Building a new way to reason with LLMs (we're also paying contributors to the repo)

Training reasoning models is really expensive and I had a suspicion that there was a lot of performance to be gained by exploring the models states better.

I’ve open-sourced a lightweight framework for latent-space reasoning, and the results have been more interesting than expected. With no fine-tuning and no access to logits, it consistently outperforms baseline outputs across a range of tasks just by evolving the model’s internal hidden state before decoding (including being able to solve problems that the base model struggles with). This uses a minimally trained judge (200 samples on a simple scorer; cost less than 50 cents to do completely) and preexisting models with no other tuning.

It works with any HF model, and the entire pipeline is intentionally simple so people can tear it apart, extend it, or replace pieces with better ideas. I’m putting up bounties for improvements because the goal here isn’t to claim we’ve solved reasoning, but to build a shared playground for exploring it. We're already collaborating with researchers in 2 of the top 5 AI Labs in the world to extend this with more sophisticated mechanisms (especially around aggregation and projections) but would love to have you guys in as well.


Let's make sure the new generation of reasoning is open source--

https://github.com/dl1683/Latent-Space-Reasoning

https://redd.it/1pimeg1
@r_opensource

How to get started with open source as a new CS grad?

Hey what's up y'all. I just graduated with a undergrad in CS and have been working as a software engineer at a mature tech company for about 6 months. I've learned quite a lot about how large scale applications and services are built and engineered, and I'm very appreciative of it.

However I'm soon going to a different company (better pay + standby flight benefits) where I'll work as a data engineer, but the actual engineering is much weaker there, and the projects I work on will be smaller scale and internal. I'll also be more accountable for my own work so I won't really have much senior help in engineering and designing of solutions.

But I still want to become a better software engineer overall as I see myself eventually going back into big tech/AI or quant (I'm doing a masters degree in ML, have undergrad degrees in applied math and CS).

I think the best way to hone my skills at that point is to become an open source contributer to well maintained projects, but I honestly don't know where to start. Just picking up issues, or reading forums all seems so daunting and hard to even begin.

For starters, my biggest problem is understanding large codebases. At my current job, I eventually understood mine better due to extensive architecture notes and just working on stuff for 40 hours a week. Obviously I wont have that same time or support level in open source software. GPT makes it easier to get started and reason about a codebase, but past that, it's still hard to work on software I'm not familiar with at all, my current job is my first experience with that, and its about to end :(


Second is the long term motivation. I think my job is very interesting, and the product I'm working on applies the concepts I learned in college very well, but ultimately I'm still doing it for the salary. I have a lot of hobbies outside of work, and staying motivated to stick to a project long term, for free, may be an issue. I dont know if that means this type of work just isn't for me, but I'd appreciate tips on how to actually stay committed to this stuff for no extrinsic reward.

https://redd.it/1pimwc8
@r_opensource

Async web scraping framework on top of Rust
https://github.com/BitingSnakes/silkworm

https://redd.it/1piodn1
@r_opensource

Is there an opensource dataset/app that shows national factory farms?

Im thinking of creating a dataset of U.S. factory farms since there isnt any good dataset or website that shows that so far from what Ive seen. But before I start I was wondering if anyone knew of one already?

If I end up making one then it would be completely opensource and would make a website displaying that information on a map.

https://redd.it/1pitree
@r_opensource

Releasing AnthroHeart: A Public-Domain Animation Project (Seeking Hosts for 8GB Bundle)
https://anthroentertainment.com/

https://redd.it/1pito7f
@r_opensource

Building a markdown based browser

Taking inspiration from my Kindle, I'm hobbling together a browser for hyperlinked markdown documents. I'm writing it in Python, and using Pyglet as the UI.

Why?

Honestly. . . I'm tired of getting online and having everything vying for my attention. I just want to read. To read documentation. To read news articles. To read blogs again, instead of Facebook.

Pages where I set the styling. And there aren't floating boxes everywhere. Where I'm not straining to see tiny Xs which need to be clicked with the precision of military marksman.

I'm tired of being fingerprinted and tracked from one domain to the next, like livestock.

I'm tired of a document standard so convoluted that Google's the only company capable of implementing it in its entirety.

What's your solution?

So, I'm combining the feel of a modern web browser with the simplicity of gopher, and a text styling somewhere in-between. Document-oriented formatting, like Kindle, where you can flow from page to page on a "website." Probably more like a webbook.

It doesn't block ads, but it shouldn't have to. Since most of its content will be in-line.

There is a query box at the end of the URL bar (think Firefox search box before they unified search and URL). Anything you enter into that box is appended to the end of the URL request as: ?q=query. Other than that, there's no other way to send information to the server. No headers. No cookies. Nothing.

What do you hope to accomplish

I don't plan to replace the web. More like. . . encourage people to blog again. Bring back directories (instead of search engines), where people can learn how to find their own information, instead of relying on what an AI tells them. Give documentation a space of its own. Encourage people to use other protocols to interact (email, FTP, Bittorrent). Lower server bandwidth requirements.

Basically, type out an email in Thunderbird to post to your blog, or post a classifieds listing.

My main goal is change how people use the web, from just logging onto Google and entering the information they want, to actually making them look for it and reason out how they got there.

So many people are asking Google for medical advice. Google is showing every single one of them custom tailored results. No one can tell what's real and what isn't. Whereas, if we went the card catalog (online directory) route, it'd actually force people to be aware of what they were doing and looking for. People wouldn't be zombies online anymore.

So. . .

1. Do you think anyone would actually use it?
2. Do you have any suggestions for it?

https://redd.it/1piveey
@r_opensource

Open Source Without Borders: Reflections from COSCon’25
https://opensource.org/blog/open-source-without-borders-reflections-from-coscon25

https://redd.it/1pix9az
@r_opensource

How (almost) any phone number can be tracked via WhatsApp & Signal
https://github.com/gommzystudio/device-activity-tracker

https://redd.it/1pizdi5
@r_opensource

Rephole - semantic code-search for your repos via REST API
https://github.com/twodHQ/rephole

https://redd.it/1pj0wfq
@r_opensource

How to Cultivate an Open-source Platform for learning Japanese from scratch
https://github.com/lingdojo/kana-dojo

https://redd.it/1pj1s1j
@r_opensource

AutoCAD LT Replacement?

I know this question has been asked multiple times, but I'd like an update from people that know about the advancements in the past few years, as well as what I'm looking for specifically. We use LT, which I believe is strictly 2D only, so no need for 3D. I believe the biggest thing we'd like are simplicity and similarity moving from AutoCAD LT in terms of UI layout and workflows. DXF and DWG support would be nice but I don't think it would be a deal breaker. I'm willing to pay for a perpetual license, but I'd like to stay away from adding subnoscriptions if possible.


I've seen people recommend FreeCAD, QCAD, LibreCAD, and nanoCAD. FreeCAD seems to have a focus on 3D which I don't believe we would need. I like the idea of QCAD having a one-time purchase perpetual license and having DXF/DWG support. LibreCAD seems to have a closer UI to AutoCAD LT? nanoCAD seems to mimic commands but it's subnoscription based. I know it would still be much cheaper than paying AutoDesk.

https://redd.it/1pj31pj
@r_opensource

I made a site that turns your GitHub history into a cinematic 2025 recap

Try : https://gitstory.pankajk.tech/
Repo : https://github.com/pankajkumardev/gitstory-2025

https://redd.it/1pj487y
@r_opensource

Chemical plant equipment cost database

Hi y'all. One of the major things that holds back green tech and climate tech hardware startups is making good estimates for the cost of their tech. This is my specialty.
One of the big reasons for the difficulty in making good estimates is the lack of good equipment cost data available. Large engineering firms keep this proprietary information in house.
I know it's a long shot, but does anyone know of a database of chemical plant equipment that is or could be open sourced?
I'm currently using notebooklm to search through textbooks but it's not ideal, especially as the data is very old.

https://redd.it/1pj6dyp
@r_opensource

Is there a real proof of an open weights AI model sending data to an external server

I'm wondering if there was an incident with a self hosted open weights AI model leaking data. I find it hard to believe that these companies are releasing their open weights models for free with no security threats.

https://redd.it/1pj6wfw
@r_opensource

Made a small project to turn images into pixel art using edge detection to preserves significant features
https://github.com/yogthos/pixel-mosaic

https://redd.it/1pj9fdn
@r_opensource

Luke Kenneth Casson Leighton

This dude was behind EOMA68 and Libre soc, a year now he seems to be in a really rough spot, his project funds got cut because of some unclear reasons maybe betrayal also he is in a very bad health, he is suffering with all kind of brain health problems strokes etc... ,i wonder how doesn't he get any attention or any help,

There is a whole drama in the libre-soc-dev Archives

https://redd.it/1pj50vp
@r_opensource

Portfolio and a Blog Template with a Dashboard Feel

I found the normal portfolios kinda boring, so made one that feels like a dashboard.

I tried to make it modular and relatively easy to customize, so y'all can try using it. If there's anything you guys don't like, roast me in the issues, I'd be happy to learn.

Has some whimsy text here & there, not sure if it's not too unprofessional, hah.

I genuinely hope it's useful

P.S. there is a page for writing blogs when ran in development mode

Sorry forgot to put a link https://github.com/AdiKsOnDev/dashboard?tab=readme-ov-file


https://redd.it/1pjca53
@r_opensource

Building an open source expense tracker that reads your bank emails. No bank login needed. Would you use it?

I hate tracking expenses manually. Tried apps, spreadsheets, everything. Always give up after a few weeks.

But here's the thing – my bank already emails me every time I spend money. Credit card charge? Email. Subnoscription? Email.

So I'm building an app that just reads these emails and tracks everything for me.

**What it does:**

You install a Chrome extension. It creates a filter in your Gmail that forwards only your bank emails to our app. We read those emails, pull out the amount, merchant, date, and categorize it automatically.

You get a dashboard showing where your money is going. That's it.

**What you don't do:**

* No typing expenses manually
* No connecting bank accounts
* No sharing any passwords
* No scanning receipts

**On security:**

The whole thing is open source. You can read every line of code and see exactly what we do with your data. We only see the specific bank notification emails that the filter sends us. Nothing else from your inbox. We grab the transaction details, then delete the email content.

If you don't trust our servers, self-host it.

**What I want to know:**

* Would you use this?
* Is the extension setup a dealbreaker or fine since it's one time?
* What would make this actually useful for you?

Building it for myself either way. Curious if others want the same thing.

https://redd.it/1pje5tz
@r_opensource

My first OSS project! Observability & Replay for AI agents

hey folks!! We just pushed our first OSS repo. The goal is to get dev feedback on our approach to observability and action replay.

How it works

Records complete execution traces (LLM calls, tool calls, prompts, configs).
Replays them deterministically (zero API cost for regression tests).
Gives you an Agent Regression Score (ARS) to quantify behavioral drift.
Auto-detects side effects (emails, writes, payments) and blocks them during replay.

Works with AgentExecutor and ReAct agents today. Framework-agnostic version coming soon.

Here is the -> repo

Would love your feedback , tell us what's missing? What would make this useful for your workflow?

Star it if you find it useful


https://github.com/Kurral/Kurralv3

https://redd.it/1pjg797
@r_opensource

I built a new open source to self-hosted Excalidraw on your own VPS, which focus on personal usage.

Recently I'm in love with Excalidraw, it helps a lot to showcase my idea and explain to my colleagues. The case is, I have multiple computers at work and at home, and want to be able to view/edit my drawings where I am.

Excalidraw Free only save the data in local storage, to have cloud storage you have to purchase Pro plan. But Pro plan also come with features I don't use, like collaborating, advanced component, present mode, etc...

So I think of self-hosting it in my own VPS. Excalidraw open source there core component in React, so just need to made a simple CRUD around it and you have what necessary for personal use.

Here is the oss repo: https://github.com/lukenguyen-me/personal-excalidraw

Looking forward to receiving feedback, issues, or any improvement you think of, focus on daily personal usage.

https://redd.it/1pjkz39
@r_opensource

ovr@6.0.0 - Streaming Fetch Based Multipart Uploads
https://ovrjs.com/06-multipart

https://redd.it/1pjkhgb
@r_opensource