you didn't need these things, AI would already be superhuman — that would be insane." Computer use capabilities are therefore critical to the future of programming.

---

Continual Learning is a core challenge in machine learning, also known as "lifelong learning." Traditional deep learning models have fixed parameters after training, and when forced to learn new information, they often experience "catastrophic forgetting" — where absorbing new knowledge overwrites existing knowledge. For coding agents, this means they must continuously internalize a specific codebase's history, organizational conventions, and team culture — all without forgetting general programming capabilities. Current mainstream solutions include Retrieval-Augmented Generation (RAG) — storing historical information in an external vector database for real-time retrieval — and fine-tuning, which involves continued training on specific data. Both have significant limitations: RAG is constrained by retrieval accuracy and context window size, while fine-tuning is expensive and prone to forgetting. Truell's point that the industry has "no truly good solution yet" cuts right to the core engineering bottleneck in this space.

---

Taste: The One Irreplaceable Skill

This is the most thought-provoking part of the conversation. Truell argues that no matter how technology evolves, one thing will never disappear: taste — the ability to define what you actually want to build.

People typically discuss taste only at the visual level, but he emphasizes that taste problems exist at the logical level too — how software logic should work. Current programming behavior bundles two things together: figuring out what product you actually want, and mapping that to implementation details on a physical computer. The latter is essentially "human compilation" — you know what you want, but you have to spell it all out in for loops, if statements, variables, and methods for the computer.

His view is that this "human compilation" step will increasingly disappear as computers become capable of filling in the gaps. But the taste for "what is truly useful and aligned with your intent" will never disappear. He borrows a phrase to describe this: great people can help you reach a certain standard, but truly masterful people can reach a height you can't even see yet.

From this, he concludes that future programmers will become logic designers — shifting toward intent-driven programming.

Broader Implications: Developer Productivity and Long-Tail Software

Truell sketches two second-order effects once the technology matures. The first is a massive leap in professional developer productivity. He notes that large software projects are "shockingly slow," largely because of the weight of existing logic — change one thing and a cascade of other things break. When programming is automated, the next distributed training framework, the next database, the next AI model will all be built much faster. And from what he understands, the major labs are largely bottlenecked by engineering capacity.

The second is the emergence of a vast long-tail of niche software. He recalls that his first job was as the first software engineer at a biotech company whose core competency was wet lab science — yet the company had to go through enormous effort to build a software team just for internal tools. In the future, companies like this will have far more options. "Whatever you want to happen on computers — it can happen."

From CAD to Cursor: A Pivot Worth Billions

Cursor wasn't built overnight. The four MIT-bred founders started in 2022 with something entirely different: a mechanical engineering play — building a copilot for CAD (computer-aided design), training "3D autocomplete" models to predict a user's next geometric operation in SOLIDWORKS or Fusion 360.

This "wilderness period" lasted about a year and was eventually shelved for two reasons: the team was far more passionate about programming than mechanical engineering, and the underlying science wasn't ready for 3D — pretrained models underperformed, and CAD data on the internet was orders of magnitude scarcer than code data. But the detour wasn't wasted: the hands-on experience of training billion-parameter-scale models (including forking Megatron-LM and Microsoft DeepSpeed and rewriting kernels) laid the groundwork for the inference capabilities that now handle over 500 million model calls per day.

The core belief that drove their pivot, Truell calls "follow the line" — the scaling laws: given more data and compute, models will keep improving. In 2022, this was a belief that "nobody believed." It preceded Sam Altman's public emphasis on "always bet on models getting smarter" by a full year.

---

Scaling Laws were systematically formalized by OpenAI researchers (Kaplan et al., 2020). They describe a set of empirical regularities: the performance of neural language models (measured by loss) improves in a predictable, power-law fashion as model parameter count, training data volume, and compute increase — with a quantifiable optimal ratio between the three. The revolutionary implication was that "bigger models get smarter" shifted from intuition to an engineering belief grounded in mathematics, providing the theoretical foundation for models like GPT-3 and GPT-4. Truell's 2022 conviction to "follow the line" — against prevailing skepticism — was thoroughly validated by the market over the following two years. Megatron-LM is NVIDIA's open-source large-scale language model training framework; DeepSpeed is Microsoft's open-source distributed training optimization library. Both are designed for efficiently training models with billions to hundreds of billions of parameters across thousands of GPUs.

---

The Key Product Decision: Why an Editor, Not a Plugin

One decision that was deeply unpopular at the time — and has since proven correct — was building a full editor rather than a VS Code plugin. Truell explains it came from the belief that "all programming will flow through these models, the future form will look completely different, and you need control over the UI."

He also shares a secondhand story about the birth of GitHub Copilot: the team wandered in the wilderness for nearly a year, trying all sorts of wild ideas like automating PRs, before finally landing on simple autocomplete. And even then, just to display ghost text (the gray preview text), they had to modify VS Code's main branch and expose new APIs. The Cursor team concluded: "If something as simple as ghost text requires modifying the editor, we'll definitely need to change a lot more down the road."

Moats and Metrics: Beware the Siren Song of Demo Optimization

In terms of growth metrics, Cursor focuses not on DAU/MAU, but on heavy paying users — those who use AI four to five days out of every seven and pay for it. Truell explains that as a tool serving professionals, delivery has real costs, and getting users to "graduate" to a paid tier is the only sustainable path.

He particularly warns against a trap common in AI products: optimizing for the demo. It's easy to splice together a handful of cherry-picked examples into a seemingly revolutionary video — but between a polished demo and a genuinely usable product lies enormous work on speed, reliability, intelligence, and product experience. Cursor's approach is to turn the editor into a tool the team uses every day internally (dogfooding).

On the question of moats, Truell sees the AI coding market as more like late-1990s search than traditional enterprise software: the product ceiling is extremely high and can keep improving for a long time. The data flywheel from scaled distribution is crucial — by observing where users accept versus reject AI suggestions, and how they correct after rejecting, Cursor feeds that signal back into R&D and the underlying models to build better products. He compares the ChatGPT moment to the iPod/iPhone moment of this era, and believes there are still several such breakthrough moments in this space waiting to be claimed.

---

The Data Flywheel is the core moat logic in platform competition: users generate behavioral data by using the product, data improves the model, a better model attracts more users, forming a self-reinforcing positive feedback loop. For AI coding tools, this flywheel is especially critical — the accept/reject signals from users responding to AI suggestions are high-value labeled data for training models that better match real development scenarios, and this kind of data is almost nonexistent on the public internet. The larger the platform, the more such signals it collects daily, the faster it can iterate on models, and the higher the barrier for competitors to replicate. This explains why Cursor, despite competing against giants like Microsoft's GitHub Copilot, can still differentiate on product quality: different user populations produce different behavioral distributions, and the flywheel spins in meaningfully different directions.

---

Hiring Philosophy: Interviews Are Still AI-Free

An interesting detail: despite making an AI coding tool, Cursor's technical screening interviews still only allow candidates to use autocomplete — not AI. The reasoning is twofold: coding without AI remains an excellent timed test of skill and intellect; and they've hired many top programmers who didn't have experience with AI tools and don't want to unfairly disadvantage them — these people, in turn, contribute valuable "beginner's perspective" product insights.

For the final two-day onsite, the team simply has candidates join actual projects, share meals, and work alongside the team — assessing genuine passion and enthusiasm for the problem space. Truell stresses that the quality of the first 10 employees is critical: "If you want to move fast on a year-long timescale, slowing down on a six-month timescale is extremely helpful."