The Exponential Puck Framework
by Benjamin Mann • Co-founder at Anthropic
Former architect of GPT-3 at OpenAI; currently serves as tech lead for product engineering at Anthropic, focusing on aligning AI to be helpful, harmless, and honest.
🎙️ Episode Context
Benjamin Mann discusses the trajectory of AI development, predicting a 50% chance of superintelligence by 2028. He details Anthropic's departure from OpenAI due to safety concerns, the mechanics of Constitutional AI (RLAIF), and how to build products for exponential technologies. He also introduces the 'Economic Turing Test' as a metric for AGI and offers advice on future-proofing careers.
Problem It Solves
Prevents teams from building products that are obsolete by launch time or failing to leverage upcoming intelligence leaps.
Framework Overview
A strategic approach to product development in rapidly accelerating fields. Instead of building for current model capabilities, teams should analyze exponential scaling laws to predict future capabilities (6-12 months out) and build products that rely on those future states, even if current success rates are low.
⚡ Step-by-Step Framework
Internalize the exponential curve; progress looks flat until the knee of the curve.
Don't build for today's model limitations.
Identify capabilities working 20% of the time today; assume they will work 100% of the time soon.
Design the product experience for the future capability state.
Internalize the exponential curve; progress looks flat until the knee of the curve.
Don't build for today's model limitations.
Identify capabilities working 20% of the time today; assume they will work 100% of the time soon.
Design the product experience for the future capability state.
When to Use
When developing software on top of rapidly improving foundational technologies (like LLMs) where capabilities double frequently.
Common Mistakes
treating new tools like old tools (e.g., just using autocomplete vs. ambitious coding agents); giving up when a prompt fails once.
Real World Example
Anthropic's 'Claude Code' was built assuming users would move away from IDEs to terminals, even before the model was fully perfect at it.
Don't build for today, build for six months from now... things that aren't quite working that are working 20% of the time, will start working 100% of the time.
— Benjamin Mann