#83
Waymo Expansion, Flutter Fork Flock, Claude Code Execution, Sarvam AI Release, GitHub Copilot Update, Celestia's Ginger, Robinhood's Election Bets, Based Agents, SQLite rsync, 8b Llama, SynthID & more
👋🏻 Welcome to the 83rd!
🪔 Have a happy festivity week ahead!
🤷🏻♂️ Not sure how we moved to mid-week shipping.
📰 Read #83 on Substack for the best formatting
🎧 Here’s a link to the engaging podcast version of this edition → NotebookLM
What’s happening 📰
🚘 Waymo raises $5.6B in Series C from Alphabet (parent), a16z, Silver Lake, and more. Yup! it’s not a typo, it’s a “B”. Waymo has now completed over 100k rides and servicing mainly in San Francisco, Los Angeles, and Phoenix, and is expanding to Austin and Atlanta.
🍴Frustrated about the Flutter team's limited labor and slow response times, the community has created a fork of Flutter called Flock. They hope that this will bring overall improvements and unlock the potential of Flutter.
✨ AGI Digest
⚓️ Model Drops:
🪽India’s Sarvam AI released Sarvam-1, a 2B model specially designed for Indian languages. It is trained on a custom tokenizer and 2T Indic tokens covering 10 Indian Languages along with English. In the small model range (< 4B) and even when compared against the Llama 2.1 8B, Sarvam 1 achieves SoTA performance in Indian Languages.
Source: Sarvm Blog 🇮🇳 Plus, we also got the Nvidia Nemotron-4-Mini-Hindi-4B in Base and Instruct variants during the Nvidia Summit in Mumbai. These 4B models were made from the previously-released Nemotron-Mini-4B-Base which was subjected to continuous pre-training on an additional corpus of 400B tokens that makes it better on tasks in Hindi, English, and Hinglish.
🍂 Talking of multilingualism, Cohere released Aya-Expanse in 8B and 32B sizes which excel across 23 languages with the 32B model achieving SoTA performance against other leading open-weight models. They attribute these improvements to their multilingual data arbitrage and techniques for merging multilingual models better.
Source: Cohere Bog 🦙 Meta released quantized versions of Llama-3.1 1B and 3B models which were done via Quantization-Aware Training using QLoRA. They also released the models using SpinQuant to make it easier to do post-training quantization in other quant sizes while ensuring much better model quality. These make deploying these already small models much easier on even more edge devices.
🗂️ Product Updates:
💻 In GitHub Universe last week, GitHub announced support for the new o1-mini and o1-preview in GitHub Copilot, starting with Copilot Chat with support for Claude 3.5 Sonnet’s exceptional coding abilities and Gemini 1.5 Pro’s 2M context length coming along the way. This was a much-needed facelift for GitHub Copilot which was losing its charm against ursor these days. Though it still doesn’t match head-to-head with the other features that Cursor has to offer, being ready to embrace other model providers is a welcome step.
Also, taking hints from Vercel’s v0 and Claude’s Artifacts, folks at GitHub Next also demoed GitHub Spark, an AI-powered tool for creating and sharing “micro apps” which allows you to create small independent runnable apps with just a simple description.👨🏻💻 Anthropic folks didn’t stop shipping this week as well and introduced an analysis tool in Claude Chat, an answer to OpenAI’s Code Interpreter. This tool uses JavaScript (not Python, unfortunately) to run small analyses in the sandboxed environment in the user’s browser that can do small calculations which LLMs usually fail at.
🗣️ Eleven Labs Voice Design through which you can generate a unique voice for a character from a text prompt alone, be it a raspy orge or a sweet-talking fairy or a hideous monster — your imagination is the limit.
💬 Perplexity now has a MacOS app, allowing you to ask anything with ⌘ + ⇧ + P. And, ChatGPT’s Advanced Voice mode is now available on the MacOS as well as Windows apps, while also reducing the API costs of the Realtime API.
🔬 Research Updates from big labs:
🍎 With the rollout of the new iOS 18.1 and MacOS 15.1 this week, Apple has introduced a lot of Apple Intelligence features. And since Apple puts such a great emphasis on “privacy”, certain features like Enhanced Visual Search and other Private Information Retrieval (PIR) that need access to private information had to be done right. Addressing these concerns, Apple put up a blog talking about how they combine Homomorphic Encryption with ML to improve privacy while improving quality across their services.
🪪 Google DeepMind’s SynthID, which they use to watermark AI-generated text, music, images and videos and detect AI-generated content has been open-sourced (only the text part though). The folks also discussed in a separate Nature paper their process of generating and detecting AI-generated text using this technique and in their code demonstrated its implementation on the Gemma and GPT-2 models.
Source: Google DeepMind’s Nature paper 🏞️ OAI demonstrated a continuous-time consistency called sCM, a new image-generation approach which can generate samples with quality comparable to diffusion models using only two sampling steps. Unlike diffusion models, which generate samples gradually through a large number of denoising steps, the consistency models aim to convert noise directly into noise-free samples in a single step.s
📖 Microsoft released OmniParser, a tool for parsing user interface screenshots into structured elements, which significantly improves generating structured labels accurately grounded in the corresponding regions of the interface. Through they compare various models, we’d have loved to see the new Claude 3.5 Sonnet and the Qwen2-VL in the comparison to accurately figure out how distant from the current SoTA models they are.
🔐 0x Digest
🤖 Folks at Coinbase were getting bored and saw the news of OpenAI releasing swarm, they tinkered and built an experimental playground for autonomous on-chain interactions. They called it “Based Agent”, with Replit you can set up this in 3 minutes. This agent allows you to automatically execute actions, deploy tokens, request from the faucet, and more.
🫚 Celestia’s Ginger upgrade coming in November, decreasing block times from 12s to 6s, this unlocks major UX improvements and sets the stage for a bigger block size. Just like the previous upgrade this one also implements 5 CIPs.
📜 Scroll announces drand VRF is now available for all EVM chains which is a public good by Scroll. This makes free verifiable randomness more accessible, allowing you all to freely roll a dice.
🗳️ Robinhood brings Presidential election contracts to its customers, making betting on elections accessible to all. 😏 (riding on Polymarket’s PMF eh!)
😵💫 A hacker drained $20M worth of assets from a wallet that was probably owned by the US Government, which they funnily used to store 12,000 stolen BTC that were seized after the Bitfinex hack in 2016. (hacked → seized → drained, 1 month before 5-year sentence of original hackers.)
👻 Solana’s favorite Phantom faced another downtime (3rd this year), during the GRASS airdrop.
💵 Denmark plans to propose taxing unrealized crypto gains in upcoming bills, this is worse than putting a crypto ban, it’s a full-on war now.
💼 Consensys announced it will lay off over 160 employees (20% workforce) ongoing legal battles with the SEC as one of the key factors.
🛠️ Dev & Design Digest
🆕 Svelte 5 went live last week after 18 months of development efforts. What’s new and different this time? The compile-time reactivity which made it different from React, turned out to be causing issues when developing complex applications, so in v5, they introduced runes, an explicit mechanism for declaring a reactive state, it uses signals under the hood, keeping the good DX on the table. And yes even with all this it’s backwards-compatible with Svelte 4.
Fun fact: ALL OF THESE FRAMEWORKS ARE CONVERGING.▼ Netlify joins OpenNext to support Cloudflare and SST, ensuring Next.js runs well everywhere. OpenNext aims to increase Next.js's portability and flexibility across different providers. It seems the Vercel monopoly bothers everyone. Vercel, being the creator, had the chance to do optimizations before anyone, but with competitors catching up, it’ll be a big win for developers. On that note, Seb from NextJS wrote about their Journey with Caching.
🐰 Jared Sumner in his quest to optimize JS runtime (Bun) revealed event emitters are one major performance bottleneck and how Bun manages to not use that. And he keeps teasing the stuff coming in the next version bun like Bun.randomUUIDv7() that returns UUID v7, which is a sorting & database-friendly UUID.
🔄 sqlite_rsync is now available allowing efficient backups over SSH, the initial prototype was added back in September, and seems like it’s ready now. The most asked and important question is can I have connections while this utility is running? Duh! Of course, quoting from the feature notes: “Programs can have active connections to the databases on either end while this utility is running”.
😼 NestJS released Mau, a new service that provisions and manages your infrastructure on AWS without the hassle and extra DevOps work.
What brings us to awe 😳
🔢 The Fibonacci series was first described and explained by Acharya Pingala, an Indian Mathematician and Poet, in his work called Chandaḥśāstra which dates back to 300 BCE. (ofc they weren’t explicit about it)
⚒️ Simmon Wilson (the creator of Django) shared how he built 15 projects using Claude Artifacts in a week and has provided steps, prompts, and a full transcript for all of them.
⛓️ JavaScript’s setTimeout breaks after ~25 days. So, Evan built setBigTimeout, a package that just chains those setTimeouts to allow bigger timeouts.
Today I (we) Learnt 📑
⚾ Famous “Foo” and “Bar” in code examples have some serious origins, they come from the US Army WWII acronym FUBAR, "F-ed Up Beyond All Recognition".
🥊 The reordered phrase “freeze, flight, fight, or fright” is a more complete and nuanced alternative to the famous one “fight or flight”. Fight or flight coined by physiologist Walter Cannon doesn’t cover the whole spectrum of options. There is a sketchplanation for this.
🏎️ Until iOS 17.4, all apps were only allowed to use Apple’s WebKit as a browser engine, so technically all iOS browsers were (and most still are) just Safari forks with a different skin.
🌞 Some of us let our shower thoughts out in the open on Astronomy Substack and someone who spent time thinking about “Why does Jupiter spin so fast but the Sun doesn’t?” sparked a discussion where people enlightened them and us that the Solar Wind and Sunlight come at the cost of angular velocity and that sun is facing a mid-life crisis.
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying 🔖
🖥️ lcl.host: easiest and most functional way to run localhost https servers for free as powered by Anchor.
👨🏻💻 Vim for React Devs: A bite-sized course to get you quickly productive with Vim. This is the course file.
💤 zzzuckerberg: One of Legal Lullabies, soothing white noise made with Instagram's and TikTok's terms of service (they put us to sleep).
🔠 Best Free Fonts: a curated selection of 172 free fonts. Including serif, sans serif, script, and monospace.
Builders’ Nest 🛠️
🖇️ dead-or-alive: a npm package to check if URLs are dead or alive. (atwood peaks every day)
🫣 is-my-node-vulnerable: a package that checks if your Node.js installation is vulnerable to known security vulnerabilities.
📃 route-list: CLI tool to show Express, Koa, Hapi, or Fastify routes.
🤖 agent.exe: an Electron app that lets Claude control your computer leveraging the new beta capability “Computer Use”. A project with a name ending in .exe and which lets AI control my computer sounds completely safe.
Meme of the week 😌
Off-topic reads/watches 🧗
‼️ Important Problems by Seth, breaks problems into Important, Trivial, Unlikely, and Tempting, talking about which is worth solving.
🫵🏻 Version 1 is for you by Jason Fried, sharing how he regretted sharing all these alternate use cases launching v1 of a product and distracting the team and why you shouldn’t do that.
🎓 What "Follow Your Dreams" Misses by Grant Sanderson on Harvey Mudd's Commencement earlier this year. (ngl gives very much Stanford 2005 vibes, he narrated stories like Steve)
🦵 This is a movie recommendation for a Crime-Horror Thriller, LongLegs (2024), that Nibbler P watched recently. Apart from the story that keeps you on the edge, it makes ample use of color contrasts, match cuts, and sound bridges adding to the overall suspense and feel of the movie.
Wisdom Bits 👀
“It is not the answer that enlightens, but the question.”
— Eugene Ionesco
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev
Weekly Standup 🫠
Nibbler P went for a drive with family, did lots of shopping, painted stuff, and is in a completely festive mood thinking about all the tasty food he’d have by the time this edition hits your inbox.
Nibbler A is back home now, traveled, and watched a few movies on the flights, he’s been busy wrapping up some work before the festive season begins. He is playing around with Grok and Apple Intelligence (Siri that can do 5 more things).
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻