Full Width 
Shortcuts 
Sign Up 
Log In 
Articles 
Reading List More from A Beautiful Site
As software engineers, we take pride in hand-written code and knowing how everything we build works in detail. While there are many developers out there who build good software, there are many, many more who don’t. Yet we balk at the idea of vibe coding, even when it yields decent results. You don’t need a scientific study to realize the mean quality of human-authored software leaned towards slop even before AI was around. How many times have you been frustrated by an app or website that just didn’t work right? (This is especially true for software produced by big companies.) The quality of software humans can build is absolutely incredible. But the quality of software humans actually do build is usually not. So no, AI can’t build perfect software. Maybe that will improve in the future, or maybe it will remain just as messy as the human-authored code LLMs have been trained on. I don’t think any of that matters, because if AI can build software that’s at least as good as the mediocre code that came before it, that will be acceptable for most industries. Customers don’t care about the craft, they care about the result. And if the result is faster, cheaper, and good enough…well, that’s a better combination than anything they’ve ever had before. Why wouldn't they choose it?
I launched ColorCopy on May 18. It was my second macOS app, but the first one I launched on the Mac App Store. Today, I decided to pull it. The app isn't a new idea, but it bakes four color tools into a single menu bar app: an eyedropper, a color picker, a palette manager, and a contrast checker, each one a hotkey away. I made it available for free, with a one-time in-app purchase to unlock unlimited use. No subscription, no recurring fees. It's simple, stable software with the lowest possible barrier to entry…exactly the kind of utility the App Store seems made for. I wasn't sure what to expect, so I ran an experiment to answer some questions: Is there any real benefit to having your Mac app on the App Store? Is the Apple tax worth it? Will customers just come flooding in? For ColorCopy's release, I did zero marketing. No Product Hunt launch. No Hacker News post. The only things I published were this blog post and a tweet. If the App Store delivers on its promise of discoverability, that should be enough to see some kind of traffic…right? You be the judge. In nearly three months on the App Store, ColorCopy got a little over 1,000 impressions, 149 product page views, 85 first-time downloads, three in-app purchases, and $21 in proceeds. Not per day. In total. Broken down over 81 days, that averages out to: 13 impressions per day 1.8 product page views per day 1 download per day 0.04 in-app purchases per day (about one per month) $0.26 in proceeds per day Extrapolated to 12 months, ColorCopy would have earned about $95…not even enough to cover Apple's $99 annual developer fee. If discoverability isn't a part of the App Store deal, what exactly is the benefit? Why limit your Mac app to the sandbox?* Why spend hours in App Store Connect filling out metadata, screenshots, and localization information? Why wait an arbitrary number of days for someone to review your app with the consistency of a coin flip? It doesn't seem worth it for Mac developers. Adios, App Store 👋 As of today, ColorCopy is self-distributed and sold through Polar (the new Stripe, which I highly recommend). This is the same way I sell TongueType, which makes pretty much everything easier on my end: one dashboard, real customer relationships, no finicky review process, and updates ship the moment they're ready. The move required swapping out in-app purchases from StoreKit to Polar, and updates now ship through Sparkle instead of the App Store's built-in update mechanism. Both are tried and true solutions for self-distributed Mac apps. To be fair, you can't really do this on iOS. Most users aren't jailbroken…walled garden and all. But on Mac, where self-distribution is still a first-class option, there seems to be zero incentive to be in the App Store, especially if you're counting on discoverability. In three months, the App Store sent me barely a trickle of customers. Had I marketed the app myself, the traffic would have flowed the other way. I would've been sending my customers to Apple and paying a tax for the privilege.** Self-distribution gives me the freedom and control over my apps that the App Store's sandbox never will. If it's on me to drum up all of my own traffic, I'm going to send it to my own website. *I originally used a third-party library for the eye dropper because NSColorSampler is meh, but the sandbox forbids it so I was forced to remove it. Yes, I had to make my app shittier in order to put it on the App Store. **I fully acknowledge this may not be the case for every app, but it was my experience and worth sharing. Your mileage may vary. Aside: ColorCopy was available in 10 languages on the App Store. For some reason, the listing always showed FR as its primary language (it wasn't). The app is available in many languages, including English.
I built the dictation app I wanted. It's called TongueType, and my daughter did the voice over for the video. (Family business.) It hasn't gotten much traction yet, and I think I know why: dictation has been bad for so long that most people stopped paying attention. I don't blame them. But I don't think most people realize how good local models have gotten. The thing that was flaky and frustrating five years ago is genuinely good now, and it runs entirely on your Mac. Downloads are low. But the conversion rate is great. The people who actually try it tend to stick around, which tells me the problem isn't the app, it's getting someone to give dictation one more honest chance. The real hurdle is the habit The hard part isn't accuracy. It's that talking instead of typing is a new habit, and new habits are awkward before they're automatic. For the first week it feels strange. You catch yourself reaching for the keyboard out of muscle memory. Then one day it clicks, and you realize how slow typing was making you for certain things. I still write code by hand. That's thinking, not transcribing, and I want my fingers on it (plus saying HTML tags and attributes out loud seems counterintuitive 😂). But for almost everything else, I talk. I prompt LLMs I send emails I reply on Slack I write commit messages I do most other text with my voice The common thread is that the thinking is already done and the only thing left is getting the words out. That turns out to be a surprising amount of my day. My desk setup On the go, the MacBook's built-in mic works just fine. You don't need fancy hardware to get good results. But when I'm at my desk, my laptop is docked, so I pair TongueType with a Tula mic. It's small, portable, sounds great, and looks the part! (Kuru Toga mechanical pencil positioned for size comparison.) One tip: use a wired mic if you can. Bluetooth adds latency, and latency is the enemy of a tool you reach for dozens of times a day. A wired connection makes the whole thing feel instant. Honest caveat: it's probably not a great look in a co-working space. 😂 "Why is this better than macOS dictation?" More accurate. Less flaky. More customizable. More fun. I could list the features, but the honest answer is the same one I'd give for most things: just try it and you'll feel the difference immediately. It's free to start. What's your experience been with dictation? If you wrote it off years ago, I'd genuinely love for you to give TongueType a shot and tell me what you think.
I just launched my first macOS app called TongueType. It's voice dictation that runs entirely on your Mac. Hold a key, speak, release, and your words appear wherever your cursor happens to be. I build small, simple, stable software. TongueType fits that description, and it scratches an itch I've had for a while. Why I made it I type fast, but I often think faster than I type. When an idea is fully formed in my head, the bottleneck is my fingers. macOS has had built-in dictation forever, but I never liked relying on it. Accuracy aside, I didn't love the idea of my voice taking a trip to a server and back just to write a sentence. There are many dictation apps in the wild, but I want one that's privacy focused, doesn't send data to the cloud, doesn't charge a monthly subscription, and gets out of the way. TongueType uses OpenAI's Whisper model running locally on Apple Silicon. Nothing is uploaded. Nothing is logged. There's no account to create. Zero telemetry. Your voice never leaves your Mac. How it works The whole interaction is one key. By default it's the Right Option key, because it's sitting right there and your thumb isn't doing anything important. Hold it, talk, let go. The transcribed text is inserted at your cursor…in your editor, your email, a chat box, a search field, anywhere text goes. You can also drop in an audio or video file — WAV, MP3, MP4, MOV — and get a transcript back. Handy for meeting recordings and voice memos. A few things I sweated the details on: A grace period so a quick accidental tap doesn't start recording. Double-tap to latch for when you want to keep talking without holding the key down. Cancel phrases — say "scratch that" at the end and the whole thing gets discarded. You will use this more than you expect. Spoken symbols — say "new line" or "question mark" and you get the symbol, not the words. Post-processing — for common terms that seldom get dictated properly. TongueType speaks twelve languages and includes automatic detection, so you don't have to tell it which one you're using. How I actually use it Building the app was one thing. Using it every day turned out to be another. A couple months in, it's quietly worked its way into most of what I do at the keyboard: Prompting LLMs. Talking to an AI assistant is conversational by nature, and typing out a long, detailed prompt is tedious. Speaking it isn't. I get more context into a prompt because I'm not rationing my words to save my fingers. Email. Replies that used to sit in my drafts now get spoken out in a single pass. I still read them before sending, but the blank-page friction is gone. Code comments and commit messages. The parts of coding that are just writing. It's faster to explain why a change exists out loud than to stop and type it. Direct messages. Quick replies in chat without breaking flow. Hold the key, say it, done. The common thread: TongueType is best wherever the thinking is already done and the only thing left is getting words out. That's a surprising amount of my work day. A fun personality TongueType is minimal and fun. It lives in the menu bar. The recording overlay is small and out of the way, and you can configure its position on screen. There are twenty accent colors including Rainbow Mode. None of these extras were necessary, but all of it was fun to build. Accessibility I want to call this out specifically. Voice dictation isn't only a convenience. For some people it's the difference between using a computer comfortably or not. If typing is painful or difficult for you, TongueType is built to be a genuine alternate input method, not an afterthought. That mattered to me, and it shaped a lot of the decisions above. Pricing TongueType is free to try, and the free tier includes every feature. You get 30 minutes of live dictation each month and short file transcriptions. If you want unlimited, TongueType Pro is a one-time $19.99 purchase that covers up to five Macs and unlocks unlimited dictation and full-length file transcription. No subscription. Buy it once, keep it forever. Requirements TongueType needs macOS 14 or later on an Apple Silicon Mac (M1 or newer). The local model is the whole point, and that's what makes it run so well. If any of this sounds useful, give it a try at TongueType.app. It's free to start, and I'd genuinely like to hear what you think.
I believe artificial intelligence is a powerful and valuable tool that can significantly improve how we create, solve problems, and bring ideas to life. I didn't always feel this way.. But these days, I use AI regularly in my work and I expect that to continue. We may not have chosen this reality, but it's the reality we're in. The tools and their benefits — costs be damned — can no longer be ignored. That said, when people ask "was this made with AI?" the honest answer is rarely simple. AI can speed up many parts of the process, but it doesn’t replace human judgment, creativity, or responsibility. What appears effortless on the surface rests on deliberate human direction, critical thinking, and careful review. Getting good results from AI requires active guidance and oversight. The nuances of context, ethics, user needs, and real-world application are simply too varied given the current technology. Moreover, AI doesn’t generate meaningful ideas or elegant solutions on its own. Strong human vision, architecture, and decision-making are still essential. There is no prompt, model, or service that can deliver finished, trustworthy work without substantial human input. My commitment to you is this: everything I create will be driven by human ideas, architecture, verification, and final review. I will use AI as an assistant to do what I would have done anyway, but more efficiently. I will not let AI replace my intelligence, but I will use it to turn my intelligence into code faster. — Cory LaViska
More in programming
A clip of me singing a funny song from Gilbert and Sullivan’s Ruddigore back in 2013
In this video, we look at why fork() needs copy-on-write, how it works inside the kernel, and a memory usage problem that Instagram encountered with Python.
Comments require commitment, but they’re worth it.
An aggregation is some kind of summary of a set of data. This can be the sum, length, minimum, etc. It is quite common to want to calculate such a summary repeatedly, e.g. “the maximum noise level in dB for the past 30 seconds” for a nuisance detector. In such a case we say there is a sliding window over our data, and we want to aggregate over our window. If our aggregation is a binary operator with an inverse, like integer sums, there is a very easy solution using a double-ended queue: from collections import deque class SlidingWindowSum: def __init__(self): self.sum = 0 self.elems = deque() def push(self, x): self.sum += x self.elems.append(x) def pop(self): self.sum -= self.elems.popleft() def eval(self): return self.sum But what if our operator has no inverse? This is actually the case for most interesting summaries such as minimum, quantile, approximate unique count (for example using HyperLogLog), etc. In fact, even something as simple as a floating-point sum suffers from the fact that floating-point addition is not invertible. For example, if you ever have a NaN in your input data with the above naive algorithm your sum will forever remain NaN, even long after the bad value has left your window. Six years ago I came up with an algorithm for maintaining just the minimum/maximum in a sliding window and posted it to cs.stackexchange. I now consider this algorithm pointless, because it turns out there is a simple and efficient algorithm that solves this problem for a very wide class of aggregations. I’m writing this blog post to spread the word, because I feel it should be more widely known. Folklore I came across this algorithm while reading a far more advanced paper, Low-Latency Sliding-Window Aggregation in Worst-Case Constant Time by Tangwongsan et al. Why is this paper titled low-latency? Because it does the same as what I’m about to describe, but in O(1) time for each step. However, in it they also described a “two-stack” algorithm, which does it in amortized O(1), and is far, far simpler. Amortized O(1) means that across many operations the total amount of work per element is constant, but an individual operation can take much longer. This is almost always fine, unless you absolutely need a low upper bound on latency. Funnily enough that paper attributes this algorithm to “adamax” from a 2011 Stack Overflow post. They in turn credit a 2001 lecture note by D. Sleator for the inspiration. However, this lecture note does not describe a sliding window aggregate, it describes the classical two-stack algorithm for implementing a FIFO queue and does amortized analysis on it. Ultimately I would not be surprised to find that this algorithm was already described in an obscure paper from the 1970s, seeing how simple and brilliant it is. Two stacks Like the authors of the paper, I will generalize the two-stack algorithm to arbitrary associative aggregation functions. By abstracting the aggregation as a set of functions, empty(), unit(x), combine(x, y) and finalize(x), you can describe many possible aggregations, for example a mean: empty = lambda: (0, 0) unit = lambda x: (x, 1) combine = lambda x, y: (x[0] + y[0], x[1] + y[1]) finalize = lambda x: x[0] / x[1] if x[1] else None I’d like to note here that these functions have the following signatures: fn empty() -> Agg; fn unit(x: Value) -> Agg; fn combine(x: Agg, y: Agg) -> Agg; fn finalize(x: Agg) -> Out; I’m making a distinction here between Value, Agg and Out because while they seem superficially similar for something like an integer sum, for an approximate unique count on strings you would have (Value, Agg, Out) = (String, HyperLogLogSketch, u64), three wildly different types. Without further ado, the algorithm: class TwoStackAgg: def __init__(self): self.values = [] self.values_agg = empty() self.cum_aggs = [] def push(self, x): self.values.append(x) self.values_agg = combine(self.values_agg, unit(x)) def pop(self): if not self.cum_aggs: cum_agg = empty() while self.values: cum_agg = combine(unit(self.values.pop()), cum_agg) self.cum_aggs.append(cum_agg) self.values_agg = empty() self.cum_aggs.pop() def eval(self): return finalize( combine(self.cum_aggs[-1], self.values_agg) if self.cum_aggs else self.values_agg ) That’s it, the entire algorithm. There’s two stacks (values and cum_aggs) and one more aggregate, values_agg. At any point in time values_agg holds the aggregate of values, and cum_aggs contains the cumulative aggregates of all values in our window that aren’t in values, in reverse order. From this we can get the aggregate over our entire window in constant time by by combining the last value of cum_aggs with values_agg. The neat part is that (assuming w is our window size) every wth operation we drain all of values and maintain a running aggregate while pushing the partial cumulative aggregates onto cum_aggs. This is what makes it amortized O(1), doing O(w) internal operations every wth pop bounds the total amount of work per element to O(1), even though a singular operation might not be constant time. I think this is best visualized. Suppose we sum [1, 2, ..., 10] with a fixed-size sliding window of four elements, then the state on each eval() call would look like this (values_agg not shown as it is simply the aggregate of the values): cum_aggs values out [] [] = 0 [] [1] = 1 [] [1, 2] = 1 + 2 [] [1, 2, 3] = 1 + 2 + 3 [] [1, 2, 3, 4] = 1 + 2 + 3 + 4 [4, 3 + 4, 2 + 3 + 4] [5] = 2 + 3 + 4 + 5 [4, 3 + 4] [5, 6] = 3 + 4 + 5 + 6 [4] [5, 6, 7] = 4 + 5 + 6 + 7 [] [5, 6, 7, 8] = 5 + 6 + 7 + 8 [8, 7 + 8, 6 + 7 + 8] [9] = 6 + 7 + 8 + 9 [8, 7 + 8] [9, 10] = 7 + 8 + 9 + 10 [8] [9, 10] = 8 + 9 + 10 [] [9, 10] = 9 + 10 [10] [] = 10 [] [] = 0 In total the memory usage is O(w), where w is your maximum window size. Note that for simplicity of analysis and the example I assumed a fixed-size window w, but there is nothing about the two-stack algorithm that requires this. You can call push(x) and pop() as many times as you’d like between each eval(), growing and shrinking the window size as needed. Floating-point non-associativity Note that we required above that our aggregate combine is associative, meaning: combine(combine(x, y), z) = combine(x, combine(y, z)) Technically speaking, floating-point addition doesn’t respect this. Nevertheless, the above algorithm is still very useful because the results closely match the expected outcome, even more so if you use a compensated summation algorithm like Kahan summation. Another neat thing about the two-stack algorithm is that it doesn’t require commutativity, if you follow the above implementation precisely. The order of operands is maintained, which can matter for things like string concatenation. However, there is a second very useful property of the above algorithm. Each aggregate is strictly a combination of the elements in the window, and none outside the window. This means if your window contains a NaN or infinity (or some other outlier), that value only poisons the windows that contain it rather than the rest of your computation. But even without NaN or infinity it is useful, due to not propagating errors endlessly. E.g. if your sliding window starts with [1e20, 1], this is what would happen with a naive rolling sum: >>> 1e20 + 1 - 1e20 - 1 -1.0 Compensated summation will reduce these effects, but not making your result depend on values outside of the window will eliminate long-term error accumulation entirely.
