According to the linked PR, the original model does come with MTP which is another "head" (=output path) in the same model and (supposedly) runs faster.
The current implementation ignores that head but the PR let the tool recognize it, plus does proper integration (run the MTP while running the slower main path then compare the result, I believe.)
The standard way of doing MTP is to run the drafter autoregressively for k steps, and then (not concurrently) use the larger model as a verifier for those k tokens at the same time. The larger model can then accept a prefix of those k tokens, and in any case generates one more token (which is needed in case you accepted zero tokens from the drafter). The larger model can effectively use this k as a "batch" dimension, reducing the penalty of large weight loading. Meanwhile the drafter is much smaller, so it's fine for _it_ to be autoregressive, as long as the main model is parallel.
I really think there is something here. It's a very low level general primitive, and you're right, OOP has been around for a long time. Objects and protocols, e.g. in objective C.
If you want, I'm happy to jump on a zoom and talk. You can use calendly.com/safebotsai
There's a well known theory on this exact concept! The Medium is the Message. Or, the very act of defacing a public building is meant to sledge-hammer the artist's work into the viewer's consciousness. Compared to say, some quiet exhibit most people would never encounter.
You are not supposed to get any attention and you are not supposed to have any say in how the city and the world looks. If you buy the building you still don't get to paint.
I hope HN can appreciate what a game changer (and paradigm shift) Zed can be.
To the Zed developers: CONGRATULATIONS! I have been following your project with great interest since your speed demo years ago. And since it’s AI-first, I’m interested to see how we can integrate it with https://safebots.ai (Safebots, Safebox, and Safecloud).
I would love to see how we might be able to increase the safety of agents in Zed, use local models like Qwen/Deepseek and we also have Grokers which can turn any codebase into a graph with tree-sitter and help your agents far more than RAG and similarity search (https://grokers.ai/deck.pdf)
What’s the best way of getting in touch? (If you want, my profile has a way of emailing me).
The thing is, technology is either enabling something or not. The exploration space might be huge, but once an exploit is found, the exploitation code / strategy / plan can trivially proceed and be shared worldwide. So you have to deal with this when you design and patch systems.
Example: preserving paths in URLs. Safari ITP aggressively removes “utm_” and other well-known querystring parameters even in links clicked from email. Well, it is trivial to embed it in a path instead, so that first-party websites can track attribution, eg for campaign perfomance or email verification links etc. In theory, Apple and Mozilla could actually play a cat-and-mouse game with links across all their users and actually remove high-entropy path segments or confuse websites so much that they give up on all attribution. Browser makers or email client makers or messenger makers could argue that users don’t want to have attribution of their link clicks tracked silently without their permission. They could then say if users really wanted, they could manually enter a code (assisted by the OS or browser) into a website, or simply provide interactive permission of being tracked after clicking a link, otherwise the website will receive some dummy results and break. Where is the line after all?
reply