Should AI product builders build scaffolding around current model limitations, or design for the capabilities models will soon have?

Sam Altman · Apr 15, 2024

Roughly 95% of AI startups should build assuming the underlying models will keep improving at the current rate, rather than assuming the models are frozen and patching around their limitations

The models are on a continued improvement trajectory, so gap-filling products get absorbed as the base models advance

Scope: framed as two available strategies, with the 'assume no improvement' bet being the wrong one

0:00 20VC: OpenAI's Sam Altman and Brad Lightcap on The Future of Foundation Models: Will They Be Commoditised | How to Solve the Problem of Compute | Open vs Closed: Which Dominates and Why | Which Companies and Verticals Will Be Steamrolled by OpenAI

Sam Altman · Apr 15, 2024

Roughly 95% of AI startups should be built assuming models keep improving at the current rate, but most have instead been built assuming the model won't get better.

Startups that patch around current model limitations get steamrolled when OpenAI does its fundamental job of improving models and tooling, while startups that benefit from a much better GPT-5 will thrive.

Scope: assumes AI progress continues as OpenAI expects

21:16 20VC: OpenAI's Sam Altman and Brad Lightcap on The Future of Foundation Models: Will They Be Commoditised | How to Solve the Problem of Compute | Open vs Closed: Which Dominates and Why | Which Companies and Verticals Will Be Steamrolled by OpenAI

Amjad Masad · Apr 25, 2026

Founders building on AI must continually build scaffolding to plug model gaps, then delete it as models improve, and upgrade their vision accordingly to stay ahead

Replit's own cycle: heavy guardrail infrastructure in Sept 2024, deleted much of it for agent v2 in March 2025, rebuilt for agent 3 autonomy, now autonomy is native to Opus 4.6

7:13 20Product: Replit CEO on Why Coding Models Are Plateauing | Why the SaaS Apocalypse is Justified: Will Incumbents Be Replaced? | Why IDEs Are Dead and Do PMs Survive the Next 3-5 Years with Amjad Masad

Harry Stebbings · Jul 13, 2026

Sam Altman has said that you have to build for the model of tomorrow or next year, not today's model

12:48 20VC: Wix's Founder on What Wall St Gets Wrong About AI and Wix | Will Base44 Win the Vibe Coding Wars | The Truth About the Economics of Vibe-Coding | The Buyback Disaster: Lessons Learned with Avishai Abrahami