site banner
ArjinFerman  Tinfoil Gigachad  1mo ago (text post)   610 thread views

Tinker Tuesday for July 14th, 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.

1
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

I added switching between weapons, which ended up being a bit more work than expected. I had to rearrange some data buffers, because it turned out some stuff I'm passing to the material shader isn't useful, and other things that I don't need aren't being passed. I'm also considering switching the projectile collision model from circle-circle to circle-AABB, as this will allow me some flexibility of projectile shapes, and hacks for high-velocity projectiles passing through an enemy in a single frame (just make a long rectangle, lol).

I also played around with some benchmarks. Fired off my old desktop to see how the game performs there, and it was interesting. It seems to be free of the issues plaguing my new laptop like overheating, and the sudden choppiness when I do GPU-CPU copies of the background texture, but since it's pretty old hardware, the total amount of monsters it can handle at 60 FPS is actually lower. Still, I think I'll start recording the dev journal videos from there, as I seem to be able to get better recording framerate and quality out of it.

How have you been doing @Southkraut?

Wrote even more design documentation. It's a purpose unto itself.

Other than that, I spend quite a few hours trying to blender together a humanoid character, but ended up getting frustrated with my inability to make Blender do what I had in mind. So I did some thinking. Would it be possible to construct character meshes procedurally in Unreal Engine, for which I already have the technique down and only ("only", he says!) need to write up the exact specifications for the character? And then animate the product thereof manually? My LLM Copilot says yes, and as far as I can judge it sounds credible.

Now all I need is time and concentration. Hard to come by lately.

By generate procedurally, do you mean use unreal's tools like metahuman and just export a model? Or like, write code to build you a model? If the former, yeah, I would definitely recommend using some sort of preexisting humanoid avatar model that has all your blend shapes and replacement slots configured, then just building from that base. Building new dedicated humanoid characters is really an activity for people doing very art-driven projects with particular aesthetics. I personally work in Unity, so I just import UMA and that gives me good default humanoids out of the box that I can always replace with custom models later.

Write code to build me a model. I've already built myself a small toolkit for building procedural meshes, and I have more code from older projects that I could port over or take as inspiration for expanding those tools. The main challenge as far as that goes is that I'm not kitbashing together subdivided primitives, or building fairly simple mechanical shapes, but trying to sew together a full humanoid, give or take a few simplifications. That's obviously quantitatively more work, but there's also more to take into account aesthetically. Beyond that, I've never done any rigging or animating, so that's just entirely new.

I built a first prototype of my build-procedurally-and-edit-in-unreal pipeline, and it didn't work out. Mostly because I failed to consider various details - like whether I can accept having uniform proportions on all characters, for example. I'm putting the whole animation business on the backburner for a bit while I sort out my priorities on that one.

And since I haven't accomplished anything visible in too long, I've started doing a little UI work instead. Just be able to see some results at all.

Yeah, that's a really hard goal. It's only gonna get worse once you add blend shapes, equipment, and attachment points. You'll start to find that you really need to have unified skeletons, because if you don't, you'll need new copies of models for everything. IE you'll need to reskin and possibly even mess with the low-poly mesh for every pair of pants x every base mesh you build. If your dwarves/elves are just humans modified with aggressive blend shapes though, you only need pants for men and pants for women. It's definitely a cool idea, but I tend to think it's way too big for a single person to accomplish, not one of the low hanging fruit for procedural generation, like terrain or animating a spider.

It's a good thing I won't need dwarves or elves, normal humans will do for me. Along with being able to elegantly handle characters of different ages and builds, I especially wanted to avoid the usual video game failure mode of every character having the exact same proportions and looking weird by the sheer excessive uniformity of it - but I guess that's just a technological/economical necessity for now. I'll go with a uniform skeleton as you advise.

Blend shapes will allow you to mess with proportions. That's actually pretty easy, and the industry standard solution for stuff like scaling fat or muscle. The problem is that animations wont play nicely with blend shapes. If you increase the length of legs or forearms rather than scaling height uniformly, all your physical interaction animations begin to break. If you really want truly unique builds, you basically need to do VR style procedural IK animation for all your interactions.

Which would be the objectively correct solution - but that exceeds my abilities for the forseeable future. I need to learn not to go chasing down overly-ambitious rabbit-holes. Keep it Simple and all that.

GPT5.6 Sol is something ... I needed a simple 3d printer leather belt skiver jig. The bastard two shotted it.

If you are into 3d printing the models are quite good at creating functional prints.

Can you describe the whole process? Did you upload an image with dimensions? Did you verbally describe what you wanted? I'm not at all sure what the work flow looks like.

I have ocp viewer opened in VSCode, and I have also opened a folder with venv. It also has access to freecad and couple of other python libraries.

› Here is totally unrelated problem for you. I need a 3d printing jig for skiving leather strip - up to 30mm wide. It may use standard utility or box cutter blade. The goal is leather thickness of 2.3 mm. You may use additional imbus bolts and nuts for clamping or so. It should be easy to print sturdy and safe.

It generated a solid first design. and send it to the viewer. from there I just sent one or two more prompts to make the clamps better - it had forgotten some features, there was stuff that was impossible to print, but nothing drastic.

Those are the following prompts

I like it but the clamping mechanism is a bit off. Looks weak between the layers, the holes for the both is unaligned with the hex nut - probably you didn't do the right trigonometry because of the slant, and doesn't looks secure - looks like it is prone to snapping.

Is there a cutout for the bolt head - if you want you can make the end clamps upper end, paralel to the main plane. with a proper cutout. we will still have the slanted clamp but mounting will be more secure.

It's final answer

Reworked the blade seat to eliminate the unprintable parallelogram slots.

The blade now:

  • Drops onto two completely open sloped ramps from above.
  • Has no enclosed gaps or trapped support cavities.
  • Is supported beneath both ends, outside the leather path.
  • Registers against small open lips integrated into the removable clamps.
  • Has its sharp ends covered when the clamps are installed.

I also moved the exit guide behind the complete raised blade envelope, removing another hidden intersection previously carved out by the slot.

Validation confirms one valid base solid, zero blade-seat obstruction, and an unobstructed 30 × 2.3 mm path. Updated files were regenerated and sent to OCP port 3939. Changes remain uncommitted for inspection

Everything else - proper positioning and depth of nut, holes, the guiding rails for the leather the whole size of the object, the mounting holes so that it could be bolted to a worktable and a few other unasked for but proper features were all its idea.

I've been putting off working on my next album for too long. My plan is to pick up my guitar every day, and keep playing it until I come up with one interesting riff or chord progression. Hopefully if I keep at it for a month or so I'll have enough ideas that I can start assembling them into songs. Yesterday evening was productive.

I have noticed a lot of creative pursuits are like this; getting anywhere is far more about fostering mundane, day-to-day discipline than it is about feverish bouts of lightbulb-brained inspiration like the stereotype would suggest. The people I know who do this wait far too long to know what they want to make, and in the end barely make anything. Far more important is to get into a routine of just working on something everyday, honing your skill as you go on, and even if 95% of it is shit that 5% of good, worthwhile material will accumulate overtime.

100%. Every successful novelist will tell you the importance of writing 1,000 words a day, every day, even if (especially if) you don't feel like it or don't feel "inspired".

There's definitely other processes that work. I know some people who do manage to get books done by ruminating over the story in their head for ages and only writing when inspiration strikes them. Met one at a party last night actually, was an interesting talk. But they're the sort of people who actually are struck by inspiration regularly. For the rest of us, there is daily discipline.

Whatever one's process is, if it does not result in words on a page month over month, it will never result in books.

Last week I whined that I really should have gone the Full Monty right away on my graphics card, mostly because I'm having so much fun dicking around with local LLMs but want more out of them. Today I stumbled upon the elegant solution of buying a used 3090 and a 1200 watt power supply (which I would have needed for the better card anyway) to get myself to a nice respectable 40 gigs of VRAM, which is oh-so-tantalizingly close to being able to run a full 70b model at the desired Q4_K_M. Odin dammit but do I ever wanna pull the trigger and spend another ~$1500 to net myself that headroom!

That is all. For now...

Semi-related graphics card question: My PC is 10+ years old right now and my mac can run some games but emulation and minimum graphics only get you so far. In the next 2-3 years will the current expansion of chip fabs lead to cheaper cards in the future? I'm hoping to build a cheap, medium-tier gaming PC for less than say $1500 (cheap peripherals). Thoughts?

It's been a while, but back in the day, Logical Increments was often pointed to as a pretty good guide for approximately how to get in at various price points. It looks like it's still being updated. Not sure if the specific recommendations are now dominated by ads or anything, but even if so, it helps give you a sense of what is a reasonable expectation at different price points. If you're interested in gaming, they give some FPS performance at various resolutions in some games. I have no idea if the games they're referring to are meaningful today. I think they get their performance numbers from other sources. You can use it as a jumping off point for any other specific tradeoffs you prefer. When I last did it, I used them to get myself into the ballpark of about what I wanted and then used PC Part Picker to narrow in on specific components and do a bit more optimization on current prices.

As far as predicting the future, I can't do that. But just briefly looking, it seems like capable enough gaming PCs are currently not obscenely unreasonable. Probably still on the "somewhat overpriced" side, but honestly not as insane as one might think.

With the caveat that my semi-informed opinion plus experience will, with $7, buy you a latte at Starbucks, I think that the new fabs will help, but prices will still be above average unless an AI crash tanks the chip market in that time frame. I don't think such a thing is out of the question, and I expect some sort of market correction is inevitable in the long run, but I now grok why AI demand for wafers in general and memory in particular is so insatiable, which in turn leads me to believe that looking for lower prices in the medium term will be a coin flip at best. I'm personally glad that I went the other direction and decided eff it, I'm getting in now on a good deal and I'll let the chips fall where they may, and my only ragret is that I didn't spend the extra ~$1600 when I had the chance and go whole hog. I expect that if you're willing to buy used, you should be able to find something more than suitable when you're ready to make the purchase, and even if you want new, the budget/mid-tier stuff is plenty capable of suiting your needs. The performance of the Steam deck is instructive there.

What's the best I can do with a 32gb gpu? I've been running gemma 4 31b in LM Studio which lets me load the model into vram. Loose rumors say you can fit more into it, or use ram (64 gb) or the cpu.

Gemma4 is quantization sensitive enough that 31B-Q5 is probably the best bet, closely followed by Gemma4 26BA4B around Q4-Q6 where you want (much) higher speed and simpler questions. That said, Gemma4 is great for writing, mediocre for coding, and lackluster for complex spatial reference.

Qwen 3.x is really good for local coding, especially 3.6 35B-A3B and 3.6 27B. Depending on context length, you'll want either Q6 or Q5, make sure to get the MTP variant setup where available. If you just want to code, these two models honestly cover 99.9% of anything you'd be able to do locally with any practical single-GPU machine.

Nemotron 3 Nano Omni 30B-A3B (Q4-Q5) is a little weird and it's not as good for code generation as Qwen, but it does handle really big contexts for analysis better. It's a bit annoying to set up properly, though.

Llama.cpp can run GGUFs of models in mixed-inference mode, where some layers operate on the GPU and some run in system RAM. This can be slightly slower for MoE models (anything with AXB as a suffix), or much slower for dense models, but it lets you run stuff that'd otherwise be impossible. I've got GLM-4.5-Air running on a single nVidia 3090 and 128 GB system RAM at Q8... admittedly, at <5 tokens/sec. Still can be useful for things like draft review if you let it run overnight.

((If you're really desperate, you can even offload to NVME, but this is a very bad idea for drive wear reasons.))

The general rule-of-thumb is that even if you're willing to accept slow inference, quants under Q2 are usually useless and under Q3 are marginal at best, so this doesn't mean it's worth the bandwidth and drive space to go with something like GLM-5.2 at Q1 just because the parameter count is high. But there are some useful options, still. For your setup, some that might be worth evaluating:

  • GLM-4.5-Air (355B A12B). Was the big standard for writing for quite some time, and while it suffers from quantization, it's still pretty useful.
  • Solar-Open-102B A12B, probably Q4-Q5. Kinda general-purpose, nothing outstanding, but it's unusual for its size point and fits your specs pretty well.
  • Mistral 3.5 Medium 128B, probably Q3-Q4. More meant as a strong reasoning code model, unfortunately not a great one, but it's the best Euro model and for language support that sometimes helps.
  • MiniMax 2.7 (229B) at Q3 (or variants like IQ3_S). Not a huge fan of it, but has a following.
  • DeepSeek V4-Flash (284B A13B) at Q3 or Q2.
  • Qwen 3.6-27B at Q8 can be worth having around, despite the RAM offload. Most of the time the difference from Q5 isn't massive, but I have seen logic and parsing differences.
  • GPT-OSS (120B, effectively A5.1B). This one's weird because of the native quants, but some people really like it and some people find it very meh.

Thanks! I've looked up what these q levels mean. If you're doing coding, you'll want high precision and thus at least q5, whereas for everyday chat you could get away with q4 or even less?

I'm still unsure how model size (pre q.) plays into the final "iq" or usability, after pretty heavy q. has been applied, compared to a smaller model with lighter quantization.

It's a little weird, and it depends on what model you're using, what you're doing with it, and your set up. The general rule that if you have to go under Q4, you're probably better going off with a smaller model, but that's... very approximate.

Raw parameter count alone doesn't make for a smarter model, especially across families or generations. PALM was a 540B model, for example, and it'll make mistakes that modern models in the 100B range won't. With more recent models, raw parameter count matters less for reasoning ability, but smaller models will have far more and larger gaps in knowledge. As an example, Gemma4-12B has no idea what the Burned Furs controversy was and assumed it was about someone having heatstroke, Gemma4-26B-A4B talked about a real-but-different event from 2016, and GLM-4.5-Air knew the actual thing and had a summary equivalent to the actual furry wikis histories.

Mixture-of-expert models can be thought of as having a bunch of specialized sub-models and a coordination engine, so they're usually dumber than their headline number for general purposes, but smarter than their A#B parameter number would suggest for specialized purposes.

Quantization generally doesn't make the model dumb, so much as it makes it wrong. Hallucinations, weird word choices, different languages or character sets, or 'obvious' errors tend to pop up. I've heard Q2 described as a model being drunk, and it's not wrong. This varies a lot depending on model and use, though.

Be aware that there's the separate concept of KV cache quantization. That's a runtime memory toggle, not part of the model. Most models run KV_Q8 perfectly and KV_Q4 reasonably, but Gemma and some reasoning-heavy models are a little more 'correct' at KV_Q16 and goes outright wrong at KV_Q4.

With the caveat that I am totally not an expert in this situation, it sounds like you're right in the sweet spot, as 31b will fit comfortably into 32 gigs at Q5 with a little bit of headroom. If it were me, I'd probably be sticking with 32b myself, so I'd be right there with you.

Right, though I've heard that with quantization - reducing the demand for accuracy in the memory's calculations, you can cut a model size effectively in half or more and fit something much bigger into vram.

Gattsuru has already covered this ground better than I could, but IME Q4 can be sketchy enough that I would have serious general reservations about moving down to Q3 and personally prefer Q5 when it actually fits into 16 gigs, which is... not that many models, unfortunately, hence my burning desire for moar VRAM.

I'll caution that mixing different sizes of memory and generations of card means you're really going to want to stick with llama.cpp over vLLM, and be limited in hardware dtype support. For casual home use, that's fine, but there's a lot of talk about FP8, FP4, NVFP4, that just won't be worth it for you.

It's like you know me or something!

The underlying issue that I have is that my motherboard is not ideal for dual GPU setup either, offering only PCIE4x4 for a secondary GPU, so when if I find myself that deep in the rabbit hole, the deal that I am making with future me is that I'll evaluate my options at that point with an eye towards still trying to keep the price reasonable and thus getting a motherboard that'll give me 2x8 for PCIE and probably a second 3090, because if I start thinking too much about that now, I might be a little too tempted to roll the dice on one of the many superficially good-looking but probably fraudulent eBay classified ads that'll tie up a non-trivial amount of cash for a month or two while I work a CC dispute claim to its inevitable conclusion. It's a good story I'm telling myself, right?

I didn't spend my digital fast writing much. I almost finished my first draft of the smut novel (I plan to finish it today) and then it's editing time!

Out of curiosity, what kind of smut is it? Hetero/homo/both?

Stereotypical, that is, women swing both ways, men do not.

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