In early 2025, I started seeing people turn off their brain as they use LLMs1. They would have an LLM take an action (summarize text, write some code, etc.), and just assume that it worked2. This generally didn't work in early 2025 and the result was often quite silly. As LLMs have gotten better, I've seen more of this. Sometimes, people will try to get the LLM to write some code for them and basically just assume that it works3. Sometimes there's a human in the loop and, if the thing doesn't work, they'll ask the LLM to figure out the problem and solve it. Niklas Gruhn calls some variants of…
We previously noted that, while it's easier than ever to hit a particular quality bar by having coding agents use effective test techniques, software quality seems to be getting worse, indicating that whatever defaults developers are using may not work very well. Here, we test if simple instructions to agents to use particular techniques or libraries improve implementation correctness. Here, we re-use the Zstd implementation eval discussed in this comparison of agentic programming language effectiveness and, instead, compare different testing techniques and testing libraries when agents are…
I was curious how well the predictions of the most widely cited AI skeptic I've seen (Ed Zitron) have done, so I looked at how his predictions panned out. To disclose my own biases, I've never had a particularly strong pro or anti AI progress position. For example, in 2022, I did a comprehensive look at predictions Futurists made, including well-respected folks like Kurzweil and found them to be generally wrong on both the prediction results as well as the reasoning. On the flip side, in 2015, I wrote about how people were underestimating AI's ability to displace humans in jobs and have…
I used to wonder why I see so many more bugs than most people. I easily observe hundreds to thousands of bugs per week and nothing seems to work, but most people I talk to don't see anything like this. For a long time, I thought this had something to do with how I use computers but, over time, I've realized that it's mostly that people are hitting the same bugs and don't notice. If you're not a programmer, that's probably a better way to see the world, but I think curing quality/bug blindness is helpful for programmers. I've done this with a lot of friends and acquaintances (just by pointing…
The other day, I saw a viral tweet saying that people talking about how LLMs are causing slow, bloated, code are going to eat crow once they re-write everything in super-optimized assembly. We're not quite at the point where we want to write everything in assembly, but some variant of what Nolan Lawson said about testing, you can choose how many bugs you want now, which I less eloquently noted here, is becoming more true for performance. In response to a comment in my last post that the cost of formerly specialized performance work has dropped by many orders of magnitude and performance work…
There's been a lot of talk about the vulnpocalypse, to which I don't have much to add because I'm not a security person, but I haven't seen much discussion on the closely related (and to be fair, less serious, issue), the benchmarkpocalypse. While it's become easier than ever to make serious performance gains, it's also become easier than ever to reward hack a benchmark and make fake performance gains. The former is probably happening quietly across many different companies, but the latter is something I see at least once a week nowadays. Someone will claim they optimized X and got some huge…
This somewhat widely cited post (I keep seeing it cited, anyway) suggests that dynamic languages and/or languages that represent things more concisely are more token efficient. It seems to be cited enough that LLM search results agree. For example, when I searched for "dynamic vs static language token cost" (no quotes), Google's AI summary opened with Dynamically typed languages generally have a lower LLM token cost than traditional statically typed languages because omitting explicit type declarations makes the code more compact. Google's AI cited the same post, which suggests that some…
We're going to look at three different kinds of benchmarks, one set of calculations for baseline numbers for performance "napkin math" estimates, one set of AI model evals, and one on car tires. To build my intuition for things, I like thinking about them before seeing the explanation, so these are presented with the benchmark information first and the explanation later in case you want to think about your answer before seeing my thoughts. 29. A friend of mine is reviewing performance orders of magnitude to prep for computer performance interviews and found that…
I've been using AI fairly heavily since last November and the whole thing is a funny experience. An agent will do something that, if a human did it, you'd immediately fire them. My reaction, of course, is to act as if this is great and spin up a thousand agents so they can do even more of that. Mid-last year, I had GPT (maybe 5.0 or 5.1) try to find the source of a bug. Naturally, this code didn't have tests and git bisect wouldn't work, and it was a UI interaction bug for which I'm not even really qualified to write a test for, so I asked Codex to bisect between dates X and Y to find the…
There's a common narrative that Microsoft was moribund under Steve Ballmer and then later saved by the miraculous leadership of Satya Nadella. This is the dominant narrative in every online discussion about the topic I've seen and it's a commonly expressed belief "in real life" as well. While I don't have anything negative to say about Nadella's leadership in this post, this narrative underrates Ballmer's role in Microsoft's success. Not only did Microsoft's financials, revenue and profit, look great under Ballmer, Microsoft under Ballmer made deep, long-term bets that set up Microsoft for…
About eight years ago, I was playing a game of Codenames where the game state was such that our team would almost certainly lose if we didn't correctly guess all of our remaining words on our turn. From the given clue, we were unable to do this. Although the game is meant to be a word guessing game based on word clues, a teammate suggested that, based on the physical layout of the words that had been selected, most of the possibilities we were considering would result in patterns that were "too weird" and that we should pick the final word based on the location. This worked and we won. [Click…
There've been regular viral stories about ML/AI bias with LLMs and generative AI for the past couple years. One thing I find interesting about discussions of bias is how different the reaction is in the LLM and generative AI case when compared to "classical" bugs in cases where there's a clear bug. In particular, if you look at forums or other discussions with lay people, people frequently deny that a model which produces output that's sort of the opposite of what the user asked for is even a bug. For example, a year ago, an Asian MIT grad student asked Playground AI (PAI) to "Give the girl…
From 2011-2012, the FTC investigated the possibility of pursuing antitrust action against Google. The FTC decided to close the investigation and not much was publicly known about what happened until Politico released 312 pages of internal FTC memos that from the investigation a decade later. As someone who works in tech, on reading the memos, the most striking thing is how one side, the side that argued to close the investigation, repeatedly displays a lack of basic understanding of tech industry and the memos from directors and other higher-ups don't acknowledge that this at all. If you…
In 2017, we looked at how web bloat affects users with slow connections. Even in the U.S., many users didn't have broadband speeds, making much of the web difficult to use. It's still the case that many users don't have broadband speeds, both inside and outside of the U.S. and that much of the modern web isn't usable for people with slow internet, but the exponential increase in bandwidth (Nielsen suggests this is 50% per year for high-end connections) has outpaced web bloat for typical sites, making this less of a problem than it was in 2017, although it's still a serious problem for people…
If I ask myself a question like "I'd like to buy an SD card; who do I trust to sell me a real SD card and not some fake, Amazon or my local Best Buy?", of course the answer is that I trust my local Best Buy1 more than Amazon, which is notorious for selling counterfeit SD cards. And if I ask who do I trust more, my local reputable electronics shop (Memory Express, B&H Photo, etc.), I trust my local reputable electronics shop more. Not only are they less likely to sell me a counterfeit than Best Buy, in the event that they do sell me a counterfeit, the service is likely to be better. Similarly,…
On large platforms, it's impossible to have policies on things like moderation, spam, fraud, and sexual content that people agree on. David Turner made a simple game to illustrate how difficult this is even in a trivial case, No Vehicles in the Park. If you haven't played it yet, I recommend playing it now before continuing to read this document. The idea behind the site is that it's very difficult to get people to agree on what moderation rules should apply to a platform. Even if you take a much simpler example, what vehicles should be allowed in a park given a rule and some instructions for…
This is a set of notes on the Quinn Emanuel report on Cruise's handling of the 2023-10-02 accident where a Cruise autonomous vehicle (AV) hit a pedestrian, stopped, and then started moving again with the pedestrian stuck under the bottom of the AV, dragging the pedestrian 20 feet. After seeing some comments about this report, I read five stories on this report and then skimmed the report and my feeling is that the authors of four of the stories probably didn't read the report, and that people who were commenting had generally read stories by journalists who did not appear to read the source…
There's a class of comment you often see when someone makes a popular thread on Mastodon/Twitter/Threads/etc., that you also see on videos that's basically "Why make a Twitter thread? This would be better as a blog post" or "Why make a video? This would be better as a blog post". But, these comments are often stronger in form, such as: I can't read those tweets that span pages because the users puts 5 words in each reply. I find common internet completely stupid: Twitter, tiktok, Instagram, etc. What a huge waste of energy. or When someone chooses to blog on twitter you know it's facile at…
In The birth & death of search engine optimization, Xe suggests Here's a fun experiment to try. Take an open source project such as yt-dlp and try to find it from a very generic term like "youtube downloader". You won't be able to find it because of all of the content farms that try to rank at the top for that term. Even though yt-dlp is probably actually what you want for a tool to download video from YouTube. More generally, most tech folks I'm connected to seem to think that Google search results are significantly worse than they were ten years ago (Mastodon poll, Twitter poll, Threads…
This is a transcription of videos Elon Musk's appearance on stage with Dave Chapelle using OpenAI's Whisper model with some manual error corrections and annotations for crowd noise. As with the Exhibit H Twitter text message release, there are a lot of articles that quote bits of this, but the articles generally missing a lot of what happened and often paint a misleading picture of happened and the entire thing is short enough that you might as well watch or read it instead of reading someone's misleading summary. In general, the media seems to want to paint a highly unflattering picture of…
Read at the source
Your visit, your choice.
Optional Google Analytics helps us understand visits. Microsoft Clarity records masked interactions to improve the site. Optional tools stay off unless you choose them. Privacy details.