Blog Posts

I finally started looking into what jj actually is and how to use it. To wrap my head around it, I prompted DeepSeek V4.1 to build an interactive guide that assumes you already know git well, mapping common git tasks to their jj equivalents.

It took some back and forth prompting to get it right, I had to explicitly tell it to drop D3 and write the visualizations from scratch as it looks like the model handles its own plain code much better than wrangling in a library in. The result came out really nice for a quick experiment, and you can check it out here: jj — A Visual Guide for git users.

Quoting Benedict Evans's "AI, tools and transformation" (via)

It can be very tempting to think that AI will sweep most of this away. There's an old joke that an engineer is someone who'll spend an hour building a tool to automate a task that would take 10 minutes. But with AI, now you can make that tool in two minutes, and you don't need to be an engineer, and you don't need to write code. You can just ask the model to make the tool for you, or, more fundamentally, just do the task for you itself. Instead of having to create those tools one at a time, software might be dynamic, generative, free-form, and spontaneous. Massively more tasks can be automated, with massively less software.

Second, all of this means that software is bought or chosen or created on a spectrum from top-down to bottom-up - the company buys SAP and the user makes a spreadsheet - and I think it's useful to think of this also as a spectrum from institutionalised to improvised.

Quoting the Hacker News discussion on Ed West's "Doomscrolling Ourselves to Death" (via)

Learning is friction. The tool to remove friction also removes learning.

Can individuals sustainably rewire their attention spans back to long-form reading through deliberate behavioral changes?

Quoting Brennan Kenneth Brown's "Gemini, Gophers, and Fingers. Oh My! Alternative Internets Beyond HTTPS" (via)

Solderpunk's pseudonym instantly reminded me of Solarpunk, a speculative fiction genre and a real-world political and aesthetic movement. Cyberpunk imagines a future of neon-soaked corporate dystopia and steampunk looks backward to industrial romanticism, but solarpunk imagines forward—toward decentralized communities, renewable energy, mutual aid, technology that serves life rather than exploiting. The symbol is a half-gear for technology compatible with nature, and a sun for the infinite potential of the open imagination. The movement began in sci-fi and migrated into architecture, gardening, political organizing, and Internet culture.

Solarpunk doesn't worship technology, nor demonize it. Rather, solarpunk asks how we can use technology to improve humanity. It's interested in decentralized infrastructure, community engagement, and a DIY attitude. It distrusts monocultures, suspicious of scale for scale's sake, asking: who benefits from technology, and at what expense?

Quoting Dan Shipper's big AI takeaways (via)

• 2. Automation is a lie—every automation needs a human. Dan's company doubled in size this year despite being incredibly AI-forward.
• Why? Because in order to make automation work well, you need humans making sure everything keeps working.
• This is why benchmarks are misleading—they measure AI on problems we've already framed and can score, but there's always a higher frame.

The code writes itself fast, sure. But keeping the machine aimed is a full-time human job: the review gates, the CI/CD, the project scaffolding, the file where every mistake gets written down. I live a miniature version of that forward-deployed role today, and yeah, it still feels like babysitting.

It's the part that can't be automated, because automating it would just create more of itself. The mechanical work gets shaved away; what stays is someone to guide the machine and say "this is what we're building." It's also why benchmarks are misleading: they measure AI on problems we've already framed and scored. The frame is still ours.

Quoting John Lanchester's Squillions (via)

Since fewer and fewer people are using banknotes, it follows logically that fewer banknotes are needed, and therefore that fewer banknotes are being printed and put into circulation. Right? Wrong. In the UK, there is £1300 cash in circulation for every single one of us, but the amount of cash we actually hold is one seventh of that figure. The value of banknotes in circulation has been rising sharply for decades, and not just in the UK. In 2005, the total value of all the dollar bills in circulation was $759 billion. By 2015, it was $1.38 trillion. Last year, it hit $2.395 trillion. As Kenneth Rogoff put it in The Curse of Cash (2016), the dumbfounding thing is that ‘no one quite knows where exactly most of it lives or what it is used for.’ According to Oliver Bullough, in his alarming and unsettling book Everybody Loves Our Dollars, in 2022 the average American held $418 in cash, but there was $7357 of cash in circulation for every American man, woman and child.

Quoting Simon Willison's More than just code review (via)

Sometimes this involves reviewing every line of code they have written, but there are other ways to achieve that goal. Eyeballing every line of code has never been the most effective way to validate a change to a piece of software.

I recently bought a Bfour BF-30 as a budget BLE thermometer and it works great for slow cooking meat. But its Android app is a nightmare: trying to get your attention with all sort of different ways with apparently no way of turning it off and one feature I was deeply missing was some kind of trend in temperature to know how much time to expect until it hits temperature or if the temperature delta is slowing and I need to add extra coal or open the vents on my grill.

While waiting on the BBQ, I used opencode with DeepSeek-V4 to build a browser-based replacement. It turned out to be much more straight forward than I originally thought.

Quoting My AI Adoption Journey (via)

Such a strong piece from Mitchell Hashimoto about his experiences with GenAI and coding that hits quite close to home.

I quickly discovered for myself from first principles what others were already saying, but discovering it myself resulted in a stronger fundamental understanding.

1. Break down sessions into separate clear, actionable tasks. Don't try to "draw the owl" in one mega session.
2. For vague requests, split the work into separate planning vs. execution sessions.
3. If you give an agent a way to verify its work, it more often than not fixes its own mistakes and prevents regressions. More generally, I also found the edges of what agents - at the time - were good at, what they weren't good at, and for the tasks they were good at how to achieve the results I wanted.

Quoting Dario Amodei's Machines of Loving Grace (via)

I am not as confident that AI can address inequality and economic growth as I am that it can invent fundamental technologies, because technology has such obvious high returns to intelligence (including the ability to route around complexities and lack of data) whereas the economy involves a lot of constraints from humans, as well as a large dose of intrinsic complexity. I am somewhat skeptical that an AI could solve the famous “socialist calculation problem” and I don’t think governments will (or should) turn over their economic policy to such an entity, even if it could do so. There are also problems like how to convince people to take treatments that are effective but that they may be suspicious of.

I was intrigued by this assertion and had a somewhat long discussion with Gemini about why that's the case to help me build a mental model about the problem and it used a game analogy that was quite useful so trying to reproduce as much of that conversation as possible below.

Quoting Jason Gorman (via)

The hard part of computer programming isn't expressing what we want the machine to do in code. The hard part is turning human thinking - with all its wooliness and ambiguity and contradictions - into computational thinking that is logically precise and unambiguous, and that can then be expressed formally in the syntax of a programming language.

Quoting D. Richard Hipp (via):

But once we got that and got this aviation grade testing in place, the number of bugs just dropped to a trickle. Now we still do have bugs but the aviation grade testing allows us to move fast, which is important because in this business you either move fast or you're disrupted. So, we're able to make major changes to the structure of the code that we deliver and be confident that we're not breaking things because we had these intense tests. Probably half the time we spend is actually writing new tests, we're constantly writing new tests. And over the 17-year history, we have amassed a huge suite of tests which we run constantly.

Quoting Tired People Don’t Win:

Peak performers aren’t the ones who push the hardest. They’re the ones who stay unhurt long enough to keep showing up. Careers work the same way. You stay in the game so you get the chance to win.

Prediction: AI will make formal verification go mainstream (via source), which itself points to the author's blog.

Martin Kleppmann makes the case for formal verification languages (things like Dafny, Nagini, and Verus) to finally start achieving more mainstream usage. Code generated by LLMs can benefit enormously from more robust verification, and LLMs themselves make these notoriously difficult systems easier to work with.

More with less, or is it more with the same (via source)

How automatable the work already is. Where the work is rules based, high volume, and low variation, AI may replace labour in the same way classic automation has. Think claims processing, simple customer support, structured back office workflows. These functions already lived close to the automation frontier. AI just expands the frontier a bit. This will reduce headcount, but mostly in places where headcount has been under pressure for decades anyway.

The cost and consequences of mistakes. In many industries, the limiting factor is not productivity, but risk. Healthcare, aviation, finance, law. Increased throughput also increases the risk surface area. If AI increases the probability or cost of an error, you cannot shrink the team. You often need more human oversight, not less.

From @karpathy on X:

You will never be able to detect the use of AI in homework. ... You have to assume that any work done outside classroom has used AI. ... Therefore, the majority of grading has to shift to in-class work. Using the calculator as an example of a historically disruptive technology, school teaches you how to do all the basic math & arithmetic so that you can in principle do it by hand, even if calculators are pervasive and greatly speed up work in practical settings. In addition, you understand what it's doing for you, so should it give you a wrong answer (e.g. you mistyped "prompt"), you should be able to notice it, gut check it, verify it in some other way, etc. The verification ability is especially important in the case of AI, which is presently a lot more fallible in a great variety of ways compared to calculators

I had a bit of a weird experience this weekend. I was bogged down debugging a personal tool I’m building, a sync engine between Raindrop.io and my reMarkable tablet, when I stumbled into a realization about the future of our craft. It wasn't that the AI was a genius programmer; it was that it brought the marginal cost of a wrong guess down to zero. It turned the Infinite Monkey Theorem from a philosophical absurdity into a viable engineering strategy where the marginal cost of a wrong guess is down to zero.

My former colleague Rebecca Parsons, has been saying for a long time that hallucinations aren’t a bug of LLMs, they are a feature. Indeed they are the feature. All an LLM does is produce hallucinations, it’s just that we find some of them useful.

Found out about this interview with Steve Jobs that the master was supposedly lost for a long time and released in full when finally found. The long format is a sneak peak into such a different time, he was still with NeXT and at some points during the interview very salty about the fight that forced him to leave Apple. Being able to hear what was going on in his head and his thought process at that giving time is such a blessing.

The most profound shift in my workflow isn't a new productivity app or a fancy monitor. It's the realization that my primary productivity tool is now just a collection of ongoing conversations with LLMs. This is a form of context farming: cultivating multiple, independent project contexts that I can dip in and out of with zero ramp-up time. The cognitive tax on switching between wildly different tasks was a cost we all just accepted. Now, that tax has been effectively eliminated. True multitasking, the kind the experts told us for years was a myth, is now not only possible but shockingly effective. The baseline for productivity has been redrawn, and the tools we used just a couple of years ago feel like they're built for a different cognitive era.

Notes from Write to Escape Your Default Setting.

Let's call your mind's default setting 'perpetual approximation mode.' A business idea, a scrap of gossip, a trivial fact, a romantic interest, a shower argument to reconcile something long past. We spend more time mentally rehearsing activities than actually doing them. You can spend your entire life hopping among these shiny fragments without searching for underlying meaning until tragedy, chaos, or opportunity slaps you into awareness.

The modern web has a lot of power, yet I often find myself thinking about the 90s. Tools like Visual Basic, Delphi, and even Flash, didn't just let us build personal programs; they made it fun, mainly because they made starting so easy.

I read this article and it got me thinking that this problem of measuring ourselves by our wealth is not something new but it got worse with the internet and how many peers we have to compare ourselves with.

It all has to start from the first page, right?