The Twenty Minute VCLovable CEO, Anton Osika: The State of Foundation Models, Grok vs OpenAI, and Replit vs Bolt
CHAPTERS
- 0:00 – 0:37
Cold open: Universities, model bets, and a China warning
A rapid-fire cold open sets the tone: Anton questions the value of university, makes provocative investing calls (long Grok, short OpenAI), and predicts a future leading model could come from China. The segment frames the conversation around talent, morale, and geopolitical model competition.
- •University isn’t the best place to learn, regardless of subject
- •Anton would invest in Grok; considers shorting OpenAI
- •Grok’s strength: team “slope,” morale, and data curation/AI tutoring approach
- •Prediction: an as-yet-unbuilt leading model will emerge from China
- •Concern: 50/50 chance the world uses a Chinese model someday
- 0:37 – 1:39
AI is a talent-and-trust arms race (not just capital)
After introductions, Harry asks whether AI is a capital arms race. Anton argues the real race is assembling the best team and earning user trust through brand—capital matters most for foundation-model training compute, less for application-layer players like Lovable.
- •Arms race centers on team quality and user trust/brand
- •Capital is helpful but not Lovable’s primary constraint
- •Foundation model labs face compute-heavy training constraints
- •Winning requires extreme speed of execution
- •Recruiting and iteration cadence are strategic advantages
- 1:39 – 3:10
Competing with Meta’s pay packages: different talent for different layers
The conversation turns to Zuckerberg-level compensation and what “top talent” means across the stack. Anton contrasts foundation-model experts (rare, knowledge-specific) with application-layer builders who excel at product iteration, culture, and collaboration.
- •Meta pays for scarce foundation-model training know-how
- •Application-layer engineering talent is different, not inferior
- •Hard part: identifying who will thrive in Lovable’s culture
- •Compensation follows confidence in fit; uncertainty makes it harder
- •Team effectiveness depends on moldability and cohesion
- 3:10 – 6:44
Hiring for 'slope' and building a protective layer around the founder
Anton explains a non-obvious hiring heuristic: “slope” (learning rate/adaptability) revealed through dynamic conversation and evidence of past execution. He also describes the ‘protective layer’ of close generalists that filters inputs and helps prioritize while he stays in founder mode.
- •‘Slope’ as a primary hiring signal (adaptability + learning velocity)
- •Desire to see ‘how they worked’ in past roles for real signals
- •Anton balancing scrappy founder behavior with growth-stage needs
- •Protective layer: founder-type generalists filtering and prioritizing
- •Founder mode remains core, with organizational support for order
- 6:44 – 7:19
Brand as trust: Apple-like detail obsession while shipping fast
Anton defines brand as the accumulation of trust built through consistent, high-quality user interactions. Lovable aims to move quickly while still understanding user reactions to rapid product changes, using brand to reduce friction and increase retention.
- •Brand = trust earned through detail and consistency
- •Apple cited as a model for trust-building via quality
- •Fast shipping must be paired with careful rollout/feedback loops
- •Every product update is a brand interaction
- •Goal: high trust despite rapid iteration
- 7:19 – 9:17
Defensibility in AI startups: platforms, switching costs, and 'cannon chickens'
Harry presses on defensibility; Anton argues it comes from becoming a value-accumulating platform users don’t want to leave. He likens AI startups to “chickens shot out of a cannon”—traction requires relentless execution first, then defensibility can compound.
- •Defensibility comes from user value stored/compounding on-platform
- •Lovable positioning: from technical cofounder to broader business cofounder
- •Future features: admin/ops support (finance ops, etc.) to deepen lock-in
- •Early-stage priority: execute faster than the constant influx of competitors
- •Defensibility can be layered once altitude/traction is achieved
- 9:17 – 15:41
Unit economics reality: pass-through costs, routing, and margin patience
The discussion moves to whether AI app-layer companies can be good businesses when model costs are a major pass-through. Anton explains why Lovable prioritizes speed and UX now, expects margins to improve as users stay subscribed, and explores simplifying model-provider connections and potential token-margin opportunities.
- •Paid usage has majority of revenue flowing to model compute today
- •Long-term goal: subscription retained for platform value, not just ‘pay to build’
- •Too early to optimize model choice aggressively; models change monthly
- •Future: mix of cheap ‘obvious’ calls vs expensive ‘new situation’ thinking
- •Simplifying model-provider setup; possible margin via cost reduction/markup
- 15:41 – 19:22
Lovable vs the labs: competing with OpenAI/Anthropic and evaluating GPT‑5
Harry asks whether model providers will move up the stack. Anton sees OpenAI as the more serious near-term UX competitor, explains Lovable’s multi-model agentic chain (Anthropic often for code; GPT‑5 for hard debugging), and critiques GPT‑5 as an ‘ambitious’ consolidation with trade-offs rather than a step-function leap.
- •Labs will encroach; long-term advantage is team execution + UX
- •OpenAI currently stronger UX ‘gateway’ than Anthropic
- •Agentic chain: multiple models; Anthropic for code, GPT‑5 for tough debugging
- •GPT‑5 evaluation: latency, quantitative evals, and “vibe checks”
- •GPT‑5’s trade-off: unifying many behaviors into one model limits improvements
- 19:22 – 20:05
From prompts to personalization: the next step-function for AI builders
Anton describes what would unlock major capability gains: AI systems that understand the user and situation deeply, reducing the need for hyper-detailed prompting. He argues prompting remains, but personalization will absorb much of today’s manual context-setting—like great employees who already ‘know how you think.’
- •Step-function change: richer context about the user and goals
- •Hyper-personalization requires both product architecture and top talent
- •Prompting as context provision will persist but evolve
- •Personalization reduces brittle, detailed prompting
- •Analogy: high-context employees vs low-context chatbots
- 20:05 – 26:34
How Lovable reached ~$100M ARR: who pays and why (founders first)
Anton breaks down Lovable’s revenue mix: most revenue comes from people building real, complex applications, with fast growth in enterprise prototyping use cases and a smaller long-tail of websites and personal projects. He argues starting with AI-native founders fits the mission—unlocking business creation for those without coding or capital—while enterprise adoption follows.
- •~80% of revenue from builders of complex applications
- •Growing segment: enterprise teams prototyping to communicate product ideas
- •Consumers/hobbyists use it for websites; growth but not the main driver
- •Mission-led sequencing: start with business builders, then broaden
- •Enterprise angle: speed of learning what to build, not just dev productivity
- 26:34 – 32:23
Democratizing product creation: design-first vs build-first (Figma’s threat)
Harry probes whether teams will skip traditional design/brainstorming and whether Figma’s position in design gives it an advantage. Anton argues AI will replace pixel-perfect, slow design workflows for most cases with higher-level design intent plus AI implementation, while acknowledging a trade-off between flexibility and velocity and a future of more opinionated building paths.
- •Product lifecycle compression: idea → validated product in minutes/hours
- •AI-driven design: humans set philosophy; AI executes; humans gather feedback
- •Figma remains for pixel-perfect needs, but may slow most builders down
- •Lovable’s balance: flexibility today vs faster velocity if more opinionated
- •Future: more opinionated workflows as AI UX stabilizes
- 32:23 – 44:13
Security and public spats: why the whole category is vulnerable
A competitive incident triggers a deeper security discussion. Anton claims security is a daily focus and argues Lovable can become safer than an average solo developer by enforcing reviews and automated checks—though it’s not perfect yet—drawing parallels to self-driving safety relative to average human drivers.
- •Anton’s critique of irresponsible vulnerability ‘announcements’ by competitors
- •Category-wide security shortcomings acknowledged as real
- •Lovable aims to enforce security reviews and automated scanning
- •Claim: Lovable can outperform the ‘average’ dev building alone (not yet perfect)
- •Goal: drive vulnerability probability toward 0%
- 44:13 – 49:49
10x culture and team design: impact over hours, keepers test, quality vs speed
The conversation shifts to performance culture: intense effort over short periods, sustainable balance over decades, and measuring impact rather than raw hours. Anton discusses the keepers test, the challenge of maintaining quality as the company matures, and the ‘cowboy vs farmer’ tension between innovation and long-term optimization.
- •Two-year intensity vs ten-year sustainability framing
- •Impact-centric culture: 10x output via talent, focus, and sometimes hours
- •Uses the keepers test; avoids excessive reshuffling to protect culture
- •Maturing org needs quality discipline (‘move slow to move fast’)
- •Innovation sprints are fine, but brand requires reliability and polish
- 49:49 – 54:20
Building in Europe: hard mode networks, easier talent magnet, and public storytelling
Anton explains Europe’s trade-offs: weaker networks and fewer experienced scalers, but an advantage in becoming the top talent magnet in a smaller hub like Stockholm. He credits transparency and building in public for global distribution, and highlights cultural strengths like humility, teamwork, and efficiency.
- •Europe ‘hard mode’: thinner networks and fewer repeat scalers
- •Distribution easier from SF/NY, but Lovable broke through from Stockholm
- •Key tactic: storytelling, transparency, building in public
- •Europe advantage: easier to be the #1 talent magnet locally
- •Cultural benefits: humility, low ego, efficiency, knowledge compounding
- 54:20 – 1:02:05
Founder lessons, bottlenecks, and the road to a full lifecycle platform
Anton reflects on mistakes (splitting focus between open source GPT Engineer and Lovable) and emphasizes bottleneck-solving as the path to speed. He outlines current bottlenecks (next-phase product/technical leaders, enterprise pull vs founder focus), a non-traditional enterprise sales approach, and a 2026 vision where Lovable ‘eats the whole stack’ across build, growth, and customer communication.
- •Hindsight mistake: doing two tangentially related tracks; need maximal focus
- •Operating principle: find the company bottleneck and solve it
- •Hard hiring: engineering leaders; past performance doesn’t always translate
- •Enterprise sales yes, but not ‘wine and dine’—more enablement and learning
- •2026 vision: one opinionated tool across the entire product lifecycle
- 1:02:05 – 1:14:08
Quick-fire on AI’s trajectory: benchmarks, plateauing, China, open vs closed, and risks
In the closing quick-fire, Anton argues benchmarks degrade under Goodhart’s law, expects plateauing in some dimensions while exponential progress continues in areas like bioengineering, and makes bold calls: invest in Grok due to team morale/slope, short OpenAI, and expect a future leading model from China. He also discusses open vs closed models, job displacement fears, and geopolitical competition risks.
- •Benchmarks become ‘more bullshit’ over time; metrics get gamed
- •AI is ‘smarter than humans’ with the right context/system design
- •Progress: plateau in broad nuance; continued rapid gains in science/bio
- •Model bets: long Grok (missionaries + data curation morale), short OpenAI
- •China: 50/50 chance of best model; open vs closed trade-offs; competitive race risks
