Anton Osika on Lovable

Guest: Anton Osika, co-founder and CEO of Lovable
Host: Lenny Rachitsky (Lenny’s Podcast)
Date: 2025
Source: raw/lenny/Anton Osika.txt

Overview

Anton Osika demos Lovable — an AI software engineer that generates a working product from a natural-language prompt. The episode covers Lovable’s extraordinary growth (4M ARR in 4 weeks, 10M ARR in 2 months, with 15 people), how the team operates, what distinguishes Lovable from competitors, and Anton’s thesis on what skills will matter most as AI tools democratise software creation.

Key ideas

“The last piece of software.” Lovable’s stated mission: build the last piece of software needed, because every future product can be created by talking to an AI. The end-state is near-instant translation from idea to fully working, deployed product.
Reliability as the core differentiator. Lovable’s technical advantage is not raw capability but not getting stuck. The team painstakingly identified where the AI fails (login, data persistence, Stripe integration) and systematically eliminated those failure modes. Users rank it as “works most reliably” among competitors.
Visual editing on top of AI generation. Unique differentiator: after the AI builds the app, users can edit text and UI elements directly, Squarespace-style, without re-prompting the agent. Changes propagate to the codebase. Also syncs with GitHub — engineers can use Cursor on the same codebase without conflict.
Generalist as the new specialist. Anton’s hiring philosophy: each team member should span design, architecture, product taste, and user research. When building product teams going forward, the bottleneck is taste and user understanding, not engineering output.
Prompting precision is the critical skill. Most common beginner failure: vague error descriptions. The right practice is stating exactly what you expected and exactly what happened. Prompting well is more consequential with AI agents than with human engineers.

The demo

Live demo: Lenny prompts “Airbnb clone” — a complete Airbnb-style UI renders in 30 seconds. Anton then connects a Supabase backend (one click), adds authentication, and demonstrates visual editing to change button text without re-prompting. The whole flow takes minutes. Hosting: Cloudflare for frontend, Supabase for backend.

Lovable cannot yet be used on arbitrary existing codebases (research preview at time of recording); it can import code built in Lovable into Cursor for engineer use.

Growth and scale

300,000 MAUs, 30,000 paying users at time of recording.
Revenue growth: 1M ARR per week from launch; 4M ARR in first four weeks; 10M ARR in two months.
Team: 15 at launch, 18 at time of recording; 12 of 18 write code at least part-time.
Growth driver: almost entirely organic word-of-mouth and building in public on social media.

Technical approach

Lovable’s “scaling law”: identify precisely where the agent gets stuck, tune the full system quantitatively, maintain fast feedback loops to improve those specific areas, then move to the next. The most critical areas addressed first: login/auth, data persistence, Stripe payments. The team builds Lovable using Lovable itself for applicable parts; Cursor for the rest.

What skills will matter

Anton’s view:

Taste and user understanding become the binding constraint once engineering output is automated.
Generalists are disproportionately valuable — the new ideal hire knows architecture, design, product thinking, and user research, not just one discipline.
Engineers should reframe themselves as translators of problems into technical constraints, not writers of code.
Top 1% heuristic: spend one full focused week building something end-to-end with AI tools, from idea to a thing someone actually uses. That alone puts you in the global top 1% of AI tool fluency.

Hiring and operating model

Hiring filter: obsession and ownership (not a passenger), raw cognitive capability, startup mindset that prioritises speed over process. Work trials of 1–7 days before hiring. Job posting deliberately Shackleton-esque to filter in the people who want extreme intensity and filter out those who don’t.

Planning cadence: weekly, FigJam board of ranked problems, weekly demo for team alignment. Short roadmap horizon (~1–3 months); stays flexible to major AI capability drops. Engineering-led product decisions because the right solutions are often entangled with technical details.

Key concepts

Vibe Coding — what Lovable enables at scale for the 99%
Agentic Engineering — Anton’s framing of engineers-as-translators
Tool Use — foundation of the Lovable architecture
Product Taste — the skill Anton identifies as the remaining bottleneck