site banner
ArjinFerman  Tinfoil Gigachad  1mo ago (text post)   593 thread views

Tinker Tuesday for June 16th, 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.

no tinker Tuesday for June 23rd?

It's there, I guess it just wasn't approved yet.

mods are asleep. post memes!

What's the best AI coding plan right now?

From what I can tell:

  • GLM gets you about 400M GLM 5.2 tokens / month on the Lite plan ($15 / month). 5.2 is a great model, no performance issues I've seen so far, but at max effort using the ZCode harness it guzzles tokens like crazy. My experience with their plan is still quite limited.
  • Opencode Go ($10 / month) has very transparent pricing equating to about 60M tokens using GLM 5.1, or a bit more (~100M) using slightly dumber models like Kimi 2.7. It also seems to use tokens more efficiently than any other harness I've tried.
  • Claude code's limits are harder to define objectively but it seems comparatively tiny, like 5M tokens / month for Opus. Of course smarter models use these tokens both better and more efficiently.

As far as models, GLM 5.1 is the dumbest reliable model, meaning it won't often make obvious mistakes given simple tasks. I don't think it's really worth the time to use anything dumber unless your task is very well-defined and limited in scope, in which case a free model will do just fine. Gemini has been the biggest disappointment. 3.5 flash is ok, 3.1 pro is bad and expensive, 3 flash is unusable. Opus is obviously best but doesn't have a huge lead.

I don't have much experience with chatGPT, from what I hear it currently has the best bang for your buck with 5.5xhigh. Would love to hear what you guys think as I can't afford every coding plan at once, and each harness has its own learning curve.

Developers seem to be split Claude and Codex (ChatGPT) right now. In my experience Codex seems to be a little better at thinking but a little worse at making sure it didn't hallucinate the function. Claude is a little better at doing what other developers are doing. That can be a problem with things like React where it has been trained on a lot of bad code.

The developer split seems kind of ideological right now since people have strong opinions on Anthropic vs OpenAI.

Gauntlet.ai has a free "night school" with videos of talks that will help you stay up to date.

https://apply.gauntletai.com/night-school

They push you to apply but just ignore it if you don't want to.

They turned me down because I don't have enough programming experience haha. Oh well. Their "classes" or whatever you want to call them are great.

Right now I'm just asking. I got hooked up with a GLM legacy plan and in theory have ~40 billion tokens/month of GLM 5.2, which I'm scrambling to use.

Didn't get much done in terms of code, but I figured I'll get to the bottom of performance issues. So like I said I noticed that I get better graphics performance on Windows, but I remembered that actually things weren't working half-bad back when I was on Mint, so it had to be something with configuration rather than the drivers or something inherent to Linux. Turned out I had some hybrid setup, where it was mostly using the integrated gpu, and only "offloading" specific programs to the external gpu. It's supposed be better for the battery, but goddamn does it screw with performance. Fixing it was a huge QOL improvement so I'm pretty happy just for this.

Ok, so this helped, but the frame rate was still dropping after a few seconds. The simulation I was testing was pretty massive, with 6 figure numbers of monsters, and I thought maybe it's because they get concentrated in a pretty small area. I have a spatial index grid to sort the monsters and exclude huge chunks of them when I do collision calculations, and I remember that it helped to make the grid cells smaller... curiously doing so only seed to make it worse. Then I rendered I refactored the simulation set up, and it's size is now fixed. Using smaller cells just means there'll be more of them, and that can indeed screw with performance, if it's not balanced. I tweaked the numbers a bit, and created a smaller area simulation, with small cells.

This, also helped significantly, but the frame rate was still dropping as time went by. I thought I'll try to wait it out, the player character is on auto fire, so I figured the more monsters die, the less load on the simulation, but again if anything it kept getting worse. So I thought, maybe that's the issue? Maybe they're not properly ignored once they're dead? I kept all the collision detection stuff, and took out just the part that changes their state to dead. It worked! The fans run at full steam, but I can sustain +100K monsters at 60 FPS indefinitely. I still haven't found how moving them to the "dead" state actually increases the load, but I'm positive the issue is somewhere there.

How have you been doing @Southkraut?

Stripped out large swathes of LLM-generated code that I was not minded to maintain compatibility with as I reworked my core systems. Which also means that my minimum viable prototype is no a nonviable prototype, as this has cost me my terrain generation and everything having to do with NPCs, but I'm fairly confident that I'm now in a better position to write those systems myself.

I ended up re-implementing the same component pattern for my PhysicalEntities that I had already used in my Unity days, in which each PhysicalEntity has a number of mandatory components:

  • One for Shape and Substance.
  • One for how it is to be physically simulated - on low-fi rails, in Unity/Unreal physics, or not at all.

And several optional ones:

  • One that handles any functions like Speech, Perception, Locomotion etc.
  • One that handles ongoing behavior, like walking up to NPCs and starting conversations with them, or like seeking and consuming food, or like evading other NPCs.
  • Several others that I forgot. My Unity project had around a dozen such components per Entity.

Obviously there's a lot of degrees of freedom in here. I could combine Functions and Behavior, for one, by simply making Behavior a Function. But I need to take care not to make it too granular - that was a trap I fell into back when, where I spent months building intricate networks of such Functions to generate fairly simple behavior. Like building a gun out of separate functions for each mechanical interaction possible with the gun, instead of just having one overarching "Shoot!" Function. Which was fun, in its way, but not very productive. If I go into that level of detail again, let it be after everything is up and running.

And then there's the matter of whether it even makes sense to create all this architecture, when Unreal already offers a Component Pattern through its Actors and their, well, Components. I'm very likely overengineering everything when I should just use the tools that come with the engine. But I have a slight feeling of disgust when I'm forced to use the Unreal API instead of one I wrote myself. For no good reason, as far as I can tell. Unreal's is probably the way it is for very good reasons. But as always I'm stubbornly determined to make as much myself as I can.

I abandoned/paused my last attempt at coding my own videogame since I got bogged down in tons of pedantic technical details until I lost motivation for a few months. I am self taught at programming and have to look up every little feature if I've never used it before and it's annoying.

I heard some people talking about Vibe coding, making the entire thing without writing any code yourself, and figured I'd give it a try. I'm working with Google Gemini free tier, and it seems okay so far. The idea being I'm doing all of the actual game design decisions but it can implement them way faster than I can with approximately the same number of bugs (which in turn it can fix when I point them out). I am very impressed with its ability to make a workable prototype near instantly and put a bunch of placeholder stuff so I can test things bit by bit while having background stuff functioning, but am worried that its performance will degrade as the code increases. Has anyone else tried Vibe coding? Any recommendations?

(If it matters, the game is an Idle game somewhat similar to Increlution, so doesn't have fancy graphics or physics or stuff, and is mostly clicking on buttons in various menus)

Fun! Vibe coded game development has gotten better and better over the last few months. It sounds like you're fairly new to coding with AI (forgive me if I'm off the mark with that guess).

This article has a good showcase of what the top-tier model(s) from Anthropic/OpenAI can pull off: What it feels like to work with Mythos

In your case, I'd recommend just watching the video at the top of the blog post, with the following in mind:

All of these are one initial prompt in Claude Code where Fable had to take my vague prompts and generate something workable, followed by a couple of additional prompts with minor encouragement (“make it better”) or feedback. What makes these especially impressive is that Claude cannot generate images, so every piece of art or 3D object was made with math alone, not using any external assets.

With that out of the way; the free google models are a great starting point, and I'm sure you'll be impressed at how fast & precise the outputs are. That said, there's a significant difference in capabilities between the free versions of the models & the frontier (best at the time) versions, especially when you pair those frontier models with coding "agents" (software that allows for fast, targeted control of your personal machine).

For the time being, stick with the free models. Have fun, try crazy stuff, see what works & what doesn't. If you notice that the outputs aren't coming out how you'd hoped, sink a few hours into some youtube tutorials. I'm sure there are great independent creators, but there's just so much noise that I tend to just default to the easiest route - videos from the AI labs themselves. This one looks fine: Prompting 101: Code w/ Claude

You'll squeeze more juice out of the free models if you use the techniques described in the video, but eventually you'll get fed up with the mistakes that gemini is making. At that point, it's probably worth spending $20 to get a subscription to one of the top labs. I'm partial to anthropic & claude, but that's probably just because I use it so much through work.

Good luck!

Neat! I've been burning my excess tokens from claude's $20 tier trying to vibecode my own videogame. I want to make a naval warfare sim sorta like Rule the Waves but in the modern era. I have minimal coding experience (I don't know how to use github) and have basically just been feeding claude prompts and letting it sort things out.

So far the going has been good, if a bit slow. I tend to burn through the 5-hour limit within an hourish using opus on medium thinking, with about 3-4 hours every week dedicated to this for the last 2-ish months. The sim needs to run on a mac, so I've spent a lot of tokens telling claude to plan the architecture and look for optimizations, which, to my amateur analysis, has been reasonably successful. The latest version is 47k lines of typescript (is this good? bad? I don't know :) ). Theres ships with different modules, planes, doctrines, missiles of various types, and intercept programming, as well as reasonably(I believe) accurate modelling of radar and flight performance.

I've tried using Sonnet to code, and it works surprising well for things which don't require too much forethought, like "make a model and loadout for an arleigh burke destroyer". It's not so great at things like "The tomahawk missiles keep falling into the sea after launch, find out why and fix it", where it tends to get fixated on one aspect or metric and ignoring others. In the latter case, the launch booster didn't have enough thrust, but it kept trying smudge the lift surface size to get the missiles to fly without it. As a whole, I've found that using Opus (and fable, for the brief moments it was available) lead to net faster progress and implementation of systems I wanted, with at least an order of magnitude less mistakes. A few weeks ago I migrated from a browser interface to electron I believe and Opus made the switch painlessly while I'm sure Sonnet would have fumbled.

I'm decently happy with the current program: ships follow orders, missiles fly, some are intercepted, theres cool sprays of CIWS and 5 inch guns when cruise missiles get to close, and airplanes are mostly capable of launching standoff munitions and doing air combat. That beiing said, it's pretty rough on the edges, theres no main menu or situation setup: I have to add ships/planes via lines of code if I want them in the simulation.

If anyone with experience wants to check it out I'll happily post the code to github. I'd love feedback, I'm sure the llms have missed some big things, somewhere, but I'm to dumb to ask the right questions.

Bambu Labs started their annual 3D printer sale this month and I bought their P2S combo with the AMS 2 Pro, which finally positioned it further down the product stack with a less offensive price point no longer too close to the recently released and superior X2D, which has some hobbled secondary support nozzle head gimmick. I’m told the industry is moving to hot-swap toolhead designs soon anyway.

https://us.store.bambulab.com/products/p2s

I was on the fence because of their AGPL violations but then I realized I already converted from Stallman gNewSense GPL to De Raadt OpenBSD ISC long ago.

So far I’ve only printed diagnostics and cat statues. I was going to print wargame or D&D figurines but then I realized I don’t actually play that stuff.

Anyone here 3D print? Have any fun suggestions, experiences, tips and tricks before I just start printing myself ghost guns?

https://fortune.com/2026/06/15/new-york-3d-printer-ghost-gun-law-feasibility-industry/

I recommend a mechanical engineering project, and doing the CAD 100% yourself. You'll waste a metric ton of filament on prototypes, but you'll learn a lot. You can look up designs online, but not download STLs.

Depending on your skill set you can try building a simple water pump or a pair of tank treads (arbitrarily complex - you can skip road wheels, suspension, return rollers/idlers, and tension wheels for as long as you like). Sometimes, you'll end up buying a small motor to drive your contraption, but also just building the stuff and turning the moving parts by hand is fun.

We have a pair of H2Ds at work, and I have an Ender 3 at home.

I don't find most 3D models (e.g. on Printables) that impressive. If I'm going to go searching for something to get, I might as well search a store-store for physical objects and buy them. What you gain from cutting out the shipping/storage/infrastructure costs of a traditional store, you lose in restricted materials, personal labor, and quality control.

If you want to get the most out of a 3D printer, you need to know how to design 3D models to match your needs. This starts with problem definition, goes through measurement and design-for-manufacture, and ends with actually using the new thing. Train yourself to notice when a custom object would fix a problem, learn how to design it, then enjoy.

Get TPU. Having the option to make flexible things in addition to rigid ones gives you a lot more options, if you know how to take advantage of them.

I've found PLA to be good enough for most stuff, but PETG or ABS are fine as daily drivers too. They're harder to work with, but last better (impact resistance, heat tolerance, and mechanical fatigue are all better. Strength is near-identical.) You can get fancy engineering filaments if you want, but I haven't found much to be in the niche of too demanding for PETG, but not demanding enough for someone else to injection-mold or machine it instead.

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