Field Notes

Latest · July 12, 2026

Blog Post -1

This week, Paul Graham posed a thought-provoking question that has sparked extensive debate online: "Imagine what it will be like if 5 years from now models have improved as much as today's models improved on GPT-3." Consider the implications of this. GPT-3, released in 2020, could generate a decent paragraph. Today's leading models can plan projects, write production code, and manage multi-step tasks for hours without supervision. This represents the significant leap we've already experienced. Now, envision that same leap occurring again. The responses have divided into two perspectives: → The skeptics argue that "S-curves flatten. Scaling is hitting walls. This is peak hype." → The believers counter with, "It won't even take 5 years — AI is now improving AI. Compound, not linear." As someone who works with these tools daily, my perspective is straightforward: I don't need to determine who is correct. Even the most conservative scenario - where today's frontier capabilities are available at a fraction of the cost - fundamentally alters the economics of every service business, agency, and development team. Thus, at my company, we don't ask, "Will AI reach that level of capability?" Instead, we ask, "Is anything we're building designed to withstand the curve?" Here are three changes I've implemented as a result: 1. We no longer build around the limitations of current models; workarounds have a short shelf life. 2. We focus on the elements that remain constant - client trust, domain knowledge, and distribution. 3. We view every workflow as temporary; the process may be disposable, but the outcome is not. Betting against the curve has proven to be a losing strategy for six years. Where do you stand - S-curve flattening or compound takeoff?