AI shenanigans

[AI burnout alert]

As a response to the insanity that was January, I invested a lot of (company) time dabbling with AI software. It would be dishonest not to say that this was driven by insistent corporate suggestion and the collective fear of irrelevance of the last few weeks.

Meta commentary Almost everything I describe as “clunky”, “esoteric”, or “distracting” isn’t really about output quality. It can do many useful things. It’s the kind of thinking this tool force me to do in order to use it. The subject matter is asserting itself through my nervous system. Writing about it ease anything. The cognitive loop is metabolically expensive: AI forces me to design in a medium that is hostile to how design thinking emerges. In practice, it becomes less a production tool than a thinking environment with a supervision tax. The resulting attention conflict is structural, not personal - which is unsettling, and not particularly fun.

Impressions

Claude desktop app is clunky and stuffed. The soft warm subtle vibe isn’t my jam. I appreciate that it keeps me away from image manipulation. Landing-page vibe-coding is fun for a second until you generate three that look the same. Editing is a struggle. Line-by-line instructions feel like standing behind an intern’s back. I’m not the target for Claude. For developers willing to set up integrations and keep them updated it seems to be fun. All the design skills I tried felt like refurbished js plugin demos. As a designer I’m not impressed but I’m not a dev. Claude is in high demand, especially during PST working hours, it tends to spin for hours on big requests.

Codex UI looks great. I used it for tiny CSS changes and ran into the same hurdles as elsewhere. Talking about design in plain language doesn’t go far. I tried providing flat design images. It doesn’t get it.

I’ve been using ChatGPT like a normal peon since its release. No notable changes. Trying to use it for work didn’t change my UX: lots of rewriting, summarizing, occasional image styling. Research and verbal sparring remain my favorite. Asking it to not be so sycophantic is annoying. It gets lost easily, especially with images. Asking to emulate a style or tone consistently ends up caricatural. It’s only acceptable for documentation and descriptions. Creating image is seriously fun. It’s been testing my discipline. It’s seriously good. I think I have developed that “intuition” for talking to GPT.

Gemini’s output is always too verbose; after two minutes I skim. Summarization on the other hand is very aggressive. It’s so fast.

Google AI Studio is opinionated but gave me the best results for UI tweaks from images and text instructions. Getting started was easy. Not having to setup any API key to use google AI models helped.

Flora was expensed with a pro account and 240k tokens, which I thought would allow me to get something out of it. I appreciated being able to try different models. Some are clearly better at certain tasks. But again I realized I’m not the target user. It’s for asset production and maybe ideation. My favorite image model remains GPT. I got better results with the same prompt via ChatGPT than Flora. The workflow UX with boxes and connectors is cool but esoteric. I got lost trying to fine-tune styling and composition. Talking about design involves a degree of abstraction that makes no sense to a machine clearly looking for keywords. Vector is still out of the equation for now, so Flora yielded nothing useful. I got sucked in building messy files, hoping adding more boxes would help. It feels like a game, not a professional tool.

Pencil.dev - Reminded me of sketch. Better visual design than the big american SaaS products - but still more of a toy than anything substantial. I can see their taster getting acquired more than going big.

Agent.expo.app - Dogfooding expo’s new product. Agent specialization makes too much sense for a react framework. It’s a hard product design challenge. Because I see it from the inside I see mostly the pitfalls. I can also see this the future of Expo: absorb snacks, IDE, combine Go + cloud service in one interface.

Bottlenecks

AI tools aim to be used to produce a lot more output until a new bottleneck is reached and requires human intervention. The bottleneck happened very quickly from my experience. That makes my experience very different from what I see in the news.

You can’t one-shot anything good. Quality is a function of iteration. Such plain common sense would not have been a great marketing tagline not so long ago. As a response to AI, it is today. What a leap forward.

I tried giving a few LLMs an outline and asking for a draft. The time it took to get the draft to a place I felt good about was far lower than what it took me to write my own version. My version remained superior. This is irrelevant considering global adoption of this technology, but it speaks to the loss of “quality.” Doctors, journalists, lawyers, and yes, developers — we need their best genuine thoughts. Review is a different cognitive effort from creation. Like cycling compared to running. Not better or worse but the mechanical assistance promotes a less “wholesome” engagement.

After a week of cluelessness I went over a hump. I started treating AI as a literal intern. I never asked for one but everyone tells me they’re the shit. They do have a lot of enthusiasm and we’re all speculating what to do with them. Sometimes fun, always frustrating, they teach us more about ourselves than their lacking substance.

Most of my time was spent rectifying, reframing, waiting. The wait was never productive nor creative space. At some point I had five tabs open with prompts running. The thrash made me unfocused and unfulfilled. I heard this has a name: token anxiety.

One of my major bottlenecks is attention. The constant interaction with Chats made me anxious in subtle ways. The thrash. The waiting. The reframing. The sense of always being slightly behind.

Each ecosystem wants focused attention. A chat is a conversation that needs to be steered. Like all conversations, frustration happens. The constant fast interactions disrupt my subtle solitary creative process. Chats on Slack with colleagues have a slower pace and ambiguity. AI chats are fast and precise. I’m not, and I dislike talking to INTJs (whatever the tone or personality).

I still feel like I’m using AI like a caveman. Lowering expectations while staying enthusiastic helps here too. Putting vision over execution creates an imbalance. Imbalance leads to failure or correction.

Design commoditization already started years ago with frontend frameworks. Ultra-cheap AI-generated UI is ramping up. It will barely be reviewable without a layer of AI assistance. It doesn’t seem to be evaluated by humans, at least none with decent taste. Meanwhile some people will write code manually for the sake of craft. Cleaning up AI generated design is easy and rewarding but not the creativity I signed up for.

Feelings

Machines summarizing thoughts that barely deserve to exist in full form feels like a strange loop. Most ideas are dead ends. AI keeps them alive while they should have died long ago.

Talking about design in plain language didn’t go far. It’s already tricky with humans. Editing AI output line by line feels like correcting an intern who is too confident. The abstraction layer between intention and execution is enormous.

Vibe-coded shaders are everywhere. I’m visually teased but cognitively left empty-handed. Without design experience, there’s currently no incentive to be a disciplined user. It’s very salient when seeing developers building stuff with AI. The lack of restraint is so obvious.

Hyper specialized AI tools, like humans are the most satisfying. My bet is that the great big project of AGI should or will never materialize. Instead, finely tuned and specialized tools will continue to shape the digital landscape - similarity to the highly collaborative dev tool space of today. Seeing AI as an equalizer seems foolish and yet not entirely wrong.

I realized taking these notes that my personal experience is irrelevant, even for myself. We collectively will benefit of new technologies while paying a cost for it. The attributes of « good design » will remain. The shape of it will change.

On a personal level AI is to my brain what cars are to my legs: Dramatically more effective for navigating the world we made, completely inadequate for navigating the world as it is. I’ll take the car when I have to, otherwise, I’m still walking.

---

I needed to get this post out for sanity's sake. If you read this message and felt something, you're not alone.