sourdoughness

Using Opus 4.5 through VScode/CoPilot gives so much better results than anything else I’ve tried that I kept paying when they briefly made it 3x token rate.

I really like the interaction flows better than Gemini 3 or Codex, though I can’t quite quantify why. The amount of explanation/supporting material in Opus’s output feels just right to me.

Interesting. I treat VScode Copilot as a junior-ish pair programmer, and get really good results for function implementations. Walking it through the plan in smaller steps, noting that we’ll build up to the end state in advance ie. “first let’s implement attribute x, then we’ll add filtering for x later”, and explicitly using planning modes and prompts - these all allow me to go much faster, have good understanding of how the code works, and produce much higher quality (tests, documentation, commit messages) work.

I feel like, if a prompt for a function implementation doesn’t produce something reasonable, then it should be broken down further.

I don’t know how others define “vibe-coding”, but this feels like a lower-level approach. On the times I’ve tried automating more, letting the models run longer, I haven’t liked the results. I’m not interested in going more hands-free yet.

We have a low-tech version of something like this in South Australia: we pay the wholesale rate for electricity, which updates at 5 minute intervals. During the day when there’s oversupply of wind and solar, the rate is super low or even negative, which we take advantage of to charge an EV (and we’ll be adding a home battery soon).

The power company can integrate with car chargers and battery controllers to control all of this automatically, though we don’t bother - just check the app for the cheapest/greenest times and schedule the car to charge then.

It’s allowed us to switch to an EV without even really noticing any extra power cost for charging it.

We’re not talking about AI writing books about the systems, though. We’re talking about going from an undocumented codebase to a decently documented one, or one with 50% coverage going to 100%.

Those orgs that value high-quality documentation won’t have undocumented codebases to begin with.

And let’s face it, like writing code, writing docs does have a lot of repetitive, boring, boilerplate work, which I bet is exactly why it doesn’t get done. If an LLM is filling out your API schema docs, then you get to spend more time on the stuff that’s actually interesting.

I don’t think entropy ever moves from high to low overall, it only ever distills some local low out of an higher entropy area, and in doing so, the overall entropy increases.

It works a bit like air conditioning: yeah, you can make one room cold, but only by making more heat outside the room. The overall temperature of the system increases.

My understanding of this idea is that once the universe reaches a state of maximum entropy (this is the “heath death” of the universe, where everything is a uniform, undifferentiated cloud of photons, then time stops being meaningful because there can be no change from moment to moment. In a sense, time _is_ the change from low to high entropy - if you don’t have any entropy gradient, you can’t have any time either.

The “resetable” aspect seems to be a crucial part of true “time-lock” encryption… rather than what feels like a “proof of work” mode where an amount of computation stands in as a proxy for “time elapsed”. But regardless of how good a proxy it is, “time-elapsed” is not really what we want - we don’t want a lock that takes “at least n seconds of effort to open”, we want a lock that will not open before a timestamp.

It feels like a true time-lock solution should be impossible to unlock before a particular date, but trivial afterwards, rather than assuming the opener started work on it at the moment it was released and has been burning cycles at some maximal rate since then.

I don’t think the universe contains unfakeable timestamps, which seem to be a requirement for that true solution - it feels like it’s not compatible with relativity, so maybe proxies are the best we can do on a fundamental level.

You can turn watch history off, and then you don’t get any recommendations at all - just the channels you subscribe to. YouTube doesn’t like it - and goes out of its way to make the apps feel a bit broken with out it, but I much prefer to curate my own list of channels without an algorithm trying to keep me watching longer than I mean to.

Yeah, that’s true - it’s integral because an EV retains all of its components throughout its lifespan.

If we consider an ICE car’s fuel as a “component”, then it’s an interesting comparison: fuel is basically maximally degraded - it accumulates as much “wear” as possible - and then it’s jettisoned so what remains attached to the vehicle is comparatively less worn.

Driving any car like that will result in increased stress on components and wear. A lot of EVs have high-end sports-car levels of acceleration, and those aren’t known for being low-maintenance.

I don’t agree that they’re just props: they’re analog controllers and so the form factor and manner of their use directly affects the music being played. Even the controllers that are just “big heavy knobs on bearings” are fundamentally different to a typical knob on a synthesiser in terms of the signal they produce and the sound that ends up being made.

Vacuum clamping would get tricky when the surface you’re trying to slurp onto is also the thing you’re actively cutting holes in, right?

Turning CO2 into a stable feed-stock seems to be an important part of it, given that as the article says, we need to be processing gigatons per year.

If we’re just using our captured CO2 to extract more fossil fuels to burn, thats not nearly as big a reduction in atmospheric CO2.

That is an extremely interesting and potent way of phrasing it - thank you!

Is adding side-panels to skyscrapers a solved problem? It doesn’t seem nearly as easy or mature as rooftop or grid-scale PV.

I know what you mean, but not strictly true. I cook everything from scratch - the most processed ingredient I use is probably a can of whole tomatoes or occasionally a pack of pasta. I’m a good cook, I think, and it’s easy for me to make our meals, but I do like things to taste good, and I no doubt add more butter, olive oil and cheese to things than is strictly healthy.

Agreed! This stood out:

> The laws of physics definitely allow for functional subluminal warp drives

“Definitely”? That feels like an exceedingly bold claim, given the nature of the thing we’re talking about. I don’t believe we know enough about “functional” machines comprised of multiple Jupiter-masses of exotic-if-not-quite-fictional matter to start using words like “definitely”.

The new material is an extremely thin film that sits on top of a traditional glass substrate: it doesn’t replace the window, and it doesn’t dim light or reduce the spectrum in the same way that frosted glass normally would.

This is the question I come back to: a crucial advantage of plastic is that nothing eats it. What’s the missing step that allows things to eat it, but only when we’re ready for them to?

Start at noon? That’s the exact opposite of the approach people take when it’s hot around here (I live in southern Australia where it regularly breaks 40C in summer). Start early - before the sun’s up if you can - and be ready to knock off soon after lunch.

I can vouch for this. You can tell how much YouTube doesn’t want you to do this by the way the app shows you a big, error-looking message on boot: “Your watch history is off - update setting?”. No thanks!

Only real downside (and it might be deliberate) is that the app will quickly forget the current video you’re watching. Still worth the trade off.

The way this is implemented, it’s polish over a checkbox, with a fallback to the checkbox where it’s not supported. The semantics has therefore got to be the same as a checkbox - that is, no (native) submission of “off” switches.

For what it’s worth, I have the same expectation and have been applying it to forms since IOS 3 or 4.

If we’re going to have a switch-looking thing, it’s because it brings some skeuomorphic affordance with it: I think “off/on with immediate effect” seems like exactly the thing it should bring to differentiate from a checkbox.

He’s got a deliberately obnoxious delivery, but by all accounts his research is solid.

The politician in question was the deputy premier of NSW (one of the highest state-level positions in the country) and stepped down amidst the - totally justified - stink Jordies caused.

It’s funny - the test rig in the article shows wraparound LED displays, which are already replacing a lot of green screen applications: the scene is rendered in real-time on those displays, taking the position and angle of the camera as input to get the perspective correct: there’s no masking to do, because the scene just “looks correct” in camera.

The Mandalorian was a high-profile early use of the tech. It’s pretty impressive!

You should go and read the extremely detailed posts Christian shared in r/apolloapp. The dude really considered everything, but it’s a non-starter: Reddit do not want 3rd-party clients, so it’d be foolish to build further on top of this platform even if it were practical, which it absolutely is not.

Bots aren’t people and can’t consent, sure - but they are tools that are wielded or deployed by people who absolutely can consent (setting aside whether click-wrap terms are enforceable or not). If I throw a brick through a window, it’s me in the shit, not the brick.

These kind of articles are what we get when writers are given impossible deadlines and targets to hit (multiple submissions per day in many cases), and incentives that favour clicks and engagement over actual imparting of information.

Even if you’re a good writer, you’re not doing your best work under those conditions.

It’s not removed for long though, is it? It’s at best offset against a ton of new CO2 coming up from underground, which might be valuable, but I don’t think it’s fair to describe this as CO2 removal.

I have one, and it’s an excellent machine (15 years old, no sign of wearing out), but the dryer is fairly lacklustre. It can only dry about half a full washing load, but I suspect that’s just physics: however much a machine can dry, it can wash more.

We air-dry everything anyway, so it’s not a problem for us, but I wouldn’t want one somewhere that relied on machine-drying a lot.