Aug 18, 2025 · 1h 14m · 20vc
Lovable CEO, Anton Osika: The State of Foundation Models, Grok vs OpenAI, and Replit vs Bolt · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this in-person interview on the 20VC podcast, host Harry Stebbings sits down with Lovable Co-Founder Anton Osika to discuss the disruptive future of generative AI, the strategic decisions behind building a high-growth tech startup in Europe, and how AI-native development will fundamentally transform the software engineering landscape.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 31.7% of the talking time here. How this is scored →
Harry as informed peer 4.6 Guest teaching 3.2 Guest disagreement 2.3 Harry pushing back 3.9
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 32:50 Anton justifying publicly bashing Replit
Anton aggressively defends calling out competitor Replit publicly, explaining his refusal to accept false security vulnerability claims made against Lovable.
Hardest push from Harry ▶ 33:34 Harry directly confronting Anton on AI security flawsHarry refuses marketing platitudes around AI code quality and directly confronts the guest with 'All of you guys suck at security. Is that true?'.
Biggest teaching moment ▶ 8:22 Chicken shot out of a cannon defensibility metaphorAnton educates the host on early AI startup dynamics, reframing traditional moat arguments with a metaphor about chickens shot out of cannons needing to flap fast.
Harry holds his own ▶ 14:16 Harry on the brand bell curve and funnel efficiencyHarry demonstrates deep venture expertise by articulating how growth companies transition from brand art to funnel optimization science and back to brand.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| The AI Arms Race: Capital vs. Talent | 3 | 3 | 2 | 3 | Harry asks whether AI is a capital arms race and probes the cost of engineering talent using Zuck's NFL-style contracts as an example. Anton gently reframes the premise, clarifying that application-layer talent requirements differ fundamentally from foundational model talent. Harry presses on whether application-layer builders need lower caliber talent, which Anton nuance-checks. | |
| Hiring for Slope and Personal Growth as a Founder | 5 | 2 | 2 | 5 | Harry challenges Anton's desire for operational management structure by citing Jensen Huang's 52 direct reports and the popularization of 'founder mode'. Anton acknowledges the perspective but defends needing a protective organizational layer while retaining founder-mode impact. | |
| Defining Brand and the Myth of AI Defensibility | 4 | 4 | 1 | 3 | Harry raises standard VC critiques regarding the lack of defensibility in AI wrapper applications. Anton educates Harry on startup momentum using a vivid metaphor comparing early AI startups to chickens shot out of cannons who must flap as fast as possible. | |
| AI Unit Economics, Margins, and Model Routing | 5 | 3 | 2 | 5 | Harry directly probes Lovable's unit economics, asking bluntly how much of every dollar passes directly to model providers like Anthropic and OpenAI. Anton admits the majority passes through, explaining why optimizing model routing is secondary to rapid feature iteration at this stage. | |
| Financial Patience and Delaying Margin Optimization | 6 | 3 | 1 | 3 | Harry demonstrates strong knowledge of growth funnels, unit economics, and payback periods, citing conversations about Revolut's growth strategy. Anton shares opposing perspectives on prioritizing immediate mindshare over unit economics optimization. | |
| Rethinking Software Architecture for the AI Era | 5 | 4 | 2 | 3 | Harry presses on competitive threats from model providers like OpenAI and Anthropic launching native coding tools. Anton explains Lovable's model selection architecture, describing how they route complex debugging tasks to GPT-o3-mini/GPT-5 while using Anthropic for code generation. | |
| Lovable's Explosive Growth and Incomprehensible TAM | 5 | 3 | 3 | 5 | Harry presses Anton for a granular breakdown of Lovable's $100M ARR and challenges whether targeting single-seat AI founders is optimal compared to consumer hobbyists. Anton holds his ground, arguing that building for AI-native founders naturally trickles down to consumer use cases. | |
| The Death of Traditional Design: Figma vs. AI Prototyping | 5 | 4 | 3 | 4 | Harry brings up Figma Make as an upstream competitor that could capture prototyping workflows early. Anton dismisses the traditional pixel-perfect design phase, arguing that Figma's approach slows teams down compared to high-level AI prompt iteration. | |
| Replit vs. Lovable: Security and the Self-Driving Analogy | 5 | 4 | 5 | 6 | Harry brings up Jason Lemkin's security breach on Replit and directly challenges Anton: 'All of you guys suck at security. Is that true?' Anton defends Lovable's security measures, using a self-driving car analogy to argue AI code generation is already safer than average human developers. | |
| The Future of Engineering and the Fall of Computer Science Degrees | 5 | 4 | 3 | 4 | Harry forces Anton to choose whether AI accelerates 1x engineers into 10x or 10x into 100x, while offering a sharp critique of UK university culture. Anton explains how AI bridges skill gaps for generalists but expands leverage for senior 10x engineers. | |
| Incumbent Disruption, Change Management, and Winning Culture | 6 | 2 | 3 | 5 | Harry advocates for an aggressive work culture, citing Cognition's mandatory 6-day work week policy and Revolut's focus on winning. Anton rejects rigid hour mandates, emphasizing that he evaluates employees solely on 10x output and impact. | |
| The Case for Building in Europe | 5 | 4 | 3 | 5 | Harry pushes back on European tech culture clichés and points out the lack of experienced scaling operators in Europe. Anton acknowledges the ecosystem limitations but argues Stockholm offers unique talent magnet and team efficiency advantages. | |
| Lessons in Focus: The Mistake of Split Attention | 4 | 3 | 3 | 4 | Anton reflects on split focus with the GPT Engineer open-source project. When Harry asks if open source was essential for feedback, Anton firmly disagrees, stating that extreme focus on a single bottleneck is key to speed. | |
| Hiring Challenges and Co-Founder Dynamics | 3 | 2 | 1 | 3 | Harry observes co-founder Fabian's prominence in a recent video release and asks about founder visibility dynamics. Anton speaks candidly about their introverted versus extroverted personality trade-offs. | |
| Humility, Success, and Personal Life | 2 | 1 | 1 | 2 | Harry asks personal questions regarding how rapid financial success impacts marriage and lifestyle. Anton reflects humbly on Swedish culture and maintaining an unchanged lifestyle. | |
| Goodhart's Law and the Fallacy of Benchmarks | 5 | 4 | 1 | 3 | Harry references Surge AI's founder calling model benchmarks bullshit. Anton agrees, explaining Goodhart's Law and illustrating how product metrics become gamified once targeted directly. | |
| Quick Fire: Grok, OpenAI, and the Rise of Chinese Models | 5 | 4 | 4 | 4 | Harry sets up a valuation quick-fire scenario between OpenAI, Anthropic, and Grok. Anton takes a contrarian stance by picking Grok to outperform and shorting OpenAI, citing team morale and Grok's AI tutoring data strategy. | |
| Quick Fire: Isaac Newton and Under-the-Radar Tech | 4 | 3 | 2 | 4 | Anton selects Isaac Newton for dinner and highlights AI browser startups as under-the-radar tech. Harry asks if Anton would invest in Perplexity at $18 billion, eliciting a telling laugh. |
Statements from this episode (31)
Osika: Invest in xAI's Grok and short OpenAI
“I'd invest in GroK, and I would probably short Anthropic. No, I would short OpenAI.”
Osika: 50% chance the best AI model comes from China
“I do think there's like a fifty-fifty chance they will have the best model. We'll be using a Chinese model at some point, because...”
Osika: Zuckerberg pays top AI talent for domain knowledge, not general skill
“For Zuck it's like, There's these 10 people that know everything about how to train the foundation models, and he's more paying for that knowledge than for, like, these people, the talent itself is so good.”
Osika: Meta's top AI hires would underperform at the application layer
“I think, like, I, one of those people, Saki's hiring, they wouldn't perform as well as engineers in my team doing what we're doing. So it's very different type of talent.”
Stebbings: 20VC hires candidates with extreme trauma or extreme masochism
“For us, I look for people who have either extreme trauma or extreme masochism.”
Osika: Lovable will expand into general business administration and finance
“Lovable today is your technical co-founder. We want it to be your co-founder in general that handles all the admin, setting up your finance operations”
Osika: Early-stage AI founders should prioritize growth over defensibility
“I think that's my recommendation to just be like, execute fast, grow faster and then when you're starting to get up there, you can start, maybe start thinking a bit about the defensibility.”
Osika: Majority of Lovable's paid usage revenue goes to AI compute
“I don't give you the exact numbers, but if you look at the paid usage, it's majority, it's not everything.”
Osika: Lovable applications generate over $10M ARR for AI model providers
“We looked at, like, the number, this was a few months ago, but we looked at how much Revenue is flowing through the AI from lovable applications. Okay. And it was more than ten million dollars in ARR.”
Osika: OpenAI will be a bigger competitor to Lovable than Anthropic
“OpenAI is doing that better than Anthropic. So I see them as a more serious competitor in, in, in 12 months.”
Osika: OpenAI's GPT-o1 falls short by optimizing everything into one model
“The biggest part of GPT-V that's disappointing is that now they have to optimize all these different things into one model. Before it was like different models, and they had to do it really fast. So it's inevitably going to fall short in some dimensions.”
Osika: Lovable uses Anthropic for coding but OpenAI for complex debugging
“And then we use for code writing, we usually use Anthropic. And right now you can say like, I want to use GPT-V and that's better when you're solving a really hard debugging problem.”
Osika: Lovable plans to spend $100M hiring top model-training talent
“And we have to do it both with like how we build this agentic chain and over time in building absolutely world-class paying a hundred million dollars for getting the people that train the models. So, so that's on the horizon for us to get it to like be hyper-p…”
Osika: 80% of Lovable's revenue comes from complex app builders
“And 80% of people are in the first category. They're building real complex applications.”
Osika: Enterprise AI strategy should target product velocity, not dev productivity
“If I was a CEO or like a CTO of a large enterprise company, I wouldn't think in terms of, oh, how can we make our engineers more productive? I would think in terms of how can we get the most information about what we should build as quickly as possible into ne…”
Osika: AI will replace traditional, detailed manual software design workflows
“The way of doing it right now with the design, what like one person does all the design very slowly, very detailed is going to be replaced by AI doing, you talk much more high level, and you talk about your design philosophy, and then the AI does the implement…”
Osika: Hyper-personalization will eliminate detailed AI prompting within five years
“Yes, I think so, but maybe it will evolve in terms of how you do it. It will, like, hyper-personalization takes care of a lot of detailed prompting. That we have to do today.”
Osika: AI makes generalist skills far more important than deep specialized expertise
“Like, I think being a generalist becomes more and more important with everything, in everything with AI, so that you can understand how things come together as a larger whole, and then you use AI for the, like, deep expertise that you don't need in the future …”
Osika: People aiming to maximize current earnings shouldn't go to university
“If you want to have a job that, where you make the most money now, you should, they shouldn't go to university.”
Osika: Prioritize intense work over balance for short two-year stretches
“I think over a 10 year period, I would advocate for some balance, but over a two year period if you really care about something, then you should make sure that you have, like, you get your exercise, you sleep in really, really well, and maybe something that, y…”
Osika: European tech culture fosters superior teamwork and capital efficiency
“A culture of like humility and low ego and working really, really well together as a team that I think is stronger in Europe and like this way of thinking in terms of efficiency and doing much, much more with less that's also stronger here.”
Osika: Lovable will act as a full-stack AI co-founder by 2026
“I mean, it's your perfect co-founder that you go to with your idea from the idea stage, but also all the way up to growing your business once you have customers and taking care of like what Elena is doing, optimizing the product for growth, optimizing the prod…”
Osika: AI model benchmarks decay over time due to Goodhart's Law
“I mean, they turn more and more bullshit over time. There's something called good hearts law. So when you start optimizing for a number, that number stops being a good measure for success.”
Osika: AI is already smarter than humans when given proper context
“I think AI is smarter than humans, and most people don't agree. I think most people don't agree, and the reason is that it's oftentimes it's very, very stupid, but that's very stupid, but if you give it all the context or you have, like, you build a purposeful…”
Osika: General AI progress will plateau while bioengineering advances exponentially
“Yeah, we're going to see a plateauing. There's some sigmoid curves where I think we're still in this, like, exponential phase of the sigmoid curve, so it and those could be something like science and engineering, and like bioengineering, where AI is just going…”
Osika: Grok has better team morale and trajectory than OpenAI and Anthropic
“I think it's more the slope on the Grok team. They have, they're doing something which I respect a lot, which is to hire missionaries for the data curation part, and they call it AI tutoring. And they I think the morale is much, much better in that team than b…”
Osika: The best AI models will always remain closed-source
“I think the best ones will always be closed but if you want maximal flexibility and some kind of open ecosystem around it might be that open ones are the ones that most people choose.”
Osika: Perplexity building its own phone is a good bet
“So they're, they want to create their phone, I think, and I think that's a good bet.”
Osika: AI startups must build broad UX flywheels before launching autonomous agents
“In the context of lovable I thought we should be building an agent Before, like, the models were ready for it, and because the models were starting to get optimized for an agentic system, and what I realized is that, no, no, no, you need to have a product that…”
Osika: AI will devalue knowledge work just as fine art lost profitability
“I think we're going to have, like, Maybe a shift away from some very glamorous jobs, which people will be get depressed by, like similarly to how being like an artist was like so fucking cool, but clearly you can't make any money as an artist. I think we're go…”
Shorts cut from this episode
▶ Analysing Chat GPT-5 · 20VC with Harry Stebbings (@16:57)
▶ Could China Win the AI Race? · 20VC with Harry Stebbings (@0:24)
▶ How AI Upscales Engineers · 20VC with Harry Stebbings (@37:44)
▶ The Future of Design · 20VC with Harry Stebbings (@28:24)
▶ Open AI vs. Anthropic · 20VC with Harry Stebbings (@16:35)
▶ Succeeding as an AI Startup 📈 · 20VC with Harry Stebbings (@8:29)
▶ “I Would Short OpenAI” · 20VC with Harry Stebbings (@0:00)