site banner
ArjinFerman  Tinfoil Gigachad  1mo ago (text post)   813 thread views

Tinker Tuesday for June 9th, 2026

This thread is for anyone working on personal projects to share their progress, and hold themselves somewhat accountable to a group of peers.

Post your project, your progress from last week, and what you hope to accomplish this week.

If you want to be pinged with a reminder asking about your project, let me know, and I'll harass you each week until you cancel the service.

2
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

Continuing on RP2350 development, fighting the software on a Pico 2 while waiting for a board fab to get my first-run prototypes back (and inevitably show me how badly I fucked up several subsystems and how much I'm going to regret QFN for prototype-gen2).

I used to wonder why software debugging support on edutech microcontrollers was either deep in the SWD/JTAG weeds or absolutely nothing. All of the components exist: micropython supports tracing with a compile flag, DAP means you don't need to build a full IDE (albeit at the cost of writing Yet More Json), there's more than enough computer and memory overheard on normal boards that the processor and memory cost doesn't really matter; dual core and multi-processor systems are so common that you can buy them by accident.

I don't wonder anymore. Holy shit.

I had to scrap my previous approach to the bullet simulation optimization, as I did everything in one go, and manage to break everything else (not terribly surprising, as I extended I bunch of existing data structures, via a lot of copy-pasting). So I reverted that and started over with a more gradual approach, and managed to at least not break anything, but the bullet collision detection still was not working. I avoid using AI for this project, as part of the point is keeping my mind sharp, but after finding a few stupid copy-paste mistakes manually, and stalling after that, I figured it's either building some elaborate debugging tool, or asking Claude, and I caved. It indeed pointed out some other stupid copy-paste mistakes, and now everything works.

Disappointingly, it doesn't seem to be much of an improvement. Ok, so I get 60 FPS with +/- 360K monsters for a few seconds, whereas previously the maximum was 250K, but the GPU get's hot and stalls out rather quickly. I was hoping it would make the simulations with lower monster counts more sustainable, but I can't say I see the difference. One day I have to test it on Windows (my Linux distro is acting a bit weird with the GPU, and most games are kinda choppy), or my old desktop which actually has a decent graphics card.

How have you been doing @Southkraut?

Not much this week. Some refactoring. Not a lot of time spent on it, but what I did spend generally yielded good returns in terms of improved code quality. I also did some thinking about my architecture and refined it a little, but it's slow, gradual, incremental improvements, with a very long way to go yet before anything interesting happens.

I bought a new standup jetski, I've wanted one since I saw them ripping around as a kid. Unfortunately I didn't quite do due diligence before buying, so I have a cylinder with low compression that looks like its got either frozen rings or a marred up cylinder wall. Goal is to get it going in the next couple weeks, so tier list is, hopefully the ring clears up with some penetrant I put into it last night enough so it starts and runs and frees them the rest of the way, or if not that, a hone and new rings and a gasket kit, and if not that, hopefully the cylinder can be bored at a nearby shop, and a new piston sends me on my way. All in all, I got a good enough deal on it I can afford to put some new parts into it, but its certainly eating into my profit/fun time on it.

Do we have many local LLM users here? I'm curious what people are doing: what models are people using? For what jobs? For what reason? On what hardware? With what runner?

As I mentioned to @WhiningCoil a few days ago, I mostly run a Qwen3.6-A3B-Q4_K_XL on llama.cpp's llama-server and connect to it from https://pi.dev/, using a Radeon 780M in my laptop. It's been decent for grinding through smaller coding jobs under close observation, though like any Chinese model it'll just give you the party line if you start asking it about Taiwan or Tiannemen Square. I've also been using a gemma4-26B-A4B for general questions about the world when I'm at session quotas. The other big reason I'm getting into this stuff is that I never want to be locked out by a subscription. Haven't looked at image or video generation at all.

I have a small ML server that I initially set up for some work stuff, and have since retrofitted for LLM and diffuser use. nVidia 3090, i5-14400, running between 128 GB RAM to 192 GB RAM depending on what else I've shut down. Squeaked in just before RAM prices spiked (and am kicking myself for not grabbing three or four more of the 64x2 kits), if you want to know the why on the weird RAM numbers.

I'll caveat that you just shouldn't expect Claude or even Grok-level outputs from local models on their own.

For LLMs runners, I've mostly stuck to llama-server (and forks) as well, after an initial and short-lived love-hate relationship with LMStudio. I have a few custom bits of code for sequencing larger grouped requests, but they're worse-than-vibe-code level stuff and basically just a UI and for loop. Toyed with SillyTavern, just in the hopes of getting better organization, but it's really heavily built for roleplay and I'm not that interested in it. I've looked at and played with some agentic-ish stuff in heavily sandboxed and airgapped environments, but when the best options are nanoclaw, hermes and odysseus, but when the least obnoxious one is powered by pewdiepie, there be dragons here.

Writing:

  • gemma4-26-A4B is a great editor, beta reader, and brainstorm sounding board. It's the closest to okay prose from a local model in its class, although beating the obvious AI tells out of it takes some effort and it's seldom very interesting. Also seems to have the best MTP assist (though I had the to build the atomic-turboquant variant llama variant to get MTP to work when it first came out; don't know if the situation has changed there).
  • [Cydonia](https://huggingface.co/TheDrummer/Cydonia-24B-v4) (24B, mistral-based) and ```Strawberry Limeade (70B, llama-based) are older models that were pretty useful and I'll still pull up on occasional to sanity-check stuff against.

Coding:

  • Qwen 3.6 is hard to beat for simple and fast work, especially things like bringing in an image sketch and converting it into XAML or webdev, or beating some simple file munging into shape. In addition to 35B-A3B, I'd also point to the 27B-MTP dense variant. It's not as fast as A3B, but the gap's smaller than you'd expect, and in some use cases it comes across as much smarter in my experience. I'd also recommend it any time you have a sketch or powerpoint art-level design you want converted into a GUI representation, and can't use a cloud model -- far from perfect, but easily saves hours of work.
  • GLM (ranging from 4.5-Air at 106B up) can be good for complex work, refactoring, and troubleshooting -- but even at moderate quants, it can be fifteen minutes per turn. Great where I've got a ton that needs to go into the hatch and can work on something different; terrible for anything where a fast OODA loop is important.

Standard vs uncensored/abliterated models is a hard question. Qwen is very refusal-prone, and not just on political topics. While Gemma4 is surprisingly willing to play along for a variety of topics, it still has some hard refusal points, some of which can come up surprisingly rapidly. And not just for weird smut, either. I've had .

For Image Generation, your two power user options are Automatic1111/Forge WebUI and ComfyUI. WebUI is the easier option to get started with, and still has a good level of support for things like img2img, swapping models out, or using various plugins or controlnets. ComfyUI's much more capable and eventually lets you do things like switch between models for different stages of a pipeline, but there's very much a 'who wants to drink from the firehose' moment every time you get started, and managing workflows sucks. On the other hand, if you want to run something like TRELLIS2 or Wan3d, ComfyUI's a lot easier (though not easy!) to set up.

In terms of models:

  • SDXL based models like the Illustrious and NoobAI family are good for producing general 'vibe'-ish scenes with one or two actors, so long as you don't need precision, and they're pretty fast.
  • The current new hotness popular options in the furry fandom are [Chroma](https://huggingface.co/lodestones/Chroma) (9B, FLUX.1-schnell-based) and [Anima](https://huggingface.co/circlestone-labs/Anima) (2B), which favor natural language over the SD-style "throw a bunch of words at it" approach. Much slower, though. Qwen Image 2512 variants also fall here, although their workflows can be a lot more annoying.
  • Qwen Image Edit and Flux2-Klein are the best edit models, especially for keeping consistency in a scene while tweaking it, or moving a character from one setting to another.

2D->3D Models:

  • TRELLIS(2) gives the nicest-looking outputs for a given input image, and supports(ish) transparency.
  • Wan3D gives more 'whole' models that require less post-processing to ship to a 3d printer, but tends to be a little fuzzy.

Animation:

  • WAN2 has the most support and has been out the longest.
  • LTX-2.3 is much faster and comparable or better quality.
  • SCAIL2 just came out, but I haven't even tried to set it up yet. Initial reports look good.

I had a hell of a time getting any non-trivial animation model working in Forge WebUI, and the ComfyUI workflows get nutty pretty fast. If you want to experiment with them and not go leaping into the deep end, WAN2GP gives a lot of workflow options, at the cost of sometimes serious performance costs and a bad tendency to automatically download a model without warning.

Well, I switched to gemma4:12-it-qat and it worked a lot better. Made an opencode command that takes a specific file and asks for unit tests on the class in it, and it worked ok. I need to explore setting up a good agents.md file to jump start them on understanding a project. I may check out some random open source project and start fucking around with custom commands. I know the workflow we want is to have some automated review and test creation, so if I can cobble together the right opencode commands, agents, skills, etc to make that easy for anyone to just go /review_and_write_tests and things start happening, great.

I still have no fucking clue what skills are supposed to be. They just look like extra prompts that hypnotize an LLM into believing it can do shit.

Skills are essentially a pointer to a loadable chunk of context, a fuzzy program written as a .md file. "If you want to know how to do X, read this file first" sort of thing. The models get more confused as their context (message history) grows, and especially when the context window fills and the initial stuff gets cut off. Then they have no idea what they're doing. So a "skill" is detected by the harness (e.g. opencode) which injects the short summary into the prompt that goes to the model, which can then ask to read the full file if necessary.

I still have no fucking clue what skills are supposed to be. They just look like extra prompts that hypnotize an LLM into believing it can do shit.

The key with skills is that they are conditional, so you can inject those additional prompts into the context window only when the skill is used rather than having them always around. This lets you write smaller generalized prompts for agents, which in turn improves the output of those agents by keeping them focused on the task at hand and not distracted by instructions unnecessary for the current task.

I looks like I might be the proud owner of the used car holy grail, an actual $500 Honda Civic (I think it's an '09...I know it's a non-Si sedan with a stick shift.). This is definitely a case of the friends and family discount (an old buddy of mine hasn't driven it in years, it's been sitting at our old boss's house, and he just wants to get rid of it) and a neat opportunity to either flip for profit or sell one of my nicer cars to net some cash, a fun little rehab project.

The good news is as follows: I almost bought the thing back in 2020/15K miles ago and did some work to it then, so I have reason to believe that the brakes are fine. It has enough miles on it (I think 210K.) that if the R18 in it was going to crack its block it most likely would've done so by now. I know that it runs and the A/C works. I think the battery is still good (It's sitting on a charger.) and believe that the clutch still works (I know that I replaced the clutch cylinder a few years ago, and it's getting long in the tooth, but it did work last I checked.). The paint is not bad by late oughts Honda standards, and I like the color. The things I know are wrong with it are relatively minor (It needs a bolt for one of the back shocks, a window switch and regulator, the tires are showing cords, and the interior is trashed and disgusting.).

The only real potential snag is that the title looks like his dog ate it, so if I can't sneak that past a disinterested county clerk I'll have to wait for a replacement.

Nice. I've been driving my 07 for almost 18 years now, and I love it to death. Just an absolutely bulletproof car, I basically give it the absolute minimum possible maintenance and nothing breaks except the UV-vulnerable plastic and the usual belts. I keep saying I'll sell it when the next problem comes up, but I think outside fluids and state inspection it's gotten 500 bucks in maintenance since 2021.

The only thing that would keep me from keeping this new Civic is that one of my other cars is that same Civic, but an Si model with a JDM K24 (the original engine cracked the block at 195K miles; it still ran for another 30K but I went ahead and did the motor swap when the original clutch died) and some other mild mods (intake/nice sounding exhaust, short shift kit). It's bad on gas by Civic standards and annoyingly expensive to insure but otherwise punchy/fun to drive and I don't have a long commute.

The third car is actually the nicest on paper but likely first on the chopping block because I never liked it that much (Hell, I sold it once and wound up inheriting it back.).

With a little bit of luck, that's the absolutely cheapest way to drive for the next couple of years. Even if you buy the cheapest, crappiest Chevy Bolt and charge it only using your solar panels, it's going to be more expensive per mile than this car. I'd try to get this thing to 300k miles out of principle.

I have no idea what I'm going to do with it. I already have an extra car that I need to sell but $500 was too good a deal to pass up. Depending on how well it cleans up and how much I like it I'll either flip it or sell one or potentially both of the other cars (I don't love either one of them.), pay off all my debt, and enjoy the beater/cheap insurance life for a year or few until something else I want comes along (I want a Civic Type R, but I don't have Civic Type R money.).

I got a new house. It's significantly bigger than my last house. I think I'm looking at like 5-10 hours of home network engineering to get it put together. Please critique my stack.

  1. Fiber hookup in my garage
  2. Router: fanless small Linux PC with dual Intel NICs in garage
  3. Omada 8 port PoE switch, also in garage
  4. Omada WiFi APs, probably 3 at various spots in the house
  5. One beefier Linux box for home services/storage on Ethernet: Immich, that home YouTube clone, etc

The house has existing coax cable runs to various rooms, which must go. I'm thinking of taping cat6 cable to the coax terminals and pulling them to try to re-run those paths as cat6. This way the coax jacks are replaced with Ethernet jacks. Avoids drilling new holes. Have them all terminate at the switch in the garage. Can put either full desktops or APs on the ends.

I'm planning on 2 vlans. One for the humans and home services. The other vlan for cloud-IoT shit that can only access the Internet and not even cross talk to each other.

I've had good experiences with Omada APs, they can be a little finnicky but most prosumer-level AP solutions can be. I have multiple SSIDs with various VLAN tags applied. My suggestion, especially since you're pulling cable, would be to use wired ethernet as the backhaul medium to connect the APs. Don't use wireless mesh backhaul, if Omada even supports that. I've heard you can use 6Ghz as the backhaul if your APs support it, but that just gives me the ick, I'm a wired-first person.

TP-Link has their Omada management platform as a software you can run on-prem, someone packaged it as a docker container that works very well. You can disable the cloud stuff and just manage it through the local network or over a VPN of your choosing.

What router OS are you planning on using for the Linux box?

I agree with gattsuru's suggestion to get a bigger switch, and bring it in the house, if possible. 16 port (let alone 24 port) managed gear can get expensive fast, but if you plan to only have one switch, it might be worth it to search ebay for older 'smart' switches with manual vlan-tagging capability. But you just bought a house, so... maybe go for gold? My fiber hookup (for 1Gig, far better than what was here before but not anything fancy) gives me an ONT that converts to copper. You may have an actual box that you'll have to set to pass-through mode. (Don't double NAT.) If you have something similar, you can run a cable from the ONT to literally anywhere in the house -- home office, network closet, anywhere -- and not have to worry about your equipment being out in the cold. You wouldn't do that to poor network equipment, would you? I guess maybe your coax terminates in the garage? In that case it might have to stay there. Sarah McLachlan judges you.

You could use one of the technically-standard home networking wall panels, but they are NOT deep, and I wouldn't put much equipment in it if any at all.

It's possible your coax is stapled in places, and so it may not be possible to pull on it. Home telecommunications wiring has always been a bit slapdash; you should see the rats nest in my parents' attic. If some of the coax can't be pulled, you can try and use MoCA to get ethernet through it... though I've had only bad experiences with MoCA, so your mileage may vary. Make sure you use solid-copper cat6 and terminate to jacks. Riser cable will work, don't pay for plenum-rated cable.

Your vlan separation is good, that's the recommended kind of segmentation for home networks nowadays. Make sure that they're not just on separate VLANs but separate SSIDs, and are blocked at the firewall level, so they can't just layer-3 route between each other. I'll note that a lot of home IoT gear can be... annoyed at the prospect of being separated from your personal devices, depending on how their system is designed. mDNS is the bane of my existence. Apple gear is particularly poorly behaving in terms of dealing with complex home networks. A lot of Apple/HomeKit/AirPlay stuff assumes a mostly flat LAN and can get grumpy across VLANs unless you set up mDNS/Bonjour reflection carefully... and I couldn't tell you how to do that, I'm still figuring it out myself. IoT devices that go straight to the cloud can probably do fine just with internet access.

One thing you could do with your BEEFY linux box is run your own DNS, with ad-block capabilities. You could use Pi-hole or AdGuard Home for that. I love it. Makes the internet feel actually usable. I'd recommend using virtualization for it, so you have flexibility. People love Proxmox, I have a soft spot for XCP-NG, not because I have any love for Xen but because their management and backup platforms are more flexible.

Congrats on the new house! I hope your home internet turns out great. Funiculus coaxialis delendus est.

Thanks for the detailed response!

What router OS are you planning on using for the Linux box?

Debian. Seems like I just need dnsmasq, nftables, wireguard and networkd to party.

Might convince myself glibc is too risky for an internet facing host and switch to Alpine.

You could use one of the technically-standard home networking wall panels, but they are NOT deep, and I wouldn't put much equipment in it if any at all.

Oh, I think I'd probably just put a patch panel on the wall so that the runs between the switch and the jacks don't directly tug on the switch itself. The switch could go in an 16U cabinet or some such.

and not have to worry about your equipment being out in the cold.

Actually, since we're planning to put our home gym in the garage (and no cars) y'all have convinced me that me and the machines would be a lot happier with a dessicant dehumidifier. Between that and an short cabinet with top exhaust we should be good. But I wouldn't have thought of this at all without posting!

Apple gear is particularly poorly behaving in terms of dealing with complex home networks. A lot of Apple/HomeKit/AirPlay stuff assumes a mostly flat LAN and can get grumpy across VLANs unless you set up mDNS/Bonjour reflection carefully... and I couldn't tell you how to do that, I'm still figuring it out myself. IoT devices that go straight to the cloud can probably do fine just with internet access.

I don't have much Apple gear and this sounds like one more reason not to start! Thanks ^_^

People love Proxmox, I have a soft spot for XCP-NG, not because I have any love for Xen but because their management and backup platforms are more flexible.

I used to virtualize stuff on my beefy box but moved away from it since taking stable and delta efficient backups was a lot more convoluted than I'd like. I ended up moving each service into its own user and hermetically sealing their dependencies to the user account so that a top-level kopia would take efficient enough snapshots. It's not quite nix-level but close enough that I should probably ask Claude to convert each one to a nix flake. (Some things like Immich are in docker and I just live with that, the important stuff is in a mapped filesystem anyway). Efficient matters because I rsync them off-site once a month. I plan to do the backups weekly once I have fiber with better upload speeds.

One thing you could do with your BEEFY linux box is run your own DNS, with ad-block capabilities.

I'm thinking this should live on the router since I'm a lot more likely to reboot or rebuild the beefy Linux box.

I will give caution that I've had some mixed results from TP-Link. The worst problems have involved their routers, which you're skipping, but their access points and especially management were pretty flaky the last time I tried them (admittedly, 2023), especially with mesh mode problems or multiple wifi network configurations. Doesn't matter a ton for the switch, but you may want to consider stepping up to UniFi.

((Or if you have concerns about China... uh, gfl.))

Get a bigger switch than you think you need. You're looking at one port for the router, three ports for the access point, one for the dedicated home server... and that just leaves three for end users. That's probably enough right now, if you don't do a lot of high-quality media streaming to televisions or want dedicated lines for a desktop computer, but it's not a lot of excess. Especially if you might want a security camera system or have a lot of PC gamers, those three remaining slots can fill up faster than you'd expect, and WiFi that works fine today might struggle when a couple more devices are all stomping the same SSIDs.

Get a box (with a fan) for the switch if you're putting it in the garage, or keep it in the house. You would be amazed how quickly these electronics pick up dust, mouse crap, sawdust, dead leaves, and insects. I'd favor keeping as much electronic equipment inside the house's vapor barrier as possible.

When you pull the coax, both tape (ideally gaffer, if not duct) and tie a good series of knots, and pull a poly line with the cat6. If you're dealing with conduit, I'd argue just pulling a poly-line run alone first, but for normal residential you shouldn't see conduit for coax.

Consider a small APC for the router, switch, and server. Less because of the uptime -- you're buying maybe 15 minutes for any not-ridiculous APC -- and more for the device protection. Voltage transients in particular are murder on single-power-supply servers, and it's as often the mainboard that dies as the psu.

Thanks btw! Excellent advice as always! <3

I set up UniFi stuff for my brother's company but the APs weren't very reliable either, though the shininess of the rest of their stack was pretty nice. I'm mostly just going to edit iptables rules by hand like a neckbeard though.

I have a few recommendations for the Omada\ APs today in 2026 so I'll take a risk with them. I would probably just disable any smart stuff they try to do and manually set their power and see if it'll let me get away with it.

A box with a fan is a solid idea. Ditto for the UPS.

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats