#90

Nvidia's new chip, Google and OpenAI continue shipping, $PENGU, waveterm, OpenVM, Pytorch shuffling issues, Claw Hand, Premise trap docx to markdown & more

👋🏻 Welcome to the 90th! (90 WEEKS!!! That’s like ~2% of the average life span) We also did a bit of cleaning in the sections this week based on the feedback we received. Thanks for being a Nibble reader 🫡
🎄 Advent Calendars are about to end with Christmas and The Holiday season around the block. Have fun. ☃️

unclog your chimneys, we are about to dive in.

📰 Read #90 on Substack for the best formatting
🎧 You can also listen to the podcast version of Powered by NotebookLM

---

What’s Happening 📰

• 🎁 NVIDIA unveils Jetson Orin™ Nano Super the most affordable Generative AI Supercomputer, priced at a sweet $249 (as compared to $499 for its predecessor). It delivers up to a 1.7x gain in generative AI performance, supporting popular models for hobbyists, developers, and students. (👉👈 hey! we are accepting Christmas presents btw!). As @championswimmer tweeted “It is the Raspberry Pi of the AI generation.”

✨ AGI Digest

• 🚀 It is impressive how Google (still) kept shipping this week, drifting away from its previous reputation of being a giant slow mover.

  • 🎥 Just when the week started and the initial Sora wave was settling, Google announced Veo 2, the latest version of its text-to-video model, and pushed some updates to Imagen 3 to better handle diverse art styles and high-fidelity details. While Veo 2 is still behind a waitlist, going by the handpicked demos posted online and by the people who had early access, Veo 2 nails photorealism and physics much better than Sora.

  • 💬 After testing the gemini-exp-1206 and Gemini Flash 2.0 in AI Studio for a while, Google went ahead and integrated them into the Gemini Mobile and Web Apps with the Flash being available for all and 1206 available for Gemini Advanced subscribers.

  • 💭 And, Google also released Gemini Flash 2.0 Thinking in AI Studio which builds on the success of Flash 2.0 and scores at par with gemini-exp-1206 in all categories on lmsys and performs really well on tough maths problems.

• ⭕️ 12 Days of OpenAI comes to an end, giving us a glimpse of what’s possible in the future

  • 🔎 Day 8 brought ChatGPT Search to all logged-in users and gave the Advanced Voice Mode access to search the web when answering queries.

  • 👨‍💻 Day 9 gave developers (yes tier-5, we are looking at you) API access to o1. The real-time API now also supports WebRTC and its prices have been slashed by 60%. OpenAI also has new Java and Golang libraries… and in case you felt useless, Sam just assigned a ticket to devs by renaming "system prompts" to "developer messages" across OpenAI’s APIs. They even sneakily changed the API docs, lmfao.

  • 🤙 Day 10 came with ChatGPT being available over a call from the US and Canada by dialling 1800-ChatGPT and chatting on WhatsApp for the rest of us.

  • 🤖 Day 11 was an update for the ChatGPT Mac App (with the same updates rolling out for Windows and Linux apps later), with support for more coding tools like Warp, IntelliJ IDEA, and PyCharm and some writing tools like Apple Notes, Notion, and Quip. You can also use Advanced Voice mode to work with these apps!

  • 🤯 Day 12 dawned on us all with a fresh dose of existential crisis when OpenAI demoed o3 and o3-mini. The o3 model which would now go safety and alignment testing (with the mini launching next month) has a 2727 rating on Codeforces making it equivalent to the #175th best competitive coder on the planet! If that doesn’t scare you, it also scored 71.7% on SWE-bench verified (the current best is ~55%) and has improved leaps and bounds on the ARC-AGI benchmark scoring 88% (there will be a much harder V2 in 2025) in the high-tuned mode though at a very high cost. And, even if you say that these models do not justify operating costs at the current state, remember that this is the worst and costliest they would ever be. We are going to see crazy times ahead.

    

    

    

    Can you feel the AGI?

  • 📽️ Bonus Day 13 was a surprise from Sam where they gave all plus users unlimited Sora access via the relaxed queue over the holidays!

• ⚓️ Model Drops

  • 🌝 MetaAI and Stanford released Apache 2.0 Licensed Apollo Multimodal Models in sizes of 1.5B, 3B and 7B and can comprehend up to 1 hour of video! However sometime after it was released, the model disappeared from HuggingFace without any explanation. Do you smell a WizardLM here?

  • 😎 XAI pushed new updates to Grok 2 and released grok-2-1212 and grok-2-vision-1212 on the API. This new model is also live on the Grok Chat on Twitter with support for web search powered by the Twitter newsfeed and updates on Aurora, their image generation model and a dedicated button to analyze a post you see on your timeline using Grok.

  • 🦅 The Technology Innovation Institute (TII) in Abu Dhabi released the Falcon 3 model family in sizes of 1B, 3B, 7B, 10B and a 7B-Mamba version which perform well across the board on various text tasks. You know that they are confident about their models because, in their benchmarks, they did not forget to compare to the Qwen2.5, which a lot of research teams shy away from ;)

    

    Source: TII HF Blog

  • ⚡️ ElevenLabs released a blazing-fast text-to-speech model named ElevenLabs Flash which can generate speech in 75ms (+ application & network latency) which is the realest real-time you can ever get right now.

  • 🐶 The folks at AnswerAI and LightOn released ModernBERT — a new model series that is a Pareto improvement over BERT and its younger siblings across both speed and accuracy and an increased context length of up to 8k tokens. These improvements stem from using newer architecture techniques that have been in the literature in the past few years after the release of the OG BERT + focusing on improving both the quality and quantity of data.

    

    Source: AnswerAI HF Blog

• 👩‍💻 GitHub Copilot is now free for all and also includes access to the new OpenAI o1 model. While it lacks agentic coding capabilities like the ones in Cursor and Windsurf, it is a good way for more people to get access to and experience AI coding capabilities and then slowly figure out other tools of their liking. BTW, our friend

  Sankalp
  recently wrote an article on The Evolution of AI-assisted coding features and developer interaction patterns that explores how the paradigms have been shifting since we first got the humble Tab autocomplete.

🔐 0x Digest

• 🐧 The Pudgy Penguins launched their ecosystem token $PENGU on Solana on December 17th. They’ll have a total supply of 88,888,888,888. Pudgy Penguins have been keeping people engaged for 3 years now, you might have come across multiple nerds or KOLs¹ using Penguin as their pfp now. The airdrop faced a little privacy hiccup, as a developer found out that public API returned all the Solana claim's linked ETH wallets.

  

  here’s approximate tokenomics from their tweet

• 💻 From custom circuits → RISC-V zkVMs → 🥁 OpenVM. Axiom & Scroll launched OpenVM, an open-source zkVM framework with no-CPU architecture (as in you bring your own), an extensible ISA², and unbounded on-chain verification all served in Rust frontend via intrinsic functions.

  

  Architecture of OpenVM (from their GitHub)

• 📣 French MP Sarah Knafo urged the EU Parliament to think about Strategic Bitcoin reserves like the US and said out loud that we shouldn’t opt in for the Digital Euro and its fight for freedom against a totalitarian European banking system that wants all control.

• 🐒 Succint brings Zk to Solana with a banger announcement video. They published the SP1 Solana verifier crate which can verify Groth16 proofs generated with SP1. Here’s the blog for how to get started.

• 👛 Privy introduced Server Wallets, they are like embedded wallets that can be scaled as they can be created and controlled using APIs, but at the same time keep the cryptographic systems of embedded wallets in place. It leverages secure enclaves and key splitting to ensure secure key reconstitution and appropriate custody for all use cases.

• 🇩🇪 Deutsche Bank to build an L2 on ZKSync, calling it Project Dama 2. Banks building their L2 was not on our bingo cards for 2024.

• 🔄 Monda Labs rebranded top Category Labs (because it rhymed with Gonad?) before launching a chain yet. They notoriously picked all the pain points of the Ethereum Ecosystem as their Value propositions like a 10,000 TPS, 1-second block times, single-slot finality, and low-hardware requirements. But hey, it’s not out there, yet.

• 💰 Vitalik shared a new way of funding open-source software calling it “Deep Funding” and also added $170k in ETH as a pilot to figure out which OSS dependencies of Ethereum to fund. The challenge is to develop an agent allocator function to assign weights to 40,000 identified Ethereum dependencies.

  

🛠️ Dev & Design Digest

• 🧵 Remember Meta’s microblogging platform Threads? Well, they are doing numbers, with users trying to move off 𝕏, Threads reported that they have a 100M DAU³, not only that Adam (head of Insta) wove(?) a thread they are seeing three months with more than a million signups a day.

• 📜 State of JS 2024 is out and we won’t spoil much for you, it’s a good interactive read, you should take that ride all by yourself. But people loved Vite and TypeScript a lot. (that means you’ll find AI Coders writing these two stuff more human-like.)

• 🐢 Lee Robinson wrote a thread on “Why does transition between pages when running Next.js locally take time?” and ended up calling everyone who asked as stupid explaining lazy route compilation, how Turbopack (a bundler that fixes world hunger using Rust) fixes some stuff, and hey you can still do use client and opt for useEffect.

• ⚡ e18e, the project aimed to clean up, speed up, and level up the JavaScript ecosystem published their November Contribution showcase, mentioning some package trimming and utility merging, mostly on performance gains.

---

🤝 You have read ~50% of Nibble, the following section brings some fun stuff and tools out from the wild.

---

What Brings Us T(w)o Awe 😳

NOTE: yeah, we merged the “What Brings Us To Awe” and the “TIL” sections into one.

• 🖥️ We are almost certain that someone in tldraw is a time traveler. What happened now? Well, tldraw launched the tldraw computer, which is like an infinite canvas where you can do whatever you want, like adding custom logic to these boxes or calling LLMs. (quite literally). While walking down this rabbit hole, we found a tool Paul Shen’s natto.dev, a spatial live multiplayer JavaScript environment, and an attempt to make code more tangible and interactive.

• 📊 The folks who worked on ModernBERT found an interesting issue in Pytorch where the torch.randperm infact does not shuffle the elements randomly when dealing with magnitudes of billions, while numpy’s PCG64DXSM generator does. The fix is on the way by the Pytorch team but this issue highlights how

  

  Source: Post by @alexisgallagher on Twitter

• 🌔 Here at Nibbles, we love Bartosz Ciechanowski and his immensely well-researched and interactive blog posts. Just this week, he published another one related to the cosmos, this time about the Moon, Moon its path around our planet and the beautiful dynamics behind it.

• ☕ Does walking with a coffee mug often lead to spills? In 2015, Jiwon Han published a paper on why that happens (spoiler alert: coffee likes to slosh with the same frequency at which we like to walk). He also suggested several methods to prevent that, like Claw Hand which minimizes the spill 🦞.

  

  Source: Fermat’s Library on 𝕏

• 🛐 Julia Evans listed her 20 years of using the terminal in a blog post "Rules" that terminal programs follow. It’s full of TILs, gotchas, and booby traps you should know if you are a terminal person (or not).

  

  this is just one part of one of the rules, she listed 6 more.

• 🔢 Sexy Primes are prime numbers that differ from each other by six!

  

  [Originally posted by @that_anokha_boy on Twitter]

---

What We Have Been Trying 🔖

• 👨‍💻 TypeScript Console: Run and debug TypeScript code in the Chrome DevTools.

• 📹 Article Summarizer: Lesson22 launched an extension that can convert turn text into a memorable watch. [Eg: this video explains the new Protobuf Opaque API]

• 🔥 Addy Osmani built an open-source image background remover using Transformers.js. So that you and I don't put our files on random applications, we can host our own, or trust Addy's code for images never leaving your device. (exciting times for WebGPU and AI in the browser)

• 🗺️ Timemap.org - a World History Atlas. Using this app you can navigate back in time and see regions, rulers, people, and battles happening in that region. (a tool sufficient to give a boner to both map and history chads)

  

  they say a picture sells things faster…

---

Builders’ Nest 🛠️

• 📄 markitdown: Official Microsoft library for converting Word, Excel, and PowerPoint files to Markdown.

• ⚒️ ƒun: a local serverless function λ development runtime. It’s just like AWS Lambdas fr fr, but local.

• 🌊 waveterm: an open-source terminal that can launch graphical widgets, controlled and integrated directly with the CLI. honestly, it works well with my ADHD.

  

  a screenshot showcasing all its capabilities.

• 🤗 HuggingFace chat-MacOS: A native chat interface by HuggignFace designed for MacOS. (note — use the package in the release page instead of brew install because there was some checksum mismatch issue in the latter, the last time we tried)

---

Meme of The Week 😌

Source: @Yuchenj_UW on Twitter

---

Off-topic Reads/Watches 🧗

• 🪤 The premise trap by DHH, back after ~25 days, writing a post that he (and we) hope doesn’t age well. DHH talks about how AI is much like a promising junior engineer (right now) and how the training data it has been fed is the root of problems. And maybe feeding it more robust corporate codebases will fix it?

• ✊ The fame/trust inversion by Seth talks about how in the short run, fame (or marketing) leads to trust instead of the other way around, but “well-earned trust might go out of style for a while, but it’s always going to be a useful tool.”

• 👛 Underrated Ways to Change the World by Adam Mastroianni shows us that you don’t have to be born with a silver spoon in your mouth to make a positive change in the world. Sometimes, a few unconventional ways and some courage is all you need :)

  Underrated ways to change the world

• 🪖 Nat Eliason says in Proof You Can Do Hard Things that it is the ability to do hard things is an essential skill to cultivate which builds your self-image, making you more confident and resilient when facing future challenges.

  Proof You Can Do Hard Things

---

Wisdom Bits 👀

“Most of the trouble in the world is caused by people wanting to be important.”
— T. S. Eliot

---

Wallpaper of The Week 🌁

🌌 Grab the week’s wallpaper at wow.nibbles.dev.

---

Weekly Standup 🫠

• Nibbler A is filing a complaint (no idea where) as he doesn’t feel that his week has 7 days like others. Still, all occupied catching up with his readings, AoC, and racquet-shuttle game.

• Nibbler P did some infra and code stuff this week, ran longer than he ever ran, deadlifted more than he ever deadlifted and went on a sweet night drive with his family at the end of the week.

If you have 2 minutes, please fill out a feedback form here. Your input will help us improve “The Nibble” for you and other readers!

If you liked what you just read, recommend us to a friend who’d love this too 👇🏻

1. Key Opinion Leaders (ज्ञानचोद)

2. Instruction Set Architecture

3. Daily active usersupto