Came here to say this. Also probably fairly beefy GPUs and other enterprise grade hardware, in someone's backyard. If this actually happens, it'll be interesting to observe the patterns as theft expands from people who actually know what they're after to meth-heads who figure there must be a bunch of copper in there and then are stuck trying to figure out where to fence RAM.

Terrible idea, I think. Big datacenters are all about security and economy of scale, this has neither.

I've had more of the opposite problem with eBay. Selling there kind of sucks, unless you're selling brand-new Amazon style items and prepared to offer a full warranty and free returns.

Selling anything used or asâ"is is a shit show, because idiots will just click "buy now" on mobile without reading the description or condition, then return it when it's not new/sealed, and assholes will buy things like TV parts "sold as-is" then return it after they find out that the power supply is bad and now they blew the ass out of the mainboard they bought to try it.

eBay offers little recourse in these cases. Sometimes they'll refund the return shipping. You can ban buyers, but nothing stops them just making a new account, especially now that buyer feedback is essentially meaningless. The bright side is, the seller has the buyer's address for.. let's just go with "their records."

So like American "democracy," it's shit, but most of the alternatives are worse.

Not sure what Ghostly is, but according to the first link there:

Ghostty is a fast, feature-rich, and cross-platform terminal emulator that uses platform-native UI and GPU acceleration.

Yeah, this. All of the AI platforms have been operating at insane losses -- many estimates say $25-30 burned for every $1 in revenue, some say lots more, and that's without all the externalized costs.

It was never sustainable, and it will never be sustainable barring some tremendous breakthrough in efficiency of compute, or in cost of power generation. And I'm not talking about the ridiculous data-centers-in-space thing. Energy may be nearly free up there, but we have a hard enough time with the waste heat on Earth with wind and water, what are we going to do in space? Asteroid-sized radiators?

It's more of an arms race to determine the one or two players that will survive the coming meltdown, then they'll be forced to start charging what these things actually cost to run, and the average end-user will be priced out of the market. The "free" stuff like Google/Gmail AI summaries will probably stick around, just for the value the providers get in hoovering up all your data for training, though they'll obviously start inserting ads and product placement in them to a greater degree than happens already.

And lest you think that buying an annual plan actually means getting that plan for the duration of its term... remember that we live in a free society where the Epstein class is allowed to change the terms of a "sale" unilaterally. At least they're offering a prorated conversion of the annual plans into the new-and-improved pay as you go plans, which might become somewhat attractive as the multipliers scale TO THE MOON and the annual plan becomes essentially worthless.

Oh boy did that dredge up some fun memories from an embedded-Linux project I was working on a while back. Like this gem here:

Often, but not always.

More than once I've bought from an eBay seller after finding similar items cheaper on Amazon, because they claimed their product "original" or "genuine" or something, where the Amazon one is an obvious Chinese knockoff.

Then the package arrives. Dropshipped from Amazon, the same item I passed up before. So I open an eBay return and send it back at the seller's expense, then buy the knockoff from Amazon.

Yeah, great benefits. When you send an image with RCS, instead of a boring gallery picker you get a blurry preview and a silly banana icon offering to "remix" it for you.

Also, Google's implementation of RCS is shit, it breaks randomly and just starts blackholing messages because your phone fails some element of the 215-point Authentication process and gets disconnected from the servers, which keep telling all your correspondents that you're connected and available and by all means don't fallback to SMS. Pretty lame that the only working implementation of RCS appears to be Apple's.

Sure, it's "standard" but it's damn near impossible to implement without buyin from carriers, device makers, etc. Someone should have strangled this teratoma in its crib before it could stain the landscape at all. Signal ftw.

Well, I've got a Tesla too, and I'm not at all hesitant to complain about its assorted flaws (100+ years of automotive engineering focused on keeping the water and snow OUTSIDE the car? bah!), and I wouldn't go so far as to say FSD works "great," but it's acceptable for what I paid for it. Late-term fire sale prices before they discontinued outright sales.

I've seen crazy bugs. Mine got drunk for a while. Turn on FSD on the highway, it goes fine for a mile, then starts drifting to the left, hugging the lane line. Does that for a while, then crosses the line, wobbles around a bit, and goes totally batshit -- sometimes veering rather sharply at the median. All the while displaying on the visualization screen where the car is and where the lines are -- the computer KNOWS it's driving drunk, and keeps doing it. Nothing in the diagnostics or event logs to explain it.

It's an older one, so it's still running v12, and while Tesla has said they'll be doing a "14 Lite" backport, and an upgrade path to newer hardware once it's "done," I'm not holding my breath. 12 doesn't do as many parlor tricks as 14, like backing out of my driveway or finding a parking spot for me. But it drives from the road in front of my house to the parking lot entrance at Costco, or from city to city, with little to no intervention. It would, in general, have no trouble driving from a major road to another one 50 miles away with no one in the seat, if it were allowed to. And if it hadn't guzzled a bottle of Wild Turkey first.

I've driven newer models with v14, and I believe they are technically capable of making the journey from parking-to-parking unattended, IF nothing unexpected happens, IF it's not pouring rain or pea-soup fog or blizzard. Doing it safely, maybe not yet, which is why they still require a butt in the seat and eyes on the road (they have a camera to track that), said butt being intended to prevent a tragedy or be liable for one if it happens anyway. That's kind of the big sticking point, I think -- if my car hits a pedestrian while I'm not in it, who is legally liable? Also, it's limited by charge, since there's no infrastructure in place or on the horizon for automated charging. That'd be handy even for manned trips, if I didn't have to get out in the rain to plug in a charger. But then autonomous charging might be how we get Skynet. Do you want Skynet?

I'm still waiting for them to build "banish." They have Summon, and it (usually) works. 14 has the ability to find a spot and park in it when I'm in the seat -- it isn't always the spot I would choose, but if I don't have to walk, I don't care. The pieces are there to enable me to get out of the car in front of the store and tap "park" and let the computer deal with it, then Summon it back when I'm done. I think a lot of people are waiting for that. Seems an easier lift than coast-to-coast unattended.

Lately I find myself contemplating the differences in tone between the original 99 Luftballons and the English translation. The original seemed more focused on human failings, paranoia and bravado, the English version blames the computer.

They should have used more relatable units, like "as heavy as 12 Teslas" (bonus for name checking actual somewhat-related yet totally inappropriate unit of measurement!) or "a whopping 1/18th of the International Space Station." Really, who goes around weighing humpback whales these days?

As for the cables, clearly "the thickness of a standard woodchuck" would be far better, or perhaps for the Florida audience, "Burmese-Python sized". Unless of course they've developed some technology where they just pour the energy into a cantaloupe-sized sphere and it emerges elsewhere. That'd be pretty cool.

Apparently it's a dazzling 320583 smoots in length, too. That's nearly 6000 football fields!

I've lately been using Copilot in VSCode to build a custom CRM/ticket management app to replace an ancient thing that was built in the early 2000s, in COBOL, and has various conniptions running on anything newer than XP.

I'd say the agent has autonomously written.. probably 95% of the code for a FastAPI backend and Vue.js frontend. Mostly with the model selection set on Auto. It seems to mainly use Codex and Claude Sonnet, delegating lighter tasks to Haiku. I can manually select some of the free models if I know a task is easy.

Frankly, I'm pretty amazed and slightly disturbed by how "smart" it seems -- "Here's a table schema. Make some API endpoints to get, put and post this." [working ... done] "OK, that mostly works, but (some constraint that maybe wasn't obvious from the schema)" [working.... done] "Nice. Now make a UI to display existing rows in a tabular format" [...] "Add an edit button to each row that opens a modal dialog to edit those columns" [...] "I don't like that layout, move the comments field to the bottom and put the two phone numbers side by side." [working... done] "Huh, cool. It'd be nice to have edit history. Alter the schema to include system versioning, then create some endpoints to fetch historical data." [working for quite a while... done] "Now present that in a nice UI." [working.... Here's a page that shows a reverse-chronological history of the selected row, including which columns where changed, from what to what, when, and by whom.]

It even came up with some SQL constructs that were better than what I would have done by hand (I'm competent but by no means guru-level in SQL).

Had less luck using it to do harder stuff like embedded microcontroller projects. It's all dependent on how commonplace is what you're doing -- FastAPI and Vue are everywhere, it's good at that. Arduino, pretty good. ESP-IDF or STM32Cube, less so. Needs a lot more handholding.

Would I trust the output if I didn't know how to read and critique the generated code? Absolutely not. Or maybe I would, if I were at the right point on the Dunning-Kruger graph to not even know what I don't know. Does it do stupid stuff occasionally, like implement an API endpoint correctly but forget to require authentication on it, even though all the other ones in the same file have authentication? You bet. It reimplements functions that are already implemented in shared includes, then later notices that it did that and offers to consolidate them.

Gotta watch the context windows, break stuff down into manageable chunks, generate an action plan and then execute it in steps. Gotta keep everything in manageable commits, so if it does something completely boneheaded and breaking, you can roll it back and have another go. (the VSCode UI also has some functionality for this, showing every agent change as a color-coded diff for review)

It saves a hell of a lot of time doing the drudgery of cookie-cutter HTML/CSS, and it's pretty decent with the backend stuff.

I'm using the minimum paid Copilot Pro plan, and used about 65% of one month's premium requests on this project (and a few brief digressions). I think I've gotten more value out of that than I paid for the whole year, and I'll no doubt be continuing to use it to work on little projects that have been languishing forever because I just don't have the time.

Not a lot of great solutions. Reducing the power a lot is probably the best one. Then the car needs a receiver near each wheel to get the data. Some already have this, that's how they magically figure out which tire ended up where when you do a rotation. Randomizing the IDs could work in this scenario, too, but then the car has to relearn the IDs every time they rotate, which could be error prone.

Anything involving fancy crypto is going to take a bite out of the sensor battery life, and require them to be replaced more often. AirTags do this, they randomize their BT addresses and transmit encrypted packets. And their batteries last a year or so. Could put a bigger battery in a TPMS, but the more weight, the harder to balance, etc...

This mode of tracking isn't really a threat that average stalkers are going to execute, but it's definitely something state actors and Big Surveillance could exploit. Flocks have a bunch of extra radios in them they're not using most of the time, they could easily sniff Bluetooth, and could be adapted by hardware (or possibly software) modification to intercept the 433/315 MHz signals often used by non-BT TPMS.