#012: Your AI Frustration is My Opportunity
The capability gap is the opportunity surface.
For the past couple of weeks, something has been feeling off. I couldn’t quite put my finger on it. As always, stepping away from the screen was the right first step to identifying what it is I’m feeling. I was in the middle of a pushup when my mind finally verbalized the feeling - I am frustrated. I’m not sure why - I have been sleeping and eating well, socializing, spending time with loved ones, eating 1+ ramen/week, working on things that excite me and yet I’m feeling frustrated.
But the frustration was certainly coming from things I’m working on. The projects on their own are still exciting and fulfilling overall. But I’m in the finishing stages and I can’t seem to get over that final hump. Every week is the week I expect to wrap things up. I get frustratingly close to the finish line and yet it’s moving further away from me. The last 10% is taking me 90% of the time. I’ve gone from surfing the reef of excitement to trying to climb out of the valley of frustration. And I think this has everything to do with AI.
What I’ve realized is that my frustration is coming from the gap between my expectations for the type of output I can get out of AI and the reality of what I actually get. Especially when it comes to the last stages of product development - the small details, the polish, the craft, the product’s alignment with your creative vision. If you’re a builder or a creative who cares about these things, you know that the last mile is a marathon in itself. This is not new. But when we worked with fully deterministic tools, the last mile problem came down to hidden complexity - overly ambitious timeline, improperly scoped problem, coordination hiccups, etc. As we rely more on AI in every facet of our work, we are now steering powerful probabilistic tools based on vibes towards a deterministic output (the vision). And so this last mile problem is amplified in its severity and unpredictability.
In popular psychology, the idea that “happiness = reality - expectation” is not novel. And the popular solution is easy - just lower your expectations! This is neat advice in many facets of life. But it’s terrible advice when it comes to technology, for three key reasons:
AI capabilities are opaque. The edges are murky and the capabilities are spiky. You have to stumble in the dark to find the edges through feel and experimentation.
AI capabilities are evolving rapidly. We all know someone who “tried AI” sometime last fall through ChatGPT’s free model and correctly thought it’s dumb. Not only have the models evolved significantly, they’re now powering agentic workflows rather than basic question-answer use-cases.1
You have to be an optimist to push technology forward. The greatest entrepreneurs and innovators have all been visionaries - they took the early seeds of technological progress not for what it is but for what it could be. They projected those seeds far into the future, beyond their reasonable capabilities, to shape an ambitious vision. The gap between their vision (expectation) and reality created forces that drove said technology forward.
In summary, forming realistic (or even low) expectations when technological capabilities are opaque and rapidly evolving is basically a fool’s errand. It’s even more foolish to think one can push these capabilities forward without a vision rooted in high (often unreasonable) expectations. So if we agree that you have to operate on high expectations, we also must accept that the frustration arising from the capability gap is inherent to the process.
So I’ve accepted my “AI frustration”. What now? In my mind, at a high level, there are three paths forward here:
Path #1 - Give up on AI (for now or forever). The output cannot meet your high quality bar. Revert to the tried and true tools and processes. It’ll take longer and it’ll be harder. But after all, they have gotten us where we are and you know exactly what to expect.
Many tech professionals chose this path, especially early on. But it’s becoming harder to justify. As I mentioned above, AI is improving rapidly. Where it fumbled 3 months ago, it may ace it today. Where it fumbles today, it may ace in a month. You can sit on the sidelines and wait for it to be good enough to meet your high standards. In the end, this depends on where you want to live on the adoption curve - whether you want to be the beneficiary or a reluctant consumer. And as you can already tell, this doesn’t align with my optimism principle. Giving up is not an option.
Path #2 - Don’t be precious, release before it’s ready. Don’t let the perfect be the enemy of the good. Put your work out early and get signal before you invest too much in a wrong direction.
AI kind of broke this. With AI, the principle inverts to “Don’t let the slop be the enemy of the craft”. Getting to “good enough” is now near-instant. Getting to the proverbial perfect still takes tremendous effort, skill, taste, and judgement. This is the last 10%. When everyone can now get to good enough in no time, your differentiator is that last 10% that most are not willing or not able to climb. Slop is not an option.
Path #3 - See frustration as opportunity. All this neatly leads me to the crux of my realization - the frustration is the opportunity. This is the space where we’ll need new tools, better harnesses, better orchestration, better processes, and sound judgement. This is likely why specialized direct manipulation tools, like Figma / Adobe Suite, will never go away (although their position in the value chain will change, as I argued).
But like anything in tech, this proverbial 10% will not stay static. At the pace we’re moving, AI will keep eating more and more of the process, turning arduous into effortless. But it will never eat all of it. And what remains of it is where the differentiation will shift to, where the opportunity will arise, and where the value will accrue. And the frustration inherent in the creative process is your best signal that you’re operating in this opportunity zone.
TL;DR:
Working at the capability frontier is inherently frustrating. Accept it as part of the process.
“Don’t let the perfect be the enemy of the good.” is now “Don’t let the slop be the enemy of the craft.” Slop is not an option. The last 10% is the differentiator.
Frustration is a signal. The emotional gap is the opportunity surface.
3 Bits
In the past week, there have been multiple animated AI shorts circulating around X that are getting into the dangerously promising territory. Here’s another one. I have an ongoing bet with a friend that there will be an AI short category at the Oscars in the next 5 years. So far, I haven’t lost hope.
2. The Unreasonable Effectiveness of HTML
Thariq is one of the more prominent Anthropic engineers working on Claude Code. He wrote an article basically arguing that HTML is a better default documentation format for LLMs vs markdown for many use-cases. It’s worth a read.
While there’s compelling criticism in the camp of “a token factory creator tells everyone to use more tokens”, I think it’s undeniable that HTML is becoming a new LLM primitive alongside markdown. All of the new hot design tools from Claude Design to Paper & Pencil and beyond are all HTML-based. Claude’s own “generative UI” feature that can create interactive charts, diagrams, and visualizations right inside chat is also HTML-based. It’s not unreasonable to expect that more generative UI interfaces will keep shifting to HTML going forward.
I think this is an apt observation and generally rings true in my own experience. This is also consistent with a lot of chatter I’m seeing online and from my own peers about the evolving etiquette around deliverables - it’s becoming rude to share “unsanitized” AI output. Everyone has accepted that everyone else is using AI for their deliverables, but it doesn’t mean that you can expect everyone to wade through your AI slop. It’s now an unspoken (and soon probably explicit) expectation that you rewrite it, prune it, proofread it, etc.
Carveouts
I’m wrestling with a HyperFrames plugin to create a launch video for my new product (in no small part an inspiration for today’s topic).
I finished“Ctrl + Alt + Chaos” - a book about the evolution of teenage hacking groups. I’m generally a sucker for anything regarding large-scale hacking, scamming, and conning. This was is a worthy read if you’re into similar topics.
I want to ask you for a favor - if you’ve been enjoying the newsletter, please consider sharing it with your team / colleagues / friends. It really helps, especially in these early stages as I’m posting into the social media void and looking for like-minded people. To those who already shared - thank you, it means a lot. See you next week.
Andrej Karpathy actually wrote a great post talking about #1 and #2 as a perception gap in AI capabilities.