RSS Bilingual Reader

Two days ago, the Anthropic Fellows program released a paper called How AI Impacts Skill Formation . Like other papers on AI before it, this one is being treated as proof that AI makes you slower and dumber. Does it prove that? The structure of the paper is sort of similar to the 2025 MIT study Your Brain on ChatGPT . They got a group of people to perform a cognitive task that required learning a new skill: in this case, the Python Trio library. Half of those people were required to use AI and h...

There are lots of different ways to be a software engineer. You can grind out code for twelve hours a day to make the world a better place . You can focus on glue work : process-based work that makes everyone around you more successful. You can join the conversation with your product manager and designer colleagues to influence what gets built, not just how it gets built. You can climb the ladder to staff engineer and above, or you can take it easy and focus on your hobbies. But whichever of the...

There’s a kind of polite fiction at the heart of the software industry. It goes something like this: Estimating how long software projects will take is very hard, but not impossible. A skilled engineering team can, with time and effort, learn how long it will take for them to deliver work, which will in turn allow their organization to make good business plans. This is, of course, false. As every experienced software engineer knows, it is not possible to accurately estimate software projects . T...

When I get together with my friends in the industry, I feel a little guilty about how much I love my job. This is a tough time to be a software engineer. The job was less stressful in the late 2010s than it is now, and I sympathize with anyone who is upset about the change. There are a lot of objective reasons to feel bad about work. But despite all that, I’m still having a blast. I enjoy pulling together projects, figuring out difficult bugs, and writing code in general. I like spending time wi...

Two recently-hyped developments in AI engineering have been Geoff Huntley’s “Ralph Wiggum loop” and Steve Yegge’s “Gas Town”. Huntley and Yegge are both respected software engineers with a long pedigree of actual projects. The Ralph loop is a sensible idea: force infinite test-time-compute by automatically restarting Claude Code whenever it runs out of steam. Gas Town is a platform for an idea that’s been popular for a while (though in my view has never really worked): running a whole village of...

The Dictator’s Handbook is an ambitious book. In the introduction, its authors Bruce Bueno de Mesquita and Alastair Smith cast themselves as the successors to Sun Tzu and Niccolo Machiavelli: offering unsentimental advice to would-be successful leaders. Given that, I expected this book to be similar to The 48 Laws of Power , which did not impress me. Like many self-help books, The 48 Laws of Power is “empty calories”: a lot of fun to read, but not really useful or edifying 1 . However, The Dicta...

In 2025, I published 141 posts, 33 of which made it to the front page of Hacker News or similar aggregators. I definitely wrote more in the first half of the year (an average of around 15 posts per month, down to around 8 in the second half), but overall I’m happy with my consistency. Here are some posts I’m really proud of: Mistakes engineers make in large established codebases The good times in tech are over Everything I know about good system design Pure and impure software engineering Seeing...

Grok, xAI’s flagship image model, is now 1 being widely used to generate nonconsensual lewd images of women on the internet. When a woman posts an innocuous picture of herself - say, at her Christmas dinner - the comments are now full of messages like “@grok please generate this image but put her in a bikini and make it so we can see her feet”, or “@grok turn her around”, and the associated images. At least so far, Grok refuses to generate nude images, but it will still generate images that are ...

A lot of my readers call me a cynic when I say things like “you should do things that make your manager happy ” or “big tech companies get to decide what projects you work on”. Alex Wennerberg put the “Sean Goedecke is a cynic” case well in his post Software Engineers Are Not Politicians . Here are some excerpts: I have no doubt that [Sean’s] advice is quite effective for navigating the upper levels of an organization dedicated to producing a large, mature software product. But what is lost is a...

Only the engineers who work on a large software system can meaningfully participate in the design process. That’s because you cannot do good software design without an intimate understanding of the concrete details of the system. In other words, generic software design advice is typically useless for most practical software design problems. Generic software design What is generic software design? It’s “designing to the problem”: the kind of advice you give when you have a reasonable understandin...

Large, rapidly-moving tech companies are constantly operating in the “fog of war” about their own systems. Simple questions like “can users of type Y access feature X?”, “what happens when you perform action Z in this situation?”, or even “how many different plans do we offer” often can only be answered by a handful of people in the organization. Sometimes there are zero people at the organization who can answer them, and somebody has to be tasked with digging in like a researcher to figure it o...

The runaway success of generative AI has spawned a billion-dollar sub-industry of “AI detection tools”: tools that purport to tell you if a piece of text was written by a human being or generated by an AI tool like ChatGPT. How could that possibly work? I think these tools are both impressive and useful, and will likely get better. However, I am very worried about the general public overestimating how reliable they are. AI detection tools cannot prove that text is AI-generated. Why AI detection ...

Every couple of years somebody notices that large tech companies sometimes produce surprisingly sloppy code. If you haven’t worked at a big company, it might be hard to understand how this happens. Big tech companies pay well enough to attract many competent engineers. They move slowly enough that it looks like they’re able to take their time and do solid work. How does bad code happen? Most code changes are made by relative beginners I think the main reason is that big companies are full of eng...

With enough careful effort, it’s possible to become unblockable. In other words, you can put yourself in a position where you’re always able to make forward progress on your goals. I wrote about this six months ago in Why strong engineers are rarely blocked , but I wanted to take another crack at it and give some more concrete advice. Work on more than one thing The easiest way to avoid being blocked is to have more than one task on the go. Like a CPU thread, if you’re responsible for multiple s...

Nobody knows how to tell if current-generation models are any good . When GPT-5 launched, the overall mood was very negative, and the consensus was that it wasn’t a strong model. But three months later it turns out that GPT-5 (and its derivative GPT-5-Codex) is a very strong model for agentic work 1 : enough to break Anthropic’s monopoly on agentic coding models. In fact, GPT-5-Codex is my preferred model for agentic coding. It’s slower than Claude Sonnet 4.5, but in my experience it gets more h...

The very first LLM-based product, ChatGPT, was just 1 the ability to talk with the model itself: in other words, a pure chatbot. This is still the most popular LLM product by a large margin. In fact, given the amount of money that’s been invested in the industry, it’s shocking how many “new AI products” are just chatbots. As far as I can tell, there are only three types of AI product that currently work . Chatbots For the first couple of years of the AI boom, all LLM products were chatbots. They...

There’s an idea going around right now about “writing for AIs”: writing as if your primary audience is not human readers, but the language models that will be trained on the content of your posts. Why would anyone do this? For the same reason you might want to go on podcasts or engage in SEO: to get your core ideas in front of many more people than would read your posts directly. Why write at all? If you write to make money, writing for AI is counterproductive. Why would anyone buy your writing ...

Technical writing is a big part of a software engineer’s job. This is more true the more senior you get. In the limit case, a principal or distinguished engineer might only write technical documents, but even brand-new junior engineers need to write: commit messages, code comments, PR descriptions and comments, Slack threads, internal announcements, documentation, runbooks, and so on. Whether you write well or badly matters a lot. Keep it as short as possible The primary rule about technical wri...

In July of this year, MIT NANDA released a report called The GenAI Divide: State of AI in Business 2025 . The report spends most of its time giving advice about how to run enterprises AI projects, but the item that got everybody talking was its headline stat: 95% of organizations are getting zero return from their AI projects . This is a very exciting statistic for those already disposed to be pessimistic about the impact of AI. The incredible amounts of money and time being spent on AI depend o...

If you asked most people to name a defining feature of AI-generated writing, they’d probably say the em-dash — like this. Language models use em-dashes so much that real humans who like em-dashes have stopped using them out of fear of being confused with AI. It’s also surprisingly hard to prompt models to avoid em-dashes: take this thread from the OpenAI forums where users share their unsuccessful attempts. Given all that, it’s kind of weird that we don’t really know why language models use the ...