site banner
ArjinFerman  Tinfoil Gigachad  5d ago (text post)   277 thread views

Tinker Tuesday for January 6, 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.

Physical-work side, I got to do some siding repairs. That's been a !!fun!! way to spend the holiday break.

Software-side, trying to look into the state of modern sorting-assist tools. You'd think, will all of the advances in AI tech, classifying files and sorting them would be a solved or near-solved problem. Microsoft's "agentic AI" concept drives me up the wall for a wide variety of reasons, but this seems like one of the main killer use cases. If you've ever worked tech support for either Gen Y/Z or Boomers, seeing a Downloads or Documents folder with so many loose files that it causes an SSD to slow to a crawl is a pretty common experience, and they can't find shit (or, worse, can find ImportantDocument_final_last_(1)autorecover\current.docx, for now).

So I've been trying to come up with and evaluating possible solutions to this sorta thing.

  • Zero-shot classifiers like CLIP are well-established and everyone's favorite option (eg Immich, the Bellingcat sorter if you trust them, which I don't). Trivially easy to implement in python... badly. You're effectively trying to compare one caption against another in embedding space, and there's no guarantee (and a lot of anti-guarantees) that just because two words close to each other in meaning will have any proximity. And they're pretty limited in domain; many can't compare between pptx and a png period, and those that can do so with basically zero accuracy.
  • Few- and many-shot classifiers CNNs are also well-established and trivially easy to code for a single domain... but only if you know the categories you want before hand, and have somewhere between five and five hundred examples. Even outside of the training cost, that's another thing that's a great tool for a tiny number of uses. Vision transformers have the same problem except more so; I needed close to fifteen hundred pre-classified input files. And really messy to go from one domain to another.
  • Text-only LLMs operating based on file names already have fully-developed solutions. If anyone has a use case where people name their files but don't put them in folders, I guess that'd work for all three of them on the planet?
  • Multimodal LLMs are kinda there. You can just load up GLM4.6V (-flash), throw a supported file on the input, and ask it to put the file into one of several categories. It'll burn a stupid amount of tokens if the interaction between those categories is particularly weird, or if none of the categorizes make sense for a given input, but it's not too hard to handle a decently wide subset of image-or-document files on the input side, and natural human language for categorization. Problem's that you get LLM text as the output. You can try to force the LLM to keep the answer as concise and as limited to a category as you want, but there's nothing stopping it from spitting out parts or all of your categories when describing the one matching one, even in well-constrained environments.

That's on top of other issues specific to implementations: a lot of ViTs and multimodal LLMs depend heavily on breaking, while a lot of classifiers get really stupid if you have wildly different resolution inputs, multimodal LLMs can't distinguish between prompt and content, yada yada.

On the flip side, closely related topics are nearly >98% solved off the shelf, even ones that I'd consider a lot harder.

  • GLM4.6V can convert a short comic from visual medium into prose, with all the complexity that implies. Doesn't matter if a character's name is introduced several pages after they first show up, or they have aliases, or if two characters look similar but only show up in different contexts, or if you've got stuff that's clearly outside of the likely training data. I can give examples that confuse it, but they're problems like 'did it read where speech bubbles were pointing correctly' or 'can it read these abstract shapes'? And that's something with so useless a business case that I can't believe they made much synthetic data for it.
  • Give me code corresponding to a given layout (or pencil sketch of a layout)? That's trivial to build synthetic data and has a clear business case, but XAML sucks, and GLM4.6V-flash is just fine with basic use cases.
  • It's hard to overstate exactly how far categorization has gotten in single clear domains; if you only need to handle images, or only text, you're in great situations. I built a YOLO-based model from scratch for ten categories, it took less than six hours to train on a graphics card from 2018, 1k input images, >85% accuracy. Twelve years ago, that's a pipe dream for a doctoral program; today, it's a sign that I fucked up somewhere, and I need to consider between more epochs and switching some parameters around.

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