I love using AI. I use it daily. I listen to podcasts about it. I write blog posts about why it might even be good for something. But I will also acknowledge that it’s not always a joy-inducing experience. So for this quick Sunday night post, let me briefly chat about where it’s been falling short for me recently. Let’s treat this as an amuse bouche rather than a fully considered prix fixe.
In addition to loving AI, I love coffee. In fact, I spent the first four years of my career working with some of the best coffee roasters in the country and selling coffee online. And while I’ve loved drinking coffee for my entire adult life, I’ve only recently fallen in love with making espresso at home. Home espresso is intimidating. While it’s not as complicated as some people (me) make it out to be, there are enough variables between the coffee, grinder, water, machine, and more that it can feel like a multivariate problem with far too few equations. Perfect for some all-knowing AGI assistance!
But here’s the thing — I like drinking coffee and obsessing about coffee and sharing coffee that I love with others. Coffee is a hobby. And while I do find joy in being a forum jockey or Reddit lurker reveling in the knowledge that I have, I think most hobbies do benefit from real-life experience. Coffee certainly does.
So there’s the failure mode. To give you the clarity I can only hope to achieve with my V60, here’s what I’ve found in a few dedicated Codex chats (still part of my codebrain repo — as mentioned earlier, I actually prefer Codex to Claude Code for my personal work these days):
-
Sol is really good at finding the same info that I can. And that’s actually a compliment (I might be talking myself up here). As mentioned earlier, I care deeply about coffee. I also tend to be a little obsessive about my hobbies. And I also am pretty good at using search. Turns out I’m basically doing the manual version of what a research agent does: searching the same internet, reading the same sources, and synthesizing what’s there. But does that qualifier even matter? It’s a hobby. I want to spend my time reading about conical vs flat burrs. It’s actually a bummer for me when I’ve exhausted the high-quality reading materials. Using an agent gets me to the same place, but a lot faster. That’s a bug, not a feature for this hobby sometimes.
-
And about getting to the same place — it’s a little boring. If I get there, and an agent gets there, and everyone else obsessing on Reddit gets there…does it still feel as special? I hate that this line of thinking feels a bit snobbish, but it’s kinda how I feel sometimes. It’s fun to feel like you’re part of a special club. Climbing Everest seems less special when you’re stuck in a traffic jam near the summit.
-
And finally, unknown unknowns are a thing, especially when dealing with experiential things. We talked about how important it is to understand where you’re coming from when scoping a problem. I think of this in a similar way. With coffee, I’m very lucky to have already put in a bunch of time trying different roasters, origins, and brewing methods. I’m also lucky to have been around far more knowledgeable coffee people. When working with Codex, I can’t help but worry about my unknown unknowns (which yes, is maybe a bit antithetical to not knowing what you don’t know, but roll with me). I think I know how to describe what I like. But I’ve also consumed so much coffee content that I’ll admit to parroting back jargon without even thinking about it. Give me an origin and roast level, and I’ll confidently tell you what the tasting notes are. Do I know that specific natural-process Ethiopian Guji tastes like blueberry? Nope. But that’s what I’m used to reading — and maybe what I’ve tasted a few times too. So unfortunately, the media I’ve consumed does influence how I tell Codex what flavors I’m looking for from my coffees. In other words, I probably have blind spots that I don’t even know about, and my trusty codebrain thought partner probably does too.
But what’s the problem?! You’re talking the same talk that Sol understands, right? Well — I’ve also cupped coffee with people who really know what they’re doing. And let me tell you — I’ll take anything they say over Claude or ChatGPT any day. Because they can actually smell my bullshit and correct me. And also share what they’ve spent their entire careers learning the hard way (through actually roasting, tasting, and developing relationships with farmers). Guess what, they’re also rarely writing about this online. They’re a little too busy actually doing the things. Unfortunately, this creates a bit of a blind spot for LLMs. They’re really good at finding the consensus opinion, but sometimes it’s the outlier that I’m chasing.
# Average isn’t enough.
> My agent can find everything ever written about espresso. Turns out that’s not always what I want from a hobby.
- written by: Andrew Schroeder
- edited by: GPT-5.6 Sol, Fable 5
- status: published · 2026-08-23
- canonical: https://www.moltolabs.ai/notes/average-isnt-enough/
- raw markdown: https://www.moltolabs.ai/notes/average-isnt-enough.md
---
I love using AI. I use it daily. I listen to podcasts about it. I write blog posts about [why it might even be good for something](/notes/good-for-nothing/). But I will also acknowledge that it’s not always a joy-inducing experience. So for this quick Sunday night post, let me briefly chat about where it’s been falling short for me recently. Let’s treat this as an amuse bouche rather than a fully considered prix fixe.
In addition to loving AI, I love coffee. In fact, I spent the first four years of my career working with some of the best coffee roasters in the country and selling coffee online. And while I’ve loved drinking coffee for my entire adult life, I’ve only recently fallen in love with making espresso at home. Home espresso is intimidating. While it’s not as complicated as some people (me) make it out to be, there are enough variables between the coffee, grinder, water, machine, and more that it can feel like a multivariate problem with far too few equations. Perfect for some all-knowing AGI assistance!
But here’s the thing — I like drinking coffee and obsessing about coffee and sharing coffee that I love with others. Coffee is a hobby. And while I do find joy in being a forum jockey or Reddit lurker reveling in the knowledge that I have, I think most hobbies do benefit from real-life experience. Coffee certainly does.
So there’s the failure mode. To give you the clarity I can only hope to achieve with my V60, here’s what I’ve found in a few dedicated Codex chats (still part of my codebrain repo — as mentioned earlier, I actually prefer Codex to Claude Code for my personal work these days):
1. Sol is really good at finding the same info that I can. And that’s actually a compliment (I might be talking myself up here). As mentioned earlier, I care deeply about coffee. I also tend to be a little obsessive about my hobbies. And I also am pretty good at using search. Turns out I’m basically doing the manual version of what a research agent does: searching the same internet, reading the same sources, and synthesizing what’s there. But does that qualifier even matter? It’s a hobby. I want to spend my time reading about conical vs flat burrs. It’s actually a bummer for me when I’ve exhausted the high-quality reading materials. Using an agent gets me to the same place, but a lot faster. That’s a bug, not a feature for this hobby sometimes.
2. And about getting to the same place — it’s a little boring. If I get there, and an agent gets there, and everyone else obsessing on Reddit gets there…does it still feel as special? I hate that this line of thinking feels a bit snobbish, but it’s kinda how I feel sometimes. It’s fun to feel like you’re part of a special club. Climbing Everest seems less special when you’re [stuck in a traffic jam near the summit](https://www.theguardian.com/world/2020/jun/06/everyone-is-in-that-fine-line-between-death-and-life-inside-everests-deadliest-queue).
3. And finally, unknown unknowns are a thing, especially when dealing with experiential things. We talked about [how important it is to understand where you’re coming from when scoping a problem](/notes/how-are-we-really-working/). I think of this in a similar way. With coffee, I’m very lucky to have already put in a bunch of time trying different roasters, origins, and brewing methods. I’m also lucky to have been around far more knowledgeable coffee people. When working with Codex, I can’t help but worry about my unknown unknowns (which yes, is maybe a bit antithetical to not knowing what you don’t know, but roll with me). I think I know how to describe what I like. But I’ve also consumed so much coffee content that I’ll admit to parroting back jargon without even thinking about it. Give me an origin and roast level, and I’ll confidently tell you what the tasting notes are. Do I know that specific natural-process Ethiopian Guji tastes like blueberry? Nope. But that’s what I’m used to reading — and maybe what I’ve tasted a few times too. So unfortunately, the media I’ve consumed does influence how I tell Codex what flavors I’m looking for from my coffees. In other words, I probably have blind spots that I don’t even know about, and my trusty codebrain thought partner probably does too.
But what’s the problem?! You’re talking the same talk that Sol understands, right? Well — I’ve also cupped coffee with people who really know what they’re doing. And let me tell you — I’ll take anything they say over Claude or ChatGPT any day. Because they can actually smell my bullshit and correct me. And also share what they’ve spent their entire careers learning the hard way (through actually roasting, tasting, and developing relationships with farmers). Guess what, they’re also rarely writing about this online. They’re a little too busy actually doing the things. Unfortunately, this creates a bit of a blind spot for LLMs. They’re really good at finding the consensus opinion, but sometimes it’s the outlier that I’m chasing.