close

On AI

Here are notes I want to grow into dedicated pages at some point.

“It’s always easier to grab a tool and bypass the mess of coordination, even if that means doing more and doing it alone.” — Frank Chimero, Beyond the Machine

Human traces

  • In this piece, Matt Ström-Awn explores the idea of “expansion artifacts”; they’d be like the opposite of “compression artifacts” that are just blobs of pixels inherited from saving a JPGs too many times over. We blow up our prompts with massive amounts of content. We fill the internet with this new content which feeds the next generation of models thus creating a negative training loop, a “model collapse” where the original human trace is lost.

  • Every language erreur, every tyypo, every hidden message (more or less subtly) delivered to the reader, every dissonance in our texts and creations are now a trace of the passage of Man — a way of writing “I was here” in code, same as if we’d write it on the bark of a tree far from any known path. We doubt anybody would ever come this wayy. But if they do, a part of us would surely connect.

  • Visible and obviously human mistakes are a form of “I was here”. The opposite of AI mistakes, which are often not visible straight away; they just appear under scrutiny. AI mistakes, hallucinations, always look right at first glance and just get weirder and weirder the longer you stare at them. Visibly human mistakes just look wrong at first glance and I think there’s almost something romantic about them since GenAI.

  • LLMs are everywhere even where they are not. I’m creating, I’m writing, I am the one writing this word here at this exact moment and I keep wondering: “will this read like an LLM wrote this?”. “Does my website look like it was generated by an AI?”. College students supposedly suffer from “authenticity anxiety” or “AI suspicion” these days. But I think any creator will have to outgrow that specific form of paranoia in the future.

  • I need to put a name on that moment between the prompt and the LLM response. I sometimes just contemplate my terminal window or look at the ceiling until Ghostty rings me back with its cute notification sound effect. What do we think about in these moments? To me, they are deep breaths. Sometimes they are also a chance to think ahead. Or sideways perhaps.

  • I do admit doing more and more multi-sessions with AI, leaving me with less and less of these moments. There are times for it, and maybe other times where I should probably step back from multitasking with AI. Productivity can be like a snake eating its tail. If I don’t leave myself a space to think and doubt, how can I produce any decent work?

Design engineers & market dynamics

  • A new title for designers is emerging: “design engineer”. This article from Anna Lefour says it’s nothing new; “hybrid designers” or “creative developers” always existed (I was one), but she argues AI has lowered the barrier to entry. The title only reflects what the market needs: people with full ownership of the product, from research to early conception, to project planning and finally implementation. So since titles are free to adopt, I will now be a Design Engineer. Which is another term for Product Designer. Which was another term for Software Designer. Which was another term for… webmaster. I make websites and the market makes my titles.

  • “What the market needs” reminds me of this post from Dan Luu and the phrase “markets for lemons”. In his framing, he argues that hiring developers (and I’ll extend to design engineers nowadays) is not a market for lemons because there isn’t enough information asymmetry to pin it down like that. Employers can tell what a good design engineer looks like if they use the right heuristics. And designers who haven’t transitioned to using more GenAI aren’t even lemons to begin with. People are not immutable; they learn and evolve.

  • Also, what the market needs is people who can navigate the whole value chain. The goal? No longer organizing teams in silos: project people, designer people, developer people. No islands of people. Instead we’ll have islands of features groomed by a one-man band owning everything from research to development. But I’d send us back to the Frank Chimero piece: that promise just shifts compartmentalization from a team level to an individual one. We break the friction of cross-disciplinary collaboration to go faster, to break down products into even tinier pieces that can be owned solo; but that could also isolate people within those organizations. Is the collaborative process really a kind of friction we want to get rid of?

  • There are interesting parallels drawn between AI workflows and the taylorization at the end of the 19th century. I think in my reality they are still quite different as taylorization meant to create repetitive, quantifiable tasks while the AI workflow we have now simply isolate the worker to be able to take a part of product from 0 to 1. Software also being an evolving product, or let’s not reinvent the wheel: software as a service.

LLMs as a thinking tool

  • In this article Amelia Wattenberger proposes that we use LLMs as a tool for thought:

    • Preserving thoughts outside our working memory

    • Suggesting thoughts

    • Providing perspective and increasing the cycle speed

  • I like it. I’m interested in how to use it on the upstream design process; not necessarily how to generate prototypes, which I think has been covered a lot lately. The process of writing these notes should itself be journaled so I can formalize a method I can reuse later.

  • In line with my previous point, I’m mostly interested in how LLMs can help me structure thoughts. I can store thoughts organically and quite frankly chaotically in a notepad somewhere, but once thoughts pile up, I have a hard time organizing them, categorizing them; and more importantly, creating a larger narrative or structure that a reader could actually engage with. That’s where I intend to get help from the LLM.

  • I can snapshot “live thinking notes” in very casual language and use them with AI later. Each step of the thinking helps the AI sharpen its context. Nothing is throwaway if it helps the LLM refine the thought structure. The crucial part is the thought needs to remain mine.

Software futures

  • This article discusses the consequences to org structure in the age of AI. It sounds right, but for some reason I fail to really connect with the “how”. I think I’m much more interested in the “why” in the age of GenAI. As in: “why even make that product?”, “why is this useful to anyone now?”, “why can’t this be entirely transparent in my life?”, “do I really need to do any of this?”. An AI-augmented team of great designers, PMs and engineers in some flamboyantly new kind of org, all working in concert to make the best tax return app in the world; and all I can think of is “why do I need a tax return app now?”

  • I hear it’s the age of “personal software”. Supposedly we’ll all be building our own software to do things that only cater to us. Craig Mod says it’s the time of building and indeed, there has never been a better time for anyone to start creating software. But from my personal experience, this has also been the time of using a LOT LESS software overall. Tracking calories? I got an LLM tab for that. Tracking my gym routine? I got an LLM tab for that. And sure, carefully crafted software could do all of that better than me talking to my phone; but I have nothing to download, nothing to learn, nothing extra to pay, nobody trying to sell me extra stuff. Beyond the age of “personal software”, I feel like looming in the shadow is the age of the “one software”. The one software that will check any contract or NDA I sign, the one software that will book my appointments, fix bugs on my website, create my automations, send me the five best design links of the day, make my next PowerPoint, and produce perfectly valid data visualizations for that same presentation. The age of a thousand apps on your phone survives only because AI still has memory and context issues, and because software is more deterministic for now. But until when will that be true?

  • This could be anecdotal. Maybe I’m the weird one. Maybe people use a lot MORE software overall now than ever. In which case it makes total sense to keep worrying about what future successful product teams look like. We may always need software. So we might always need competent teams that make them. Maybe what I’m saying is: it’s become very attractive to pay 20 bucks for a Claude Pro subscription and a lot less desirable to pay 10 a month for a tax return app I hate using anyway, no matter how slick you make it.

Peak dissonance

“Participating in capitalism has always asked us to make a kind of peace with dissonance – between what we value and how we actually live, between the world we want and the systems we help perpetuate. But what I’m being asked to accept – and overlook – keeps expanding.”

  • The intro to this week’s Dense Discovery newsletter resonated a lot with me.

Frictions

  • This article in ArsTechnica left me conflicted. Teaching in the age of LLMs. On one hand, I can only empathize with teachers and professors having to figure that stuff out. On the other hand, we built that education system brick by brick. We assess and score students, we rank them, we make them compete against one another. In fact, I have a hard time imagining that standardized testing (for example) was built to benefit students at all. LLMs disrupting the way we designed education systems are not a very good argument against LLMs to me, but rather a much stronger argument that schools and universities aren’t always designed for the very people seeking education.