I was musing about this recently with the recent models becoming more capable. The core of gen AI is the model, which is trained on a massive dataset. To date, gen AI has improved because the models have become larger, more efficient, the data they are trained on has become better and the software/harnesses around them has improved to help query them.

As I see it, surely the bottleneck will soon become the data they are trained on? If we imagine a scenario where a models could consume an infinite amount of training data, and there is no limit to the training time or quality. The sum of human skill/knowledge is the limiting factor. Gen AI should (in theory) never be able to out preform or push the boundary of the sum of humanity at time of training.

Or, counterpoint, is there enough randomness and speed to iterate that gen AI can actually step change and improve if training times/cost were less prohibitive? Most companies/models today will save good output and feed it back into the next iteration, but right now that's taking months. What if that took minutes?

What do you think?

Is gen AI going to take us to general intelligence?
Will gen AI get to a place where it's "intelligence" and reasoning is actually better than the sum of Humanity?

A few days ago, I posted this and quickly realized that the world of data backups is far richer than just sudo rsync -av --delete --exclude=Videos /home /home_bkup.

So now I'm window shopping the top Linux-supported backup solutions: borg, duplicacy, kopia, restic and--oh look--a core borg dev just dropped his own new-and-improved solution, vykar.

Restic was the first tool I started to research, and I thought I really liked it, got as far as installing, initializing a test repo, creating a couple of snapshots. But restic seems to be, hmm, fussy about the source and destination paths, absolute vs relative paths, etc.

The fact that merely renaming a parent directory (or grandparent, or great-grandparent, etc) causes restic to treat every unchanged byte below that as brand new ... that's a recipe for giant, bloated repos, and it's unacceptable to me ... and hey, lookit that, borg does not do that. So now, restic is out and borg is in.

But what other pros v cons are there, that I haven't even realized need to be considered? What advantages/disadvantages do other apps offer? Which ones can I easily automate with nightly/hourly cron jobs? Which ones have their own even-better automated solutions?

Do I even want encryption? All of my drives/volumes are LUKS encrypted, and anything I would store remotely would also get encrypted before it ever left my LAN ... plus, I'm just a bit nervous about having the backups encrypted, requiring working, functional software to restore/recover data from them....

That may not seem like such a big concern, perhaps, but I am currently working my way thru decrypting a bunch of 10-15 year old TrueCrypt-ed volumes, which requires using an old, outdated version of VeraCrypt and a somewhat "cross-my-fingers" effort to find KeePass repos old enough (also outdated, KeePass 1.0 repos) to still contain the various passwords I used to encrypt those ancient volumes ... but also still use new enough master passwords that I can still get the KeePass repos unlocked.

With rsync, I can literally just go into any backup, find the specific version of the specific file(s) I want to recover, and manually copy it back to my workspace. Is anything like that option available in any of these deduplicated/encrypted solutions, even if they're not encrypted? If (eg) a borg repo is created w/o encryption, the data is still all just borg-specific blobs, right? Or can I navigate into the repo and just manually grab files?

Oh yeah ... for reference, the past 10-ish years, my backup routine has been to create a new, dated, destination folder, starting with a full backup of my /home folder (excluding things like Videos, Music, VMs, other bulky stuff that gets backed up separately/differently), and then running nightly diff backups into the same folder, while also maintaining a "one-day-older" second backup of the whole thing on a 2nd HDD ... then, every 3-6 months, zipping up the current backup folder and starting a new one.

At any rate, there you go; that's the kind of stuff I'm thinking about now, as I overhaul my 20-year-old, 20TB (but could be 2TB) backup system.

Any and all feedback, recommendations, tips are welcome. Danke.

This is a recurring post to discuss programming or other technical projects that we've been working on. Tell us about one of your recent projects, either at work or personal projects. What's interesting about it? Are you having trouble with anything?

I'm running a nixos linux machine with Hyperland as my window manager and a few month back (likely after an update) I noticed that firefox started showing a emoji picker when I pressed ctrl+.. This was a bit annoying since the firefox extension for my password manager is activated by that key shortcut. I figured this was some update for firefox, but now that I dug into it to fix it it turns out that it is a gtk thing that apparently each app has to opt out of! I could disable it by flipping widget.gtk.native-emoji-dialog in about:config, but this seems like a really bad choice by gtk. Two gripes with this:

1. Them adding a global keyboard shortcut for all gtk apps that is ON by default (for a kind of niche usecase).
2. Overriding shortcuts on a desktop wide basis with no meaningful (afaict) way to disable it.

Anyone knows if this is intentional? Maybe it's already been reverted upstream and I just need to update... anyway end rant!

I have a friend that is lucky enough to have retired at 40. A year ago he was adamant he'd never work again, having been burnt out from his time at big tech. Back then he was also an absolute AI hater and wouldn't listen to anyone who claimed LLMs were useful for programming.

He finally tried LLMs when Claude Opus 4.6 released and immediately changed his mind in the face of the overwhelming evidence that LLMs can in fact program pretty well. And now with the release of Fable 5 he's giddily creating all sorts of things that would have taken far too long to make prior to AI-accelerated software development. He actually plans to try and found his own business now. He's a very smart guy, so I hope he can make something interesting that people want.

There are a lot of AI doomers and haters. In person I mostly see people doing the same thing they've always done, but now saving time on various tasks. But this is the first time I've seen someone go from grumpy and checked out to giddy and optimistic thanks to LLMs.

My daughter (11) is doing a week long Python class, which is not using LLMs.

It got me thinking about how I learned to program in the pre-internet days (laboriously, from books), and then what a marvel it was when you could just search for information, especially for troubleshooting. But for her, the first answer in the Google search is going to be the AI summary, and most of her search tools are going to be AI tools.

I wonder if it would be possible to make an LLM that has a didactic/socratic mode. So if you said, "help me write a program to do madlibs" maybe it would give you a skeleton of a function, then prompt you to come to with a plan, then critique that plan. Or if you said, "I'm getting this error", it wouldn't just fix it, it would explain what the error means and nudge you towards the answer.

Thinking in a larger sense, it could have a rubric of important concepts, even tiers of understanding. It could be using the interactions to track the user's understanding, which could let it then tune how it answers future questions, or even be used to customize assignments.

I recognize that this is potentially replacing a teacher with a machine, which wouldn't be my goal. Good teachers are more holistic in their teaching than a machine is ever likely to be. But for people who don't have access to good teachers, or need more directed support than is available from a teacher, or just want to self study, it seems like it could be a valuable addition.

Until they solve the obsequiousness problem, it would be vulnerable to prompt hacking, so really more of a tool for someone who recognizes the value of learning over just being given the answer.

What do folks think about using such a tool? What would you want it to do, or not do?

Aside: I forgot until I reached the end of this post, but this is also (somewhat) the plot of The Diamond Age, or A Young Lady's Illustrates Primer by Neal Stephenson.

In a moment of distraction, I fell for a phishing phone call and compromised my Google account. It took me 13 minutes to realize how catastrophically stupid I am and begin frantically changing passwords. I've run the official Google "secure your account" process probably 10 times (though 9 of those times there was nothing to do). I've checked all my financial info, changed passwords on all sorts of things. As far as I can tell, other than gaining access to my Gmail, I don't think anything else was compromised.

How boned am I? I've got 2FA on basically anything remotely important, and I've had decent password hygiene (although I do use the Google password manager, so that's probably comprimised). Is there something else I should do or be on the lookout for?

This is a recurring post to discuss programming or other technical projects that we've been working on. Tell us about one of your recent projects, either at work or personal projects. What's interesting about it? Are you having trouble with anything?

Much to my chagrin, the company I work for has done a lot in terms of steering/ pushing all software development be done through AI for some time now.

And what gives me much grin, GitHub changed their pricing structure for copilot. I'll skip the details the key fact is what used to be about $30/month per person + maybe few bucks in overages is now resulting in us hitting our usage cap on the 2nd day of the month. Overage costs this month will be hundreds of dollars per developer. I know this is an unexpected expense as I mentioned it casually to our CTO who had no idea.

I'm curious if this is going to force them to rethink the AI strategy. The incessant pushing to use more and more AI maybe will finally bite them on the ass so much they have to ask us to stop or pull back? Or maybe they'll just plunder our salaries, who knows.

I'm curious if anyone else is in the same situation.

This is a recurring post to discuss programming or other technical projects that we've been working on. Tell us about one of your recent projects, either at work or personal projects. What's interesting about it? Are you having trouble with anything?

This is a recurring post to discuss programming or other technical projects that we've been working on. Tell us about one of your recent projects, either at work or personal projects. What's interesting about it? Are you having trouble with anything?

I'm a web developer and find the playwright MCP to be genuinely useful. My LLM is able to navigate my site, measure the size of elements, see console errors, network requests, etc. This is the only MCP I've ever installed and haven't yet had any cause to use others. But I'm interested in hearing what other professionals are using.

This is a recurring post to discuss programming or other technical projects that we've been working on. Tell us about one of your recent projects, either at work or personal projects. What's interesting about it? Are you having trouble with anything?

Got invited to Tildes tangentially through a server hosting project but realized this space might be outside the standard tech echo-chambers (techo-chambers?). Am I the only XCP-ng/Xen user in a sea of Proxmox/KVM? Any odd homelab setups to share?

This is a recurring post to discuss programming or other technical projects that we've been working on. Tell us about one of your recent projects, either at work or personal projects. What's interesting about it? Are you having trouble with anything?

This is a recurring post to discuss programming or other technical projects that we've been working on. Tell us about one of your recent projects, either at work or personal projects. What's interesting about it? Are you having trouble with anything?

This weekend, I've fulfilled a long dream of mine and upgraded my computer to a fully custom waterloop. This is, for a number of a reasons, a complicated process, and outside of general advice, it's difficult to provide an exact guide on how to do this. Custom waterloops are, well, custom. They depend mostly on what computer case you use, and what sort of reservoir and such you've bought. As such, my advice can also only be general.

Plan ahead

Check online for custom watercooling builds in your case. Use those as a guide for radiator and reservoir placement. Sketch ideas out on paper. Measure out the places inside of your case where you intend to place components. Check the your pc case manual, those very often contain info on where you can place radiators and reservoirs.

Some cases are ill-suited for custom waterloops. Consider buying a new case rather than building in an old, ill-suited one. It will save you a lot of pain.

Some cases require modifications. I had to cut into mine with a metal saw to make space for a radiator. Minimal material was removed from the frame, invisible after the case is put back together. I also had to drill into it to place the reservoir. The holes case manufacturers place for reservoirs are best-effort guesses. Unlike for fans, and thus radiators, there are no standards for reservoirs.

Do not rush

Expect a marathon, not a sprint. There will be setbacks. My process, setbacks and all, took me 3 days. And I still fucked up assembling my GPU. The die has bad contact and I'll have to drain the loop, pull it out, disassemble it and put it back together again. A lot of this was also spent waiting for next-day deliveries to show up because I'm dumb and was missing things.

Prefer soft tubing, and do not go for PVC

There are no performance benefits to hardline tubes, and they are a bitch to measure, bend and cut precisely. That 95° angle that was meant to be a 90° is going to be evident immediately, and forever. Soft tubes are forgiving, easy to put into the system and much more time efficient. They also do not require extra equipment dedicated solely to bending hardline tubes. Think about where your build is going to sit. On the floor? Who gives a shit how it looks?

Edit: if you opt for soft tubes, don't get them made from PVC. They very often leak plasticizers into the loop that gunk up your components and reduce performance. You'll need to swap them after some time and they also tend to harden. To minimize maintenance, go for EPDM tubing. It's black instead of transparent, but it's better in the long term.

Custom waterloops are all about you, and if you insist, then you do you. Hardline tubes are the endboss of all pc builds. Be ready for a challenge. Conversely, some folks really want transparent soft tubes. Just know what you're getting into.

Tube sizes

The standard soft tube is 13 mm outer diameter and 10 mm inner diameter, or 13/10. There's a ton of other sizes as well but remember even if the inner diameter is larger, liquid flow improvemets are going to be marginal. Different sizes also need different fittings.

Respect the crink

Soft tubing is a breeze to put into your system, but don't make those corners too tight or it'll crink and cut off flow. Check this especially when you close up the panels of your pc case. Tubing is cheap, comparatively. Don't be afraid to use more than you need.

Money

Custom water loops are pricey. Full copper radiators start at 100€, water blocks are usually hundreds as well, with the tubing, fittings and all it's normal for cooling equipment alone to account for a grand. You're bolting an aftermarket cooling system onto your PC that will turn it into a racecar. A lot of that is finely machined copper. It costs.

Remember the extras

Ya know how I said that I needed to order some extra things last-minute? Thank fuck for Amazon and their fast deliveries. If you live in a larger city, there's also a good chance a specialist computer store somewhere might have what you need.

Leak tester

Those are small air pumps with a pressure gauge. You close of all ports and then pump air into the system, 0.5 bar maximum (!!!), and wait 60 minutes. If the pressure is maintained, congratulations, your system is air- and thus watertight.

Test your individual components before you put them into your case! This way, you know that the components themselves are tight, and you avoid having to pull out a radiator after screwing it in place because you forgot to tighten that one end cap you can now no longer reach. Ask me how I know.

Also test our loop when it's fully assembled. Should you have a leak, divide your loop into two halves and leak test those. Repeat (divide into halves and test) until you have located the leak. If you have tested your comps individually before, it's going to be a radiator fitting you forgot to tighten or your reservoir top 99% of the time. Have a book or a podcast ready because this is a long process with lots of downtime.

Motherboard 24-pin jumper plug

These nifty little things are incredibly cheap and useful. After you wire everything up, you want to fill your reservoir and turn on the pump, but obviously you do not want to immediately electrify your entire system. So you pull the 24-pin motherboard cable of your motherboard and put the plug into it. It bridges specific pins, tricking your power supply into thinking a motherboard is connected. This way your pump turns on without the rest of your system. Once the water is circulating and not catastrophes have occured, you can turn off your power supply and plug the mobo back in.

Common advice

This is advice that's often repeated in watercooling circles for beginners. If you're seriously considering doing this, you will likely already have stumbled upon these. I'm adding these just for posterity.

Do not mix aluminium and copper/brass

Cheaper watercooling components are often out of aluminium while pricier ones are out of copper. You do not want both in your system as they eat each other through galvanic corrosion. If your cooling blocks for the GPU and CPU use copper (they very often do), the rest needs to be out of copper or brass as well, fittings included!

Buy more fittings than you think you need

Remember, per tube you'll need at least two! Check that they have O-rings, as those provide the seal.

Put a drain port into your loop

You should generally drain and flush the loop at least once a year. This will be a lot easier if on low points you have faucet you can attach a tube to and open to drain it. Pulling the loop back apart is generally the last thing people think off when building a custom loop for the first time, so it's useful to know.

Consider quick disconnects

Quick disconnects are special fittings you can put into a tube or attach directly to a port. You can then pull them apart with minimal or no leakage of your cooling fluid without having to drain your loop. Really useful for example the GPU, which tends to be the component that's swapped out most often.

Use cooling fluids over distilled water

Obviously no fucking tap water, ever! But lots of folks also gravitate to distilled water. Cooling fluids like what Alphacool or Aquacomputer make have extra stuff in them, like corrosion inhibitors and biocides that prevent algae build up. You can also mix these yourself if you can get the inhibitors and biocides concentrated but if you're on that level I don't think you need this guide anymore.

Also, colored liquid fucking sucks. Unless you want to pull apart your water blocks and clean them with a toothbrush, use clear liquids. If you want fun colors, put RGB into your case.

120 mm of radiator length per 100 W of heat generation

The two components generating the most heat in your PC are likely the CPU and GPU. Check the specs of those to see how much heat they generate. This number is generally known as the Thermal Design Point (TDP). Radiators come in many sizes fitted to fan sizes, mostly in multiples of 120 or 140 mm, but running this calculation gives you a baseline for how much radiators you need. More is always better! Fit in as many radiators as you can into your case, but if your case can't fit the number of this calculation then you need to look for a bigger case.

Knowing the TDP is rarer for GPUs, you can also use board power or power draw as a substitute. We're doing napkin math here, no need to be precise.

Example:

CPU: 170 W

GPU: 300 W

-> round up to 500 W, which means 5 * 120 = 600. A 360 mm radiator fits 3 120 mm fans. You'd need 2 radiators with 3 fans each to cool your system adequately.

Alternatively, a 280 mm radiator fits 2 140 mm fans. You'd need 3 of those to cool the system.

Radiator thiccckness

Radiators come in different thicknesses. Since what dictates a radiator's ability to dissipate heat is the total surface of it's fins, increasing the thickness improves cooling ability. However, most PC cases, even full towers, are practically limited to 45 cm rad height at most.

Noise

A big motivation for doing this was noise. Cooling everything with a custom loop means that I've lost the 2 fans on my CPU air cooler and the 3 fans on my GPU. What remains are the case fans only, 2x180 mm ones and 3x140 mm. Those can now run at dramatically lower speeds (10% fan speed at idle, ramping up much more slowly) for a nearly silent build even under full load. The pump and reservoir combo I've chosen are isolated from the pc case through rubber standoffs which means that the pump, even when at 100%, runs dead-silent.

Chasing diminishing returns

Switching to a custom loop alone is a massive bonus to the computer's ability to be cooled, because water is a much more efficient way to move heat than air. Case radiators also have much more volume than the heat sinks on your GPU and CPU right now, improving the cooling further.

Once you step into this world, the choices open to you are staggering. Delidding the CPU. Using liquid metal instead of thermal paste, etc. etc. Unless you're planning on overclocking your system, there's no point to doing any of those things that are actively dangerous.

Liquid metal buys you a couple of degrees °C at best, at the cost of being dangerous and difficult to apply and even a tiny escaped drop having the ability to short and fry your GPU for good.

Delidding your CPU is only useful if you plan to overclock. I did it, but only because the company Thermal Grizzly sells delidded CPUs and a fitting water cooling block. If you're doing it at home, the investment is way too large to make sense. Delidding also requires liquid metal afterwards. See paragraph above for that.

If you're in this just because you want a high performing system at less noise, then using a PTM material instead of thermal paste is going to be good enough.

All of these improvements lower temperatures of your components. Delidding the CPU and cooling it directly buys you something like 20°C under load. But the thing is, a good water cooling loop can absolutely cope with a high performance CPU running at 100 per cent. With the IHS on it'll just push 80°C instead of 60°C.

Functionally, there's no difference if the CPU runs at 60°C or 80°C. The only time it matters if if you're over clocking and through that causing the CPU to approach its thermal limit. Then dropping it by a few degrees makes sense.

If not? Skip them.

I hope these help people. Feel free to ask any questions!

Edits in no particular order:

Loop order does not matter

PC custom water cooling loops are not car engines, and as such the thermal differences between coolant and components is much smaller. This means that having a radiator follow a component to immediately cool the water down is much less effective than just adding more rad volume. It also tends to make your tube runs messier and is overall not worth it.

I caught wind of the BC-250 after the Linux on PS5 post where @moocow1452 posted about them and shared a link. The BC-250 is an APU cut down from PS5s that didn't pass Sony QC, they were originally slated for waste but largely got picked up by crypto miners. I hadn't got myself anything for my birthday or promotion, so when I found a deal, I grabbed it for the HTPC I'd wanted but couldn't justify. (Long post, tl;dr at end)

Specs:
6-core AMD Zen 2, stock clock 3.5GHz
AMD custom GPU, stock clock 1.5GHz*
16GB unified GDDR6 RAM+VRAM
M.2 2280 slot
*Stock lock is weird and limited, will discuss

I'm not sure exactly when they got proper Linux support or folks started using them for gaming but I'm definitely behind the curve. You used to be able to get them <$100 but now they seem to be going for $180. I found a "pre build" for $275 shipped that had it ready to go, including a case, cooling, PSU, 256GB SSD, and unlocked BIOS... this isn't too far from what I would've totalled had I gotten the board for ~$80, and saved me some headaches. While the seller pre installed Bazzite, no way was I trusting the OS installed by a random eBay seller.

So what's it take to make one of these usable, what trouble did the seller save me? Let's start with the BIOS -- this needs to be unlocked by flashing though I'm not totally sure if it's required to use it at all but at the very least it unlocks the dynamic 512mb VRAM. 16GB total is kinda limiting today and the default split is 8/8 RAM/VRAM -- a static split isn't exactly ideal -- depending on your game, you could easily be maxing out one with unused of the other. The dynamic 512 reserves only 512mb for the GPU alone and allows the rest to be properly split as needed (mostly, there's still technically a VRAM cap that can be raised with kernel parameters).

OK next saved headache was the cooling. See, these were built as server units with fans set to blow air through them. The heat sink fins are closed off on the top. One can print a sleeve to have a standard fan push air through, but opening the fins up and letting the fan push air through from the middle is more effective, more like your typical consumer GPU. Seller did a messy job, but it is opened in the middle. Some folks use 2 fans and fully open the fins for a cooler/quieter build.

The final headache was minor but they saved me from sorting out the PSU and power button. Since the board is powered just by 8 pin PCIE, for this PSU, two pins on another header need to be shorted to stay powered on, seller already had this in place. Seller also soldered power and reset buttons to the board.

So with the hardware sorted, what does the software end look like? While it can technically run Windows, but it does not and will not have GPU drivers -- though folks have added external GPUs. The main 2 OS options folks recommend are Bazzite and CachyOS. Folks say Cachy is better but benchmarks I saw weren't that compelling, I already use Bazzite and it seemed to have less extra steps so I went with that, standard installation process went smooth and it was basically ready to go from there... So why am I futzing around so much?

Well, the dynamic VRAM can collide with the default ZRAM swap, causing crashes and other errors. So I went through disabling that, enabling ZSWAP, then tweaking config and kernel parameters according to community info. So part of my futzing was comparing speed and stability with the various RAM settings and swaps.

So the next futzing was the GPU clock. The stock is locked at 1.5GHz over 920mV always no matter if load is high or low, but PS5 runs em around 2.3GHz and higher volts. There's a community governor that can decrease volts/clocks under light load and unlock the stock. Folks easily get 2.0GHz at 1000mV but some push 2200+ at 1150. So testing with that was more futzing around, playing in Expedition 33 for half an hour to see if settings were stable. I found the defaults to be fairly aggressive, causing artifacts and crashing. I disabled all their default points over 920mV, set one at 2.0/1000, and called it a day. Didn't feel like trying to squeeze out another 100MHz. Also boosted the bottom 700mV point up by 20 since I saw some graphical artifacts at low loads and that cleaned those up. (<700 locks it right back to stock parameters so that's the effective minimum)

Next up is the CPU, stock has it at 3.5GHz at 1180mV. Folks have been able to get the same clock down to 1000mV, some mad lads pushing 4.1GHz and like 1.3V. There's another community tool for helping with this where you punch in a target clock, voltage, and temp, then it tests it out. I haven't settled this one out, I may go for a bit of an under volt to help keep it quieter/cooler since it's an HTPC, or may just leave it at stock. Folks also recommend disabling mitigations for Spectre/Meltdown as it significantly slows the CPU and you're fairly unlikely to need it for a DIY Steam Machine.

So I was stumbling through this following the community wiki, it had some good and useful info, but it didn't seem entirely consistent and sometimes was just wrong. Turns out, it's an AI compilation of community info. I'd have saved myself a ton of trouble if I'd found this repo and followed their instructions as that's what I eventually found worked best.

Expedition 33 was my main real world benchmark tool. With FSR+LSFG on medium, I was hitting 1440p@90fps, 1080p@45fps without. The site compares it to a 3060, I thought that was mine until I double checked and have a 3070. It does E33 with DLSS+LSFG at 1440p@120fps cap on medium, or 1080p@60-70fps without either. So while the .info site seems to accurately describe the GPU as similar to a 3060, their benchmarks seem to include scaling and/or frame gen based on my limited testing. However, E33 still looked bad and off compared to my Nvidia at the same settings. After a day of fiddling, I looked it up and turns out, E33 specifically looks worse on AMD.

As for HTPC apps, Jellyfin flatpak works great. Official Plex apps are deprecated, both flatpak and snap, they run terribly. Girens (unofficial Plex client flatpak) seems to work well, but requires a mouse and sub menus don't work in big picture. I plan on using a PS4 controller (touchpad works as trackpad) for the time being so this isn't too big a deal, but it's definitely less polished than the Plex smart TV app. VLC with the network drive mapped so we can play direct from file as back-up. And finally VacuumTube for YT smart TV interface with ad blocking. Hardware encoding/decoding does not work and will not so it's all done by the CPU, but it was enough for my high bitrate 4k HEVC decoding test.

It doesn't have proper sleep states. Can't wake from USB/LAN, power button only, and doesn't actually sleep CPU/GPU so it doesn't really save power in sleep. Shutdown/cold boot is inconvenient for a daily use HTPC. Estimated 60W at idle, about $6 a month left on 24/7 for us, perhaps an extra $50/yr compared to a box with proper sleep. However, most it can pull under load is 235W, less than a typical gaming PC. Not enough to realistically break even versus a gaming PC with proper sleep though, so it's probably an expense worth considering in comparison shopping for a similar usage.

Final Verdict: do I recommend it? Honestly, if you're getting bare board for $180, maybe not, but also depends on who I'm recommending to. I seen someone on the discord selling them for $150, others may be selling too, perhaps prebuilts even. Between case, fan, PSU, and SSD you're ending up closer to $400 for a running build -- and that's before wifi/Bluetooth/controller if you don't have extras of those already. At that rate, you may be better off getting a mini PC + eGPU dock + older graphics card, but I'd need to shop around to see how their price:performance ratio compares. On top of that, you'll likely need to do the trouble the seller saved me. Getting the software going is fairly simple nowadays though, so long as you have the foreknowledge to just use NexGen's repo. If you're curious, do some comparison shopping between what you find for these versus a mini PC + eGPU setup. Don't forget to consider the cost of the BC250's idle draw if you don't want to cold boot each day/use. I wanted a machine that could be a media client, do some modest gaming, and to experiment with AMD and unified V/RAM. I ended up with one that can do that and some AA/A, so I'm satisfied, despite the quirks.

I had this idea that is mostly a pun. I want to train a Love Language Model (LLM) on a large corpus of love letters and/or poems. Tricky bit is to find a large collection that is available. Any advice on sources of transcribed love letters?

Edit: so far I've scraped a few website and downloaded a dataset from kaggle. I think it's enough for a just for fun project :)