lisperforlife

I call this the bluebox problem. Back when I worked for SAP there used to be this control called the bluebox control that embedded a browser (shdocvw) into a C++/MFC app. Almost every single annoying UI bug came from that control. I see the same now in the other direction. Every 5 years we come up with the bright idea to use browser as a distribution engine for a UI surface that is rendered by a box. This has historically never worked. I just wanted to recount applets, activex controls, flash, flex (also flash), silverlight, flutter ... and now this. The issue is developer tooling, screen readers, browser extensions and all the other small things that are difficult to account for.

My guess is that it is a massive model similar to GPT 4.5 and $10/$50 pricing is for its output will discourage people from using it. I also read safety = nerfed.

This is really cool. PG has zlib compression on TOAST objects so this should still be okay even if you are not storing pack files. I am curious with your choice of hand-rolling pktline, upload-pack and receive pack implementations including rev-walking. Any particular reason you did not want to use libgit2 or something like the gitoxide implementation of pkt-line. Was it performance or is it because you wanted it to be in pure rust? Did you try running this on slightly heavier repository with a lot of commits, refs and objects?

Exactly this! Classic open source bait and switch.

Recently the v8 rust library changed it from mutable handle scopes to pinned scopes. A fairly simple change that I even put in my CLAUDE.md file. But it still generates methods with HandleScope's and then says... oh I have a different scope and goes on a random walk refactoring completely unrelated parts of the code. All the while Opus 4.5 burns through tokens. Things work great as long as you are testing on the training set. But that said, it is absolutely brilliant with React and Typescript.

The 5ms write latency and 1ms write latency sounds like they are using S3 to store and retrieve data with some local cache. My guess is a S3 based block storage exposed as a network block device. S3 supports compare-and-swap operations (Put-If-Match), so you can do a copy-on-write scenario quite easily. May be somebody from TigerData can give a little bit more insight into this. I know slatedb supports S3 as a backend for their key-value store. We can build a block device abstraction using that.

I think you can get much farther with dedicated servers. I run a couple of nodes on Hetzner. The performance you get from a dedicated machine even if it is a 3 year old machine that you can get on server auction is absolutely bonkers and cannot be compared to VMs. The thing is that most of the server hardware is focused towards high core count, low clock speed processors that optimize for I/O rather than compute. It is overprovisioned by all cloud providers. Even the I/O part of the disk is crazy. It uses all sorts of shenanigans to get a drive that sitting on a NAS and emulating a local disk. Most startups do not need the hyper virtualized, NAS based drive. You can go much farther and much more cost-effectively with dedicated server rentals from Hetzner. I would love to know if they are any north-american (particularly canadian) companies that can compete with price and the quality of service like Hetzner. I know of OVH but I would love to know others in the same space.

I don't think models are fundamentally getting better. What is happening is that we are increasing the training set, so when users use it, they are essentially testing on the training set and find that it fits their data and expectations really well. However, the moat is primarily the training data, and that is very hard to protect as the same data can be synthesized with these models. There is more innovation surrounding serving strategies and infrastructure than in the fundamental model architectures.

You can use libkrun to pretty much do the same thing.

Why is this not the top comment? FAIR published their C3MLeon paper about decoder-only autoregressive models that work with both text and image tokens. I believe GPT-4o's vocabulary has room for both image and audio tokens. For audio tokens, they probably trained an RVQ-VAE model like Encodec or Soundstream.

I am curious about models like encodec or soundstream. They are essentially meant to be codecs informed by the music they are meant to compress to achieve insane compression ratios. The decompression process is indeed generative since a part of the information that is meant to be decoded is in the decoder weights. Does that pass the smell test from a copyright law's perspective? I believe such a decoder model is powering gpt-4o's audio decoding.