Since I started using LLMs, I've always had this feeling of wanting to go back, change the history, and then go down another branch.

The lightbulb moment for me was when Google AI Studio started to let you actually edit the AI's responses. I realized, YES, I need a multiverse type of situation to branch all my thought streams and variants of the same thread. I wanted to try Gemini and ChatGPT in parallel on the same thought.

# Idea

The original concept for Alyph goes back to 2024. I had some early concepts that were a lot more sci-fi (if you look at the codebase, you’ll see names like Stargate, Hermes, etc.).

I even thought about making it 3D, like a holo deck, where you visit "planets" of topics and content. But the execution now is much more, eh, pragmatic.

# Alyph today

Alyph is a 2D canvas where you can ask LLMs stuff in parallel. When you go down a rabbit hole and finally get an answer (be it code or text), you can pull that answer up and delete all the unnecessary back-and-forth in between.

The app also has a multiplayer mode where you can do that together with others on the same board. (see demo board below)

# Background

The core problem with standard chat UIs, as I see it, is context poisoning.

If you look into how LLMs actually work, you realize they have zero memory. They are a very fancy autocomplete on your phone that (re)reads the entire complete chat from scratch with every single message you send.

This means: all your wrong turns, the insults (oops), the failed debugging loops, and the bad rabbit holes get sent with every subsequent message. Your complete history is used to predict the future. That doesn’t help when you just want a clean FaqSection.tsx.

I noticed that with the more powerful models, the choice of words when prompting isn't as important as the context that goes with it. At least it isn't for my use cases. For example, if you give the LLM a codebase with a very high-quality threshold (fully type-checked, clear front end components, strict style guide in a mainstream language) they tend to produce really high-quality stuff very easily.

Furthermore, if you give models too much context, the answers tend to degrade. Claude, for example, is amazing under 300k-500k tokens, but above that, it can degrade very quickly (whereas Gemini tends to perform a bit better with massive context, at least in my experience).

So what do you do? The mainstream answer right now is to let an algorithm try to pick and choose the context for you.

But: My strong assumption is that models will keep getting bigger and bigger context windows — we started at what, 32K for consumer models, and now Grok is at 2M these days. How would the world change if we had models with 100M context? 100B?

That's why I think (heresy!) harnesses, context compacting algorithms, agents etc. are a stop gap right now. They don't really work well in my cases (as they sometimes forget important things and go off on a tangent to produce slop), and it really feels like a black box with me as the meat proxy pressing "accept" or "reject". I'm a distant manager. I don't want to be a distant manager. I want to be in the weeds.

So what I built is, what some might say, is an anachronistic manual tool. A manual transmission, so to speak. I want to have 100% control over my baseline. I want to put in exactly what I want the model to see, experiment, and just dump a folder into it (Alyph btw respects your `.gitignore` file, which is neat). With the assumption that models will eventually allow even bigger context windows, having manual control over that baseline is important, I think.

Alyph still has some rough edges, of course, but it's a labor of love. I would love to share it to see if it actually works for others.

If you hate it, tell me. If you love it, tell me too. I'm also looking for somebody to team up with, so hit me up.

Live at: https://alyph.ai, it's free to try.

Here's a demo board for YC:

https://my.alyph.app/board/d806842f-6d52-45a4-8bb6-da46e95e5...

(no login needed)

Show HN: Alyph, a manual transmission for LLM context