Josh Quintal

Is @GeminiApp OK? This is Pro 3.1 Context: It sent me a link to a "Rainfall in Australia" Wikipedia article that didn't exist. When pressed to find the correct link, it fails every time. This attempt's CoT even began with "Pinpointing Wikipedia article link..." and still resulted in this.

AI can tell you that 100 customer requests are really the same 3 pain points, but it can’t tell you whether you built the wrong system. That deeper question is still human work.

AI is very good at the surface layers of complexity: events, patterns, repetition, similarity, anomalies, clusters. That is real leverage, but when AI finds the pattern, humans still have to ask what produced it and how it changes the deepest layers of our understanding.

I noticed a quirk in my coding agent's plans that I needed to address: hedging. For example, it would say something like: "Keep XYZ as inspection surfaces unless they have canonical mutation paths". Unless? Buddy the code is right there, just check. Have you run into this? Here's what I added to my AGENTS/CLAUDE.md to curtail it: **Resolve Hedges Instead of Writing Them**: When you're about to qualify a claim with a condition; "unless", "if", "assuming", "depends on whether", "presumably", "should be"; stop and check whether that condition is already answerable from the code or context in front of you. If it is, verify it and write the resolved, definitive statement. Keep the hedge only when the answer genuinely isn't available. When you do, say what you checked and what would settle it. Don't leave a bare conditional. This applies to deferred *facts* (knowable now), not to honest uncertainty about things you can't determine. A hedge that hides a one-grep-away answer is a bug.

I’ve been trying to understand why this announcement bothered me so much, because on the surface I agree with the problem. AI output often produces things that are technically adequate but generic. I want better AI tools for design. My discomfort comes from the word “taste” being treated as infrastructure. At a large scale, taste goes beyond aesthetic pattern recognition to a responsibility. Designs encode mental models. They shape behavior. They decide what feels normal, desirable, and human. So when a company says it wants to build the infrastructure layer for taste, my first question is “Who gets to encode this? Who profits from it? And who bears the cost if the system’s incentives are wrong?” This is where the investment framing rubs me the wrong way. I have no problem with investors; I’m a founder myself. But we should be honest that venture-backed infrastructure companies are not neutral cultural institutions. They are built to scale, monetize, and become defaults. That can be fine when the domain is narrow and the externalities are bounded. It becomes more fraught when the domain is subjective judgment. We all saw this happen with social media. The original product promises of connection, discovery, and self-expression became privately optimized cultural infrastructure. Their incentives reshaped attention, status, identity, and behavior in ways society understood too late. My worry is that this “taste” could follow a similar path. We start with the very reasonable goal of reducing slop. Then, slowly, the systems that help us avoid bad outputs become the systems that define what good outputs are. And because those systems are hidden inside models, tools, workflows, and platforms, we may not notice when judgment has moved from humans to infrastructure. Infrastructure we don’t control. So I am not anti-AI taste tools. I am pro tools that help humans express, explore, and refine their creative vision. I am skeptical of centralized systems that claim to operationalize taste at scale without a serious account of incentives, pluralism, transparency, and accountability.

"Where should I apply AI?" is the wrong question. The better one: "Where should I apply the humans?" That was the heart of my TEDx talk last weekend: Understanding Complexity.