It's being the tech lead of a team of junior to mid level developers. You design roughly what the solution should look like, split it into reasonable sized tasks so they don't go off the deep end, advise them on some of the details, then assign them the tasks and let them get on with it, keeping an eye on what they're doing, reviewing their output, and course correcting them when they go wrong.

Just like with a team of humans, you have to use your judgement as to how much supervision they need individually and how large a task you can give them without them going off the rails.

I've tried, and I feel like I've got closer with faster models, but ultimately the agentic loop excludes you. Even if you're asking the agent to do simple short tasks, it's still: prompt, wait, wait, wait, check, and you never really feel like you're the one in control.

The problem with faster models is also that they're more stupid, so that additionally breaks your flow when you have to fix something dumb it's done.

LLM-powered autocomplete is a bit more like it, but that tends to be either so dumb as to be a net negative, or slow enough to be useless. And autocomplete is pretty distracting for me.

I feel like I'm missing a mode that works more like a pair programmer. Perhaps a multimodal model that can talk to you about what you're writing, as you write it, and offer suggestions rather than trying to take over and do everything for you.

Surely if that's the case then when you move your finger to the upper left then the pointer should move to the bottom right. Because that's how it would work if it was a real object and you were pushing the pointer around with your finger. Why is scrolling a special case?

Honestly though, I wouldn't mind that much if Apple hadn't decided to call it "natural" scrolling, like you're weird if you prefer up for scroll up and down for scroll down. It's both smug and reeks of the same kinda of discriminatory attitude that made life hard for left handers.

I use a local LLM with it, but you can use a hosted LLM if you like.

The core home automation stuff can run on a potato. The LLM just writes new automations when I ask it, or acts as a natural language interface.

I use a pretty small 4B parameter local LLM, on a fairly modest mini PC. It doesn't take a frontier model to do that kind of work.

Remove the fuel elements, reprocess what's useful, and store the reprocessed materials and nuclear waste somewhere "temporarily" that isn't really suitable for long-term storage.

Remove intermediate and low level waste from site and also store it "temporarily".

Remove any non-contaminated plant and sell for scrap.

Punt the main part of the problem (scrapping the main reactors and reactor buildings) down the road for a hundred years or so until radiation levels are acceptable for demolition to proceed.

Re-use other parts of the site for projects that can use the existing HV connection, like another reactor, or battery storage.

That's essentially all you can do unless you want to risk a radiological accident.

There are a few things MacOS X inherited from classic MacOS that I don't think work that well in the modern world, and application-focused task switching is one of them. It made sense in the classic Mac context where many apps used floating windows for toolboxes and other non-document windows. You wanted to switch the whole application, with all of its windows, as a unit. It was also the right technical decision with classic MacOS's modest multitasking abilities.

But the world has since mostly standardised on SDI app design with tools contained within that window, and multiple windows representing different documents. In that context, the macOS app-then-window approach is more roundabout than pure window switching. You get used to it, but when you've got a lot of windows open, it's a small but ever-present drag on usability.

Alt-Tab is one of the first things I install on a new Mac. Hopefully one day Apple will give us a built-in option, much like they eventually did with window tiling and full-screen window zooming.

You can run x86 macOS VMs in Windows or Linux too with a little bit of technical trickery, but again, you end up with a license issue, so no-one reputable does it.

"Momentarily" can mean years or even decades, and millions of people can suffer or die as a result.

Being able to limit the influence of external bad actors is the main goal of ID verification. Age verification is a useful side effect that makes it easier to sell to the general public.

Big Tech has had at least a decade to fix this, did nothing of note, and is all out of ideas. Privacy advocates had the same time to figure out a "least bad" technical solution, but got so obsessed with railing against it happening at all, that nothing got any traction.

So governments are here to legislate, for better or worse. They know it's a trade-off between being undermined by external forces vs. the systems being abused by future governments, but their take is that a future authoritarian government will end up implementing something similar anyway.

IEEE 754-2008 defines decimal floating point arithmetic that is compatible with COBOL and is usually implemented using the Intel Decimal Floating Point Math Library on commodity hardware.

For a typical core banking ledger application, the performance cost of a software implementation of DFP (vs. having DFP hardware instructions) is pretty low, and greatly outweighed by the benefits of being able to use commodity hardware and more maintainable languages.

The OS core is deployed as a single unit and is a few GB in size, pretty small when internal storage is into the hundreds of GB.

For sure, you can squeeze a few percentage points more out if you optimize for TBDR, and there are some edge cases where it's possible to make TBDR architectures behave pathologically, but it's not that big a deal in the real world.

I also disagree that the Steam Frame is for streaming primarily. If it was, why put such a powerful SoC in it or using it as the prototype device for doing x86 emulation with Fex?

The Adreno 750 is a 3 TFlops GPU that _should be_ substantially faster than a PS4 or a Steam Deck. It'll play plenty of low-end PCVR games pretty well on its own, if Fex's x86 emulation is performant, which it is.

Like the Meta Quest 2, it's a crossover device that a lot of people will just use standalone.

There is also a Vulkan driver for the M1/M2 GPU already, used in Asahi Linux. There's nothing special about Apple's GPU that makes writing a Vulkan driver for it especially hard. Apple chooses to provide a Metal driver only for its own reasons, but they're not really technical.

My M4 MacBook Pro loads a wide range of apps - including many that have no Chromium code at all in them - noticeably slower than exactly the same app on a 4 year old Ryzen laptop running Linux, despite being approximately twice as fast at running single-threaded code, having a faster SSD, and maybe 5x the memory bandwidth.

Once they're loaded they're fine, so it's not a big deal for the day to day, but if you swap between systems regularly it does give macOS the impression of being slow and lumbering.

Disabling Gatekeeper helps but even then it's still slower. Is it APFS, the macOS I/O system, the dynamic linker, the virtual memory system, or something else? I dunno. One of these days it'll bother me enough to run some tests.

It is, however, complete nonsense, and the next few years of failed promises on AGI will eventually bring people to their senses, if a market crash and sustained economic depression doesn't do that first. It would be funny if it wasn't going to cause suffering for millions of people, whether we succeed at AGI or not.

I _like_ AI, I find LLMs and many other aspects of useful, and I am optimistic for the long term prospects of AI. But the rush to try and get to AGI is completely out of control at this point, and the fallout from when the bubble pops will set AI, and our societies, back a long time.

Maybe this will get better with time, but we're half a decade into the Matter era and the end-user experience is _worse_ than with ZigBee. In that sense, Matter has failed.

Ollama, for example, will let you run any available model on just about any hardware. But using the CPU alone is _much_ slower than running it on any reasonable GPU, and obviously CPU performance varies massively too.

You can even run models that are bigger than available RAM too, but performance will be terrible.

The ideal case is to have a fast GPU and run a model that fits entirely within the GPU's memory. In these cases you might measure the model's processing speed in tens of tokens per second.

As the idealness decreases, the processing speed decreases. On a CPU only with a model that fits in RAM, you'd be maxing out in the low single digit tokens per second, and on lower performance hardware, you start talking about seconds over token instead. If the model does not fit in RAM, then the measurement is minutes per token.

For most people, their minimum acceptable performance level is in the double digit tokens per second range, which is why people optimize for that with high-end GPUs with as much memory as possible, and choose models that fit inside the GPU's RAM. But in theory you can run large models on a potato, if you're prepared to wait until next week for an answer.

Closer to the hardware, more control, fewer workarounds because the driver is doing something "clever" hidden behind the scenes. The tradeoff is greater complexity.

Mere mortals are supposed to use a game engine, or a scene graph library (e.g. VulkanSceneGraph), or stick with OpenGL for now.

The long-term future for OpenGL is to be implemented on top of Vulkan (specifically the Mesa Zink driver that the blog post author is the main developer of).