More from The DESK Magazine
If you think about it, there is *a lot* of software that hasn’t been built across hundreds of industries because our best paid designers and engineers were occupied with building social networks and ad networks. "The best minds of my generation are thinking about how to make
If you're reading this, you're already on it! In case you're reading this on your phone, I hope you can visit on desktop soon too. You know, the one with the big screen. The bigger the screen, the better. It works beautifully on tiny
I believe our aesthetic taste is incredibly important, maybe even more important than we admit because it reveals our deepest values inside of us. If you like plastic versus wood, if you're drawn to chrome or brass, to minimalism or maximalism, to sharp edges or soft curves.... it&
More in design
I walk our dog first thing in the morning and then again at lunch, and I have done this for years, long before any of what follows. Now I walk with my phone in my pocket and my earbuds in, and I talk. Twenty minutes, usually, sometimes longer. When I’m done, the transcript lands in that day’s note in Obsidian, in a template I’ve segmented for the purpose, and then it gets read — I’ve configured a Claude Cowork project to make sense of these notes. The part that still surprises me is that I don’t have to signal which mode I’m in. I can be working through an argument, lose confidence in a date halfway through, and say by the way, Claude, check that for me, I’m not sure — and then carry straight on with the argument. I can bury an instruction inside a paragraph that isn’t an instruction. Nothing marks the boundary. No wake word, no syntax, no switching into command voice. It is also, for a verbal person, simply the fastest way to get something out of my head. Faster than typing, and far faster than typing well. And I’m a pretty decent typist. What comes back But speed is the least of it. I come back from the walk, sit down, and what’s waiting is not a transcript. It’s a response. Frequently it’s some version of this: you said at the start that your point was X, and then you stumbled over something in the middle and said Y, and Y is the better idea — and it connects to this other thing you said twenty minutes earlier and didn’t notice. Earlier this week it went further than that. I had been talking about voice interfaces, and about how strange and good it is to be able to speak to a machine without formatting my speech for it. Here is what came back: Eight years ago you named the exact capability gap and called it the thing that didn’t make sense. Today you described, as the single most powerful thing AI has given you, being able to interleave thinking aloud and issuing commands with no syntax at all — and you said “this is everything Star Trek basically convinced me to expect” without, as far as I can tell, remembering that you’d already written the piece. I had to go and look. A human idiosyncrasy In November 2018 I produced a podcast episode about voice interfaces, and I opened it with a scene from Star Trek: The Next Generation. Data is alone, talking to himself, and the ship’s computer answers him — because it cannot tell the difference between a man working something out aloud and a man asking a question. The joke is that Data catches himself and calls the habit — thinking out loud — a human idiosyncrasy. Then, in the episode, I said this: Surely a computer that is listening all the time and is able to reply to direct queries would necessarily be able to distinguish between someone talking to themselves and someone asking it a question. I was complaining about what seemed to be an obvious logic hole in the plot. What I had actually done was specify a product. Claude was right. I’m the kind of nerd who cites Star Trek frequently enough that I don’t retain a proper index. Claude did. The argument I went on to make in that episode was that our expectations had shifted in the decades between the writers’ room and the smart speaker on the counter. The Enterprise’s computer answers questions and offers nothing else. What we wanted by 2018 was something a bit warmer. I put it this way at the time: we don’t want a computer we can talk to, we want a computer we can talk with. I meant something fairly soft by that. I was thinking about loneliness and trust and the strange affection people can develop for disembodied, machine voices. What I did not anticipate is that the preposition — with — would turn out to be load-bearing in a completely literal sense. The thing I now use every day is valuable almost entirely because of what happens after I stop talking. Speaking to an editor As a writer, this is what I’ve wanted for years. I tried to make it work around the same time I recorded that episode, dictating Google Keep memos on my commute home that I’d then turn into articles. But transcription was only a small slice of the work, and not the most important part. Because the real value of all this — speaking to my phone, having it transcribed, and then having the transcription read — is that it changes how you talk in the first place. You become aware that you’re speaking to an editor. And the machinery behind that isn’t default behavior. Indexing, cross-referencing and reframing are useful, but they aren’t the only things editors do. A good editor challenges, corrects, calls you on your BS. I have deliberately trained my machine to do all of that: be blunt, don’t praise me, don’t worry about hurting my feelings, I’ll be OK, just tell me the truth. It over-indexes on criticism because I told it to, knowing the untrained disposition is to agree. I get told that an idea is trite, or that everyone is already saying it, or that it’s flat out wrong and there isn’t an essay here at all except one that might embarrass me. So I’m not describing an oracle. I’m describing an instrument I calibrated, and the calibration is the part that took effort. That has happened enough times now that I’ve stopped being startled by it, though I’m still very appreciative. An essay I published this month exists because of exactly that intervention; I had the wrong thesis for about three thousand words and the right one was sitting in a subordinate clause. One instrument Last week I published something about medieval monks that argued the opposite of what I’ve just said, and I’d rather handle the seeming contradiction now than wait for someone to raise it. The argument there was that the hermit’s real problem is epistemic. Alone in his cave he has one instrument and it reads whatever he’d like it to, so he cannot tell drift from progress. The monastery was a machine for getting a second reading. And I was skeptical of feedback that costs the sender nothing — the like that has the form of accountability and none of the function. I suppose an AI fits that bill. A model bears no cost for telling me the truth. It isn’t a brother in the order and it has no stake in whether I improve. But it has no stake in the other direction either. It isn’t managing me, it doesn’t need the relationship to stay pleasant next week, and it will not be at the dinner party. That gets me a second reading without the social tax, which is a narrower thing than accountability and still worth a great deal when what you need is someone to tell you the first two pages are throat-clearing. The calibration matters here precisely because that stake is missing. The reason I had to go in and encode bluntness is that nothing in the arrangement supplies it on its own. The monk didn’t have to configure his brothers. Not your dog walk One last note, and I think it’s important. I don’t want this to read like advice to turn every idle moment into a productive one. Not every walk should be a business trip. There’s an entire genre built on that notion and it’s miserable, and I truly have no interest in telling anyone that the twenty minutes they spend walking their dog is an underperforming asset. What makes my version of this not that, I think, is a detail of sequence. I was already thinking on those walks. For years, the same two walks a day, the same half-formed arguments assembling themselves somewhere around the third block and dissolving by the time I got back to the house. The thinking was happening whether or not anything caught it. Nothing colonized the walk. Something started catching what the walk had always produced and I had always lost. This essay began as twenty minutes of talking, about four blocks from my house, with a dog pulling toward a smell. Most of what I said that morning isn’t here. The part you’ve just read is something I mentioned in passing, on the way to a point I thought I was making, and then walked past. The computer on the Enterprise would have let it go.
I haven’t driven a car in decades (New Yorker). But from my pre-New York life, I remember how sun glare could turn a pleasant road trip into a blinding gauntlet. Enter “Sunward,” a uniquely specific demo app my Automattic colleague Peter Westwood just whipped up in Spacefast. Spacefast – for your next idea Sunward – demo app Peter Westwood – software engineer & engineering director Automattic – my work home P.S. Apologies for layout wonkiness. Trying something new. The post Better Day, Sunshine appeared first on Jeffrey Zeldman Presents.
A decade ago, working at a design agency, I lived in fear of one sentence: “this looks nice, can we see a few other options?” I was not an ‘options’ guy. I fashioned myself as Paul Rand, telling Steve Jobs off when he asked for multiple options: “No. I will solve your problem for you and you will pay me. You don’t have to use the solution. If you want options, go talk to other people.”1 But I didn’t have the guts to tell my clients off, so I dutifully designed a few other options.2 Recently, I’ve gotten to work with some of the best designers in the field. Looking at their digital workspace — in Figma, Paper, Claude Design, or something else — I see hundreds of artboards and variants, laid out in a sprawling grid, representing branching parallel universes of design directions. Sometimes I see little panels filled with dials to sweep through the possibilities, enabling a single person to see thousands of combinations in the hope of picking a single, perfect specimen. Lots and lots of options. This makes me feel like an alien. My files are just a few artboards; sure, I tweak and tune the tiny details inside, but I rarely keep the multiverse in view as I do. I’ll give myself a dial here and there, but usually for a single variable that I can’t easily visualize. Am I just a bad designer? Assuming (hoping) the answer is ‘no’, I want to revisit a thread I pull on from time to time: as designers, we’re all just trying to find the right solution in a vast multiverse of possibilities. The solution exists, and we typically know it when we see it. So most of what we call ‘design’ is groping around a dark room looking for a black cat. Different designers take different approaches. Some are changing their approaches with the advent of generative AI, but at the end of the day, design is a search problem. Computer science has worked on search problems for decades. Depending on the particularities of a given search, there are a variety of algorithms and strategies that can result in faster, more direct, or more computationally efficient paths to the solution. We can apply this wealth of research from some of my favorite scientists (looking at you, Claude Shannon) to the design process, to hone our own craft. ‘Search’ has its own formal vocabulary in computer science. I’ve put a glossary at the bottom of the post, and I’ll use dotted underline to link to a definition if you need it. To picture how a search works, imagine a Rubik’s cube. A scrambled cube is the starting state; each turn you can make is a state change, leading to an adjacent state. All the possible combinations of colors on the faces (43 quintillion for a typical 3×3 Rubik’s cube) make up the search space. Searching involves evaluating candidates that can be reached from the current state of the cube; the desired state is the solved cube with each face a solid color. Each possible version of the cube is a state. Each turn is a change that leads to an adjacent state. Now, a few basic search algorithms and how they apply to design: Depth-first search As far as search algorithms go, depth-first search (DFS) is as simple as it gets. Here’s how it works: From the current state, look for adjacent states we haven’t visited. If there is one, move to it. It doesn’t matter which. If there isn’t, go back to the previous state and start from step 1. Repeat until we reach the goal or run out of states. Depth-first search commits to the first change it finds and follows that all the way down before it backs up. Solutions evaluated: 16 In a design sense, this is akin to navigating a hedge maze by picking a random direction at each fork until you hit a dead end. It’s not a very good design process, but it’s a good intro to how these things work. Breadth-first search Breadth-first search (BFS) is just a touch more complicated, and starts looking more like the design process. From the starting state, visit every adjacent state. From each of those, visit every adjacent state we haven’t visited. Keep going, moving one step further from the start each round. Repeat until we reach the goal or run out of states. Breadth-first search tries every single change, then every pair of changes, and so on. Solutions evaluated: 21 Both DFS and BFS are naive algorithms; we don’t proactively evaluate candidates while we search. Both are used on search problems with small search spaces and/or when evaluation is cheap. But many of the most interesting search problems aren’t small or cheap. Take chess, for example; winning at chess is a search problem. But the search space is unfathomably large,3 which is why it wasn’t until the mid-2000s that computers could reliably beat grandmasters. Most product design problems I’ve worked on look more like chess than a Rubik’s cube. Design covers search spaces of hundreds of dimensions — sizing, spacing, hierarchy, color, motion, interactivity, responsiveness, etc. And evaluation is relatively expensive: beyond changing a button or color, we often need to ship something before we truly know it works. There are other algorithms meant for this kind of problem, and they resemble design more and more as they get more complex. Beam search Beam search is the first approach we’ll explore that actually evaluates potential candidates before moving. From the starting state, score every adjacent state. Keep the three best and discard the rest. Score every state adjacent to the three we kept, and again keep only the three best. Repeat until one of the states we kept is the goal, or we run out of states. Beam search scores each option by how many parameters still look wrong, keeping the best three each round. Solutions evaluated: 21 This is called ‘beam search’ because it’s like using a flashlight’s beam to look at new states; the width of the ‘beam’ above is 3. Beam search is the ‘show me some options’ approach that most of my agency work took. It’s limited by the width of the beam: a narrower beam is faster and cheaper, but may miss the best solution at a given step. Now is an important moment to introduce the idea of a heuristic. In most cases, we don’t know we have the ideal candidate until we find it. Given that inescapable fact, we have to pick a test that we think will separate the good candidates from the bad ones. Heuristics usually involve tradeoffs: they can be fast/cheap and error-prone, or slow/expensive and accurate. Here’s the problem: beam search often fails to find the best solution! It depends entirely on the width of your beam and the quality of your heuristic. For example, if the desired state is hiding behind a low-scoring upstream branch, the algorithm might never find it. This can be overcome with an infinite beam, but then we’re just back at a slightly fancier BFS, which is slow and inefficient. When I learned the shortcomings of beam search — that simply checking a few options and picking the best one might not lead to the best solution — my feelings about my option-hungry clients were validated. And yet, designers are still using the beam, because AI has given us a better flashlight: producing a handful of options for a SaaS website used to take a week, and now it takes seconds. The beam has gotten so wide that ‘give me 10 options’ feels like an effective way to prompt AI towards the best design. Still, it’s not how I design. And again assuming I’m not just in the permanent design underclass, I thought maybe there was another algorithm that explains why I tend to explore so few branching paths. A* A* (A-star) is the algorithm behind most route-finding, from maps apps to video game characters. From the starting state, give every adjacent state a score, made up of: The number of changes it took to get there from the start, plus The heuristic: our estimate of how many changes remain to reach the goal. Put them all on a list of states we could visit next. Move to the state on the list with the lowest score. This might not be adjacent to our current state! Score the new state’s adjacent states and add them to the list. Repeat until we’re at the goal. A* scores each possibility as changes made so far plus its estimate of changes left. Partway through, it’s able to revisit previous candidates. Solutions evaluated: 14 This is similar to beam search, but with a few twists: A*’s heuristic incorporates a notion of distance from the goal. A* doesn’t always move smoothly from adjacent state to adjacent state. That last bit is crucial: unlike beam search, A* can go backwards. If the path we’re on moves far from the goal, its score climbs. Eventually a previous candidate looks better and we go back to it. More importantly, how much exploring we do in A* depends almost entirely on the quality of our heuristic. With a perfect estimate, A* goes straight to the solution. A* matches how I design. I’m more deliberate with exploration, only following thoughts until I feel they’re drifting away from the solution. I typically only keep old explorations around if they’re promising; an exhaustive record of every possible variant is stressful. I’ve managed to go the whole essay without mentioning it, but with my A* approach, taste4 does in fact matter; while there are many designers with better taste than I have, what I do have is hard-won by years of designing, and I can use it to sniff out the fastest path to the best solution. Conclusion AI has given us a wider beam, but it hasn’t yet given us a better heuristic. We’re still designing for humans, so we still have a leg up on the robots: we can use empathy, or even our own biases and human shortcomings, to feel out the right solution. Meanwhile, understanding the design process is just as important as ever. This kind of introspection — learning about search algorithms out of despair over my own shortcomings — has added a few tools to my arsenal. There are many variations on these algorithms that may describe your own search process; there are even more exotic variants that may be worth trying on. It might seem reductive to call design a search problem. Where’s the creativity? But with humanity, humility, and curiosity, pursuing the optimal solution with a well-tuned algorithm can help us design faster and more efficiently. Special thanks to Josh Petersel for feedback on a draft of this essay. Definitions Adjacent Two states are adjacent when they’re separated by a single state change. ↩︎ Candidate A state we’re considering. ↩︎ Distance The number of changes that turn one state into another. Usually we measure distance as the least number of changes required to get from one state to another. ↩︎ Heuristic A test we think will separate the good candidates from the bad ones, used because we usually can’t know we have the ideal candidate until we find it. Heuristics involve tradeoffs: they can be fast and cheap but error-prone, or slow and expensive but accurate. ↩︎ Search The process of arriving at a desired state (scoring well on a fixed set of criteria) by evaluating a set of candidates. ↩︎ Search space The set of possible states. We call it a ‘space’ because often there are ‘dimensions’ to search, like the width and height of a shape, and we can make analogies to the dimensions of a physical space. Sometimes these dimensions are literal, as when a wayfinding app picks the best route from point A to point B. ↩︎ State A possible step in the movement from the beginning of a search to the end. We start with the ‘initial state’; as we evaluate candidates we might change our ‘current state’; the end of the process is the ‘final state’ (and may not be the ‘desired state’). ↩︎ State change Also called a mutation. How the system moves between states. ↩︎ Footnotes & References As recounted by Steve Jobs in “Steve Jobs interview about working with Paul Rand,” 1993. ↩︎ The client almost always picked the original (or at least the classic ‘let’s do A with a bit of B and C’). ↩︎ Claude Shannon estimated it at 10¹²⁰. For comparison, the observable universe has about 10⁸⁰ atoms. ↩︎ Say the line, Bart! ↩︎