Stuck-Point Scaling Law
Reliably improve an AI system by hunting where it gets stuck and tuning those spots with a fast feedback loop.
- Difficulty
- Advanced
- Time to result
- ~ongoing to results
- Steps
- 3
- Confidence
- 80%
A method for making an AI product reliably better as you invest more work: instead of chasing raw model capability, painstakingly identify the specific places where the system gets stuck, then quantitatively tune the whole system to fix those spots with a very fast feedback loop. Use it when building on top of foundation models and quality plateaus.
Origin
Anton described this as one of several 'scaling laws' in building AI systems — the one where more work reliably yields a better product. It was Lovable's technical unlock: AI-built products tend to start strong then get stuck, so they systematically attacked the stuck points.
Core principles
- 01More work reliably yields a better product when directed at stuck points
- 02Target the most important and most frequent failure areas first
- 03Tune the entire system quantitatively, not by vibes
- 04A very fast feedback loop is what makes the tuning compound
How to run it
- 1
Find where it gets stuck
Painstakingly identify the specific places where the AI stalls — introduces a bug it can't escape, or fails a common workflow.
Pro tip Prioritise the highest-value flows: adding login, data persistence, and Stripe payments were the ones Lovable made sure never got stuck.
- 2
Tune the whole system quantitatively
Address the stuck points by tuning the entire system with measurable, quantitative signals rather than one-off patches.
Watch out Making the model 'smarter' in general is not the lever — targeted, measured tuning of real failure areas is.
- 3
Close a fast feedback loop
Build tooling so improvements to the stuck areas can be measured and iterated on quickly, letting gains compound.
Pro tip Treat the frontier of 'where it still gets stuck' as receding — re-hunt periodically as it moves.
Watch out It will still get stuck sometimes; the goal is reliability on the most important areas, not perfection.
In the wild
Lovable specifically identified adding login, creating data persistence, and adding Stripe payments as high-value places the AI must not get stuck, then tuned the system so those flows became reliable.
→ Reliability on core workflows became Lovable's reputation — repeatedly ranked as the tool that 'works most reliably'.
Common mistakes
Chasing general model intelligence
Betting on the model simply getting smarter ignores the cheaper, more reliable win of tuning specific stuck points with data.
Is it for you?
Best for
Founders and engineers building applied products on top of foundation models
Not ideal for
Teams without the instrumentation to measure failures quantitatively or iterate quickly
From the transcript
“painstakingly identify places where it go stuck”
“having a very fast feedback loop to improve it in the areas where it got stuck”
From the episode
Building Lovable: $10M ARR in 60 days with 15 people
Anton Osika (co-founder and CEO)