Probably AI spend and increased developer cost (salary + AI subscription cost). And the increase in memory prices means procuring IT assets has sky-rocketed too.
I would assume the margins they operated on, evaporated with the AI bubble.
Basically compared to standard LLM models it is an order of magnitude cheaper and faster. (Note: I didn't get to actually try jev yet, just looked at demos/specs/pricing etc)
You can definitely do similar things with say hosted LLMs + a lib like outlines , or with API models and the proper output validation layer, but again way slower and more expensive.
And on the opposite side you can train dedicated classification models that will be even cheaper and faster to run than jev. But, well, you need to train them (costly, time consuming, and data might be hard to come by depending on target). Here you are a nice zero-short system, that can handle complex/messy data out of the box.
Thanks. How reliable is the world knowledge? Say, I use it as a delegation router for picking the best model for a task - how can I be confident it's doing that with enough intelligence?
When I use traditional LLMs I get some sense from the flagship-ness and regular usage. How do we get such confidence for Jev like models?
The Jev alternative, laya is only 334m parameters and you can post train it as the workflow is open source. The downside is that laya is only 1024 tokens, and Jev is ~32,000 tokens. I think for many tasks 32k tokens might be too small. You could probably design a task complexity router with something like qwen 0.8b with similar speed and a much larger context window (260k token). As always, "it depends". 0.8B is almost as fast as Jev without any of the limitations.
This is great. Curious: this requires users to have a host app installed on their machine that can read .capsule file right? Won't that be a point of friction for distribution?
Isn't html/css a better distribution mechanism as most computers already have the tech to run them?
You'd think so, but web browsers are downright lobotomizing the ability for local HTML files to run any meaningful javascript code. Want to run a JS file from the same directory? CORS ERROR!!!!
Hence 200MB applications that are complete copies of the Chromium browser to run under 1MB of actual web code.
couldn't you just embed the js directly into the html file though? that would sidestep the CORS error. and since most of the people using a program like this are likely to be using AI to generate their ( likely throwaway ) applets anyway, readability and maintainability for the source doesn't matter as much as it would with human authors so having a separate js file doesn't bring that many advantages anyway.
Sharing HTML files with embedded Javascript is problematic: iOS users can't open the files directly, and have to save the file and then 'open with ... '.
Isn't that the price of a relatively well sandboxed experience by browsers?
I've experienced the same limits building small tools for myself and friends that I wanted to contain in a single HTML file without external dependencies... but I understand why the same project without these restrictions could easily become malware, and ones that could be easily propagated.
I wish the browser would either allow local-only or internet-only, then only make up its mind after the first attempted request. Fetch an image from the internet? No script access for local files. Fetch a JS from elsewhere on the hard drive? No internet access. That could provide the sandbox while still being flexible.
The problem is that you can't save the user data in way that you can click a single file, similar to a Word or Excel file. I you think a about it, a Excel spreadsheet is also just UI + data in a single file.
This is a frustration of mine as well, but I vaguely remember someone showing something on HN where the file could essentially save itself, I think they were using some file APIs for that... sadly I can't find the details anymore. But I keep running into the "HTML file opened locally can't update itself" issue repeatedly now that I build tools for myself with LLM agents.
The File System Access API can do this, and it works in local ("file:///") HTML files, but it's only currently supported in desktop Chromium browsers: https://caniuse.com/native-filesystem-api
However, every browser will let you download a new version of the HTML file and save it over the old one - yes, even from a local HTML file:
I'm part of the Zoho Writer team (Google docs alternative). And the new architecture is very similar to Zoho Writer i.e edits represented as a sequence of retains/keeps followed by action.
Also the decision to forego browser's selection and draw a custom layer.
All this makes reasoning about changes a lot more saner and intuitive. I've always wondered why prosemirror's transactions & steps couldn't be simplified further so I'm one vote up for the new design direction!
The person who got inspired by prosemirror and built lexical is no more at meta. Go to the repository and you'll find it's already in maintenance mode. Its promise of cross platform extensible editor is unrealised yet.
We used Lexical intensively in my company (Crisp), and I can confirm it's definitely not in maintenance mode. What's super cool with Lexical is the package/extension system. For instance, we use the Lexical markdown package a lot; it's not the gold standard yet, but since we have a large user base, over the years we've made a few PRs to the Lexical team, and they approved/reviewed the PRs in 24 hours.
About WordGuard:
There is no community behind it; the guy building it clearly wants no AI involved and even has his own internal Git hosting system.
Finally, there is no single user on WordGuard.
I see virtually no reason to use Wordgard over lexical.
This is amusing as well as slightly scary. Prosemirror over the years has become the backbone of editors on the web. ChatGPT uses it. Gemini uses it. Linked chat and feed composers are Prosemirror powered. Literally every serious product uses Prosemirror for composing text on the web.
There is an entire YC company built on top of Prosemirror (Tiptap).
The thought that Prosemirror is no more in active development is scary.
> The thought that Prosemirror is no more in active development is scary.
Which is a thing that I never said. In fact I went out of my way in all the announcements to stress that ProseMirror maintenance is continuing as it has been.
The only scary thought is that all these important operations don't have a plan B if prosemirror, a software they don't write, don't own and presumably don't pay, goes away.
Also MIT-licensed software doesn't 'go away' all that easily. If a piano fell on my head tomorrow, interested parties would probably manage to continue development in a fork.
Takes me back to the time where I would pickup any books that taught anything about computers/programming. I was too young to understand any of it, but the fact that there's this knowledge to be unlocked, that would then let me do new things with this white boxy device on my table top - was truly addictive.
I can't say the feeling remains because most new stuff that I learn now is not runnable on my machine now. Tech has gone back to the days when owning hardware that's capable of running the latest has become a costly affair.
Just checked out VMPrint and it's crazy! Keep up the efforts. If you/someone could get a HTML/CSS input layer in front of VMPrint that would be a killer feature? Or is it possible already?
Thank you. The architecture to get an HTML/CSS input layer in front of VMPrint is already there via a component called Transmuter. I have made transmuters to allow markdown with custom syntax support as input, and they work extremely well. Full support of HTML/CSS will be still be difficult, but a well-defined subset should be easy. The sibling Layoutmaster project also demonstrates this -- you can hand it a DOM element and it automatically grabs the styles and contents, converts them into the JSON AST, then feeds them to the engine for instant layout.
Basic text styles are ok, but things like authored pagination, page header/footer, mirrored margins, margin notes, footnotes and references are basically unsupported or need to be hacked together.
reply