#79
Meta goes high on vision, Deno won't go extinct, Arc's crazy vulnerability, cz is free, Hamiltonian path poems and more
👋🏻 Welcome to the 79th!
💪🏻 If you are looking to work on awesome projects and learn this Hacktoberfest, a community member is looking for contributors for a project, go check out metz.sh.
📰 Read #79 on Substack for the best formatting
What’s happening 📰
👓 Meta held their annual Connect 2024 where they released a new and cheaper Meta Quest 3S headset, a bunch of new Llama 3.2 models, a futuristic live translation using the Meta AI glasses and the showstealer — Meta Orion AR glasses.
📱 MKBHD got quite thrashed with his new subscription-based wallpaper app, “Panels”. People didn’t miss the chance to review the reviewer at all :P
✨ AGI Digest
🍪 Google DeepMind released the pre-trained checkpoints for AlphaChip, their RL model for ASIC chip floorplanning that helped design their v5 and v6 generations of TPU. The model checkpoint pre-trained on 20 TPU blocks would allow users to get good zero-shot performance right out of the box and reach convergence faster on training on their chip designs.
Source: Google DeepMind Blog ⚓️ Model Drops:
🤟 Jina AI released Jina Embeddings v3, a 570M multilingual embedding model with 8192-token-wide context length and supporting 89 languages. It outperforms OpenAI’s and Cohere’s proprietary embeddings on many benchmarks and overall stands at rank 2 in the <1B category on the MTEB embedding benchmark.
Source: Jina AI Blog 🎉 Qwen released Qwen2.5 family of models with QwenLM 2.5 (0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B), Qwen2.5-Coder (1.5B, 7B with 32B on the way) and Qwen2.5-Math (1.5B, 7B, and 72B), and typical of Qwen models, all of them are superior performing in their weight class and stand shoulder-to-shoulder with the newer llama-3.2 models.
👀 Allen AI released Molmo, a family of VLMs based on OLMo-7B, OLMoE-1B-7B, Qwen2-7B, and Qwen2-72B with the 72B model beating GPT-4o, Gemini 1.5 Pro and Claude 3.5 Sonnet and even the newly released LLama-3.2 90B on several academic vision benchmarks becoming the best open-source vision model. You can try the models on their site.
Source: Tweet by @soldani 🦙 Meta dropped Llama 3.2 models at the Meta Connect 2024, introducing 1B and 3B text-only LLMs plus 11B and 90B VLMs adding multi-modality to the Llama model family. They also released Llama Guard 3 11B Vision which can be used as a moderation model for vision tasks (making OpenAI update their moderation endpoint to also support images, shortly afterward). Along with this, Meta also released the Llama stack which provides a collection of APIs for the developers to start working with the Llama models from scratch, along with integrations for several platforms, especially edge devices like phones for the newer smaller models.
♊️ Google ships with Gemini:
🌲 Google upgraded Gemini 1.5 Pro and Flash with the new 002 checkpoints which show significant improvements in math, coding, and long-context and vision benchmarks, while also reducing price, increasing rate limits, and improving output speed and latencies. Their new OmegaPRM paper describes how they brought about some of these improvements in mathematical reasoning using a novel divide-and-conquer style MCTS.
Source: OmegaPRM paper 💬 Gemini makes the Gmail’s smart reply contextual. No more accidental “Sounds good to me!”
🎙️ Google’s NotebookLM (which is already getting quite popular because of its podcast-anything feature) adds support for adding YouTube videos directly from URLs and even audio recordings. Interestingly, a similar product from Google Labs called Illuminate does the same thing as NotebookLM — convert text content to engaging audio podcasts.
🎥 Google has a new product called Google Vids, which uses generative AI to generate professional-looking presentations and marketing videos for you complete with content, narration, images, stock videos, music, and transitions.
🌐 OpenAI released a 14-language multilingual translation of MMLU called Multilingual Massive Multitask Language Understanding (MMMLU). Not surprisingly, they did the translation using professional human translators and not through their AI models to ensure confidence in the accuracy of the translations. Goes on to prove that we are simply not there yet where we can offload all our work to AI and expect it to work.
Meanwhile, OpenAI has also been rolling out the Advanced voice mode to all ChatGPT Plus customers and increasing the limits of the o1 models. Also, following Meta’s Llama 3.2 multimodal models, they upgraded their free moderation API to support images as well.
🔐 0x Digest
⛓️ A new decentralized on-chain social graph protocol for Ethereum accounts has dropped, it's called Ethereum Follow Protocol. It isn't a Twitter clone, there's no posting, and it leverages ENS for naming. The main agenda behind this is a layer to add context for who accounts are, contacts, etc.
⛓️💥 Binance founder CZ was released early after serving 119 days in prison for failing to implement KYC. The whole CT1 is not sure crazy over it and markets are pumping. He tweeted “gm” btw. (weird time to be alive)
💸 Celestia raised $100M making it a total of $155M that the DA has raised.
💵 4 years after releasing it for consumers, PayPal enables business accounts to buy, hold, and sell cryptocurrency. With an in-house stable coin, PayPal is really a giant that is leading the crypto adoption silently.
👜 Privy introduced “Automated account transfers”, a way to link multiple accounts to the same social/wallet identity. This makes UX even better by linking all the accounts, and when a user tries to re-use their other identity, it resolves to the same account.
🛠️ Dev & Design Digest
🦕 Deno 2.0 is out, as you know famously wanted to be different from NodeJS, the OG is now fully back-compatible with Node, and already shipped jsr (npm but TS native) a few months back and now shit tons of APIs are stable in Deno. It can now run TS out of the box and have standard libraries to prevent installing separate stuff of formatting, testing, and utils (this is cool, ngl). One bonus thing is deno compile, which makes it easy to build SEA (executable from TS/JS). WATCH OUT BUN. 😉
🤡 A vulnerability in Arc’s boosts allowed gaining access to anyone’s browser without them even visiting a website. They were using Firebase to power their “Cloud features” and misconfigured ACLs led to the above shit being possible. Arc accepted that they had this vulnerability, released a fix, and added a write-up on CVE-2024-45489 too.
⚡ The experimental
scheduler.yieldAPI will be coming soon to your nearest browser and it’s an exciting API to wait for. Here’s an article covering what it does. TL;DR → This API gives a way to hand main thread control back to the browser (scheduler to be precise) to perform another task while that one big task is being executed.🆕 A new TypeScript flag landed in v5.7 → “rewriteRelativeImportExtensions”. What does this do? Along with
allowImportingTsExtensions, this flag now fixes a major bummer for newcomers to TS, the “JS Imports”. Yeah, up until now, if you are not using some custom bundlers, you’d have to import from the `.js` file of the corresponding TS file. But with this PR merged, now you can happily import from “.ts”.
What brings us to awe 😳
🤙🏻 True Caller has ~420M monthly active users, with India alone contributing more than 259M MAUs. The most unexpected thing for us was that it is not an Indian company but a Swedish one.
😭 In this week’s “never a dull day in crypto”, we’ve another hack of course, but this one is unique in its own way. Rehold, a multi-chain trading platform, tweeted that “the private keys had been compromised by their CTO”. And that’s not it, there’s extortion of money from CEO and criminal cases are there, it’s all wild. You can read it on their website too.
⚰️ Google Cache is now fully dead, it can be an addition to KilledByGoogle, but this one affects a lot of us, remember how 12ft.io and a few other services would load paywalled articles for free for you? It was Google Cache that enabled it. Now the only way to navigate the web in the past was Wayback Machine.
♟ Sanskrit poets in the 9th century Kashmir were writing poems whose word placement provides a solution to the Knight's Tour problem which is the genesis of what is today known as the Hamiltonian path problem in Graph theory. Check this video by Donald Knuth (watch from 36:00) talking about the same.
Today I (we) Learnt 📑
☕ Ever wondered what the “third” in Third Wave Coffee means? It refers to the (surprise) third wave of the coffee marketing movement. The first wave, brought in around Nescafe’s inception, was about low-priced instant coffee. The second wave which led to the rise of Starbucks was about differentiating coffee based on its origin country. And third-wave coffee is associated with the concept of specialty coffee, with a focus on quality and craft.
🥃 The first Friday of each month is known to Distillery employees as "Good Friday" as the employer gives a bottle of Jack Daniel's Old No. 7 along with a paycheck.
🪈 You might have used
source 2>&1 | destinationto pipe STDIN and STDOUT from one command to another. Samuel reveals a neat trick in a PR about a shorter syntax to do the same 🥁, bash has|&, so your piping command can be shortened tosource |& desitnation. You can replace your2>&1|directly with|&.⚛️ We have particle accelerators small enough to fit on a chip.
Source: Tweet by @vikramskr
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been consuming 🔖
🕹️ DepartureMono: A monospaced pixel font with a lo-fi, techy vibe. The font is good, but the website is too good.
⚒️ Dev Starter Pack: The essential startup starter kit. One page to aggregate essential tools →
devstarterpack🎨 TrueMedia: Identifying political deep fakes in social media using AI, give it the link.
🫂 The Hidden Force Shaping Your Software Architecture is a small article on how “Great software isn’t just about code—it’s about the people and structures that create it”. (TL;DR → Conway's law)
Builders’ Nest 🛠️
⚡winamp: The iconic audio player Winamp is now open source.
▶︎ nano-spawn: Tiny process execution for humans — a better child_process, with features like promise interface, iteration on output lines, etc.
🗑️ trash-cli: Move files and folders to the trash instead of the usual
rm, making it safe to recover.🛰️ spacer: CLI tool to insert spacers when command output stops. (the package is from none other than Sam, he’s active on Nibble’s Discord 😉)
Meme of the week 😌
Off-topic reads/watches 🧗
🔦 Knowing the territory by Seth, relaying “As long as there are folks who are lost, we will need guides.”
🪴 Alex Schultz, FB’s VP of Growth, talks about why Startups should not have growth teams, rather they should have a North Star metric as a growth leader.
📅 Do we need Daily Scrum meetings? by Adam Rad
⏲️ When to Do What You Love by Paul Graham. In this fresh & surprisingly small essay, he discusses whether one should "follow their passion" or pursue more lucrative work, and how one can’t produce “great work” without being passionate. (btw, it’s very much why we write this, not for money, just because we are passionate about reading in public)
Wisdom Bits 👀
“All the things which cause complaint or dread are like the taxes of life —things from which, my dear Lucilius, you should never hope for exemption or seek escape.”
— Seneca
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev
Weekly Standup 🫠
Nibbler P had a slow week marked by recovery and getting back his appetite. He lost a lot of weight and is on track to get that back with proper nourishment and staying active which he started with a 2k run!! (quick recovery, eh?)
Nibbler A has been traveling for a while now, working from new places and finishing up with some official work. He is just hoping to get done with this traveling and settle at a place for a few weeks. From what we could gauge, he is not liking the whole concept of countries at this point.
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻
Crypto Twitter