The Changelog: Software Development, Open Source - Open source LLMs are catching up (News)

Episode Date: December 11, 2023

A group of researchers set out to test claims that its open source rivals had achieved parity (or even better) with ChatGPT on certain tasks, Richard Hipp and his team have rewritten SQLite's text-based JSON functions, Ratatui is a Rust crate for cooking up TUIs, Morris Brodersen built a complex app in vanilla JS as a case study & Headscale is Kristoffer Dalby's open source, self-hosted implementation of the Tailscale control server.

What up, nerds? I'm Jared, and this is ChangeLog News for the week of Monday, December 11th, 2023. Thank you to all of our listeners who left us voicemails for next week's State of the Log episode. Our inbox now runneth over, and we even have BMC cooking up some special beats for the occasion. That plus poundifying round two coming up this week. Ooh boy, we are finishing up 2023 with a bang. Okay, let's get into this week's news. One year after ChatGPT brought a seismic shift in the entire landscape of AI,

a group of researchers set out to test claims that its open source rivals had achieved parity or even better on certain tasks. In the linked paper, they provide an exhaustive overview of this success, surveying all tasks where an open source LLM has claimed to be on par or better than ChatGPT. Their conclusion, quote, In this survey, we deliver a systematical review on high-performing open-source LLMs that surpass or catch up with

ChatGPT in various task domains. In addition, we provide insights, analysis, and potential issues of open-source LLMs. We believe that this survey sheds light on promising directions of open- source LLMs and will serve to inspire further research and development,

helping to close the gap with their paying counterparts. End quote. It's becoming increasingly clear to me that the data models powering future AI rollouts will be commoditized and democratized, thanks to the competitive nature and hard work of both academia and industry. What a relief. Richard Hipp and his team have rewritten SQLite's text-based JSON functions using a binary representation, much like Postgres did years ago. This new JSONB, depending on usage patterns, could be several times faster than the original.

The technical explainer is, contiguous byte array that can read or be written as an SQL blob. This allows the internal use representation of JSON to potentially be saved to the database in place of JSON text, eliminating a bunch of overhead. End quote. All legacy functionalities preserved, the JSON functions that previously only accepted JSON text now accept JSONB binary for the same parameter. This looks like a huge step forward for everyone's favorite

and the most widely deployed in the world, SQL Engine. Phenomenal cosmic powers! that provides a set of widgets and utilities to build complex Rust terminal user interfaces, or TUIs. Rattatui is based on the principle of immediate rendering with intermediate buffers. This means that for each frame, your app must render all widgets that are supposed to be part of the UI. This in contrast to the retained mode style of rendering where widgets are updated and then automatically redrawn on the next frame. Rattatille was forked from 2ERS in 2023 in order to continue the project's development. Its outputted 2Es look really cool, but let's face it, I included it here because I love the name Ratatouille.

Build your 2Es with a Ratatouille. Cook up a 2E, Ratatouille. Something for you-y by Ratatouille. Okay, stop, Jared, just stop. You know what time it is. Sponsored news. If you had a list of commands that your engineering team uses daily, npm would likely be near the top. Shouldn't that crucial command be secure? Well, now it can be, thanks to Socket's SafeNPM CLI tool that transparently wraps the NPM command and protects developers from malware, typosquats, install scripts, protestware, telemetry, and more.

I was very impressed by the team and tech behind this tool when we first heard about it on JS Party. If you haven't given Socket's SafeNPM a good look, you owe it to yourself to follow the link in your show notes and chapter data and to check it out. Thanks once again to Firas and all our friends at Socket for sponsoring ChangeLog News. Morris Broderson built a highly interactive, complex application, a clone of the ToDo app called ToDo. That's T-E-U-X-D-E-U-X. Using vanilla HTML, CSS, and JavaScript, he did it as a case study attempting to find and stretch the limitations of vanilla web development. Morris says, quote,

there's no custom framework invented here. Instead, the case study was designed to discover minimum viable patterns that are truly vanilla. The result is maintainable, albeit verbose, If anything, the case study validates the value of build steps and frameworks, but also demonstrates that standard web technologies can be used effectively, and there are only a few critical areas where a vanilla approach is clearly inferior. End quote. The code is all open source, but Morris's write-up and its conclusion in particular is so thorough and thoughtful that you might be satisfied just by reading that alone. If you love Tailscale's offerings, but prefer the maximum privacy and

autonomy that only comes by self-hosting everything, check out this project by Christopher Dalby, who works at Tailscale, by the way. Christopher says, quote, Everything in Tailscale is open source except the GUI clients for proprietary OSs and the control server. The control server works as an exchange point of WireGuard public keys for the nodes in the Tailscale network. It assigns the IP addresses of the clients, creates the boundaries between each user, enables sharing machines between users, and exposes the advertised routes of your nodes. End quote. with an open source server they can use for their projects and labs. It implements a narrow scope, which is a single tail net, suitable for personal use or a small open source organization. That's the news for now, but also scan the email newsletter for even more stories,

such as Ars Technica pitting Google's new Gemini-powered Bard against ChatGPT, Dan Hockenmeier's proposed hierarchy of needs to use when choosing which startup to work for, and The Verge's coverage of a new low-level internet protocol called L4S. If you don't get ChangeLog news in your inbox each Monday, fix that bug at changelog.com slash news. We have some awesome podcasts coming up this week. Our final All Things open anthology ships out

wednesday and we're playing pound to find again with the same wacky cast of characters plus one on friday have a great week tell your friends about the changelog if you dig it and i'll talk to you again real soon