#89
Google's Willow, OAI vs Google DeepMind Shipping battle, WebDevArena, AndroId XR, Tips for 20s, Rupert's Drop, Street Alignment, Signals, this to that, lima, scooter, Choosing Startup & more
👋🏻 Welcome to the 89th!
What a week good last week we had, we have the youngest-ever world chess champ (undisputed) 🥳 🇮🇳♟️
The Christmas coziness is kicking in and the slow weeks of the year will start soon. 🦥
📰 Read #89 on Substack for the best formatting
🎧 You can also listen to the podcast version of Powered by NotebookLM
What’s happening 📰
🪄 Google Quantum AI labs announced Willow, their state-of-the-art Quantum Chip (*screams eternally*). Willow can do some computations in 5 minutes that would take 10 septillions (10²⁵) years (longer than the age of the universe) on current best supercomputers. This is super exciting as it unlocks the possibility of more open and large-scale quantum computers being available soon. This also aids in scaling issues with A
GI.
The major breathtakingly important changes that they bring are:The newer chip can reduce errors exponentially as they are using more (105 to be precise) and better quality qubits1.
The new chip absolutely killed it in benchmarks when done on Random Circuit Sampling. (the “Hello World!” of Quantum computing)
And don’t forget that Neven’s2 law states quantum computers are improving at a "doubly exponential" rate compared to conventional computers.
✨ AGI Digest
⭕️ OpenAI continued shipping with Days 3 to 7 of their 12 Days of OpenAI, but we still didn’t get GPT4.5/GPT-5 😞 (hopes high for next week)
📼 Day 3 started the week with a bang with OAI unveiling Sora, their SoTA video generation model and boy oh boy, it is so wild, it leaves the rest of the competition behind. After a few scaling challenges, sora.com was generally available for all Pro and Plus users.
📋 Day 4 had OpenAI make its Canvas (a collaborative split-screen writing and coding interface, similar to Claude Artifacts) available natively with all GPT-4o chats, even with custom GPTs, along with the ability to execute Python code in it. This code execution differs from its previous “Code Interpreter” because Canvas uses Pyodide (Python in WASM), which means it can now make network requests from Python!
🍎 Day 5 saw only an announcement of integration b/w Apple Intelligence in iOS/iPadOS 18.2, MacOS 15.2, and ChatGPT allowing Siri to make calls to ChatGPT to present the user with richer results for complex questions. We found the feature good and seamless for quick tasks but we’re still better off writing our own prompts, thank you very much. The Apple Intelligence section also has a dedicated “Upgrade to ChatGPT Plus” button on the Settings page that would surely nudge a lot of people to get the OpenAI sub. Smart move Sama.
📽️ Day 6 got us live video recognition in the ChatGPT App’s Advanced Voice Mode (AVM). So now along with speaking, you can also show it the stuff around you and it would consider that when responding (btw, Google beat OpenAI to this by releasing Geminin 2.0 Flash with multimodal capabilities just a day before, more on it in the following section). The best way to get it is to see it for yourself in the app or in the demo video below. Oh and they also have a new Santa voice in the AVM until Christmas, so you know what to do when your kid posters you to take them to Santa.
🗂️ Day 7 finally ended the week with ChatGPT Projects (similar to Claude projects), a way to group files, chats, and custom instructions into different folders.
♊️ Google Continues Shipping with the first release in the Gemini 20 Family
⚡️ A day before OAI shipped the AVM live video, Google dropped Gemini 2.0 Flash (different from gemini-exp-1206 btw), which despite being a much smaller model, outcompetes Gemini 1.5 Pro on several benchmarks. Gemini 2.0 Flash supports tool use and is natively multimodal with text, image, audio, and video inputs and support for audio output with image output coming along next month. You can try the live video/audio input streaming in the AI Studio.
Source: Gemini Blog 👨💻 And with that, Google has already started the integration of the Gemini 2.0 Flash, starting with Data Science Agent in Google Colab, which will have new agentic coding capabilities that would fix bugs, write code and execute them based on the description of the task.
📒 NotebookLM also gets new updates with a slick new UI and the ability to “join” the AI Podcast hosts and interact with them live and ask them for more details or to explain a concept differently.
⚓️ Model Drops
🐳 DeepSeekAI released DeepSeek-V2.5-1210, the final model in the DeepSeek V2.5 series (V3 is coming next!) This is available to use in chat.depseek.com with the ability to use “Internet Search” on the UI. Benchmark-wise, the model shows quite big improvements in both coding and creativity.
Source: DeepSeek Twitter 👀 DeepSeek also released DeepSeek-VL2, an MoE VLM in 3B/16B/27B sizes, giving equal competition to Qwen2-VL-7B with fewer active parameters. Compared to its predecessor, this was trained on higher-quality data and added dynamic image tiling for handling images across all resolutions.
Source: DeepSeek Twitter ϕ Microsoft Research released Phi-4, a 14B LLM which “outperforms” the likes of GPT-4o and Gemini Pro 1.5… but if anything can be said from the past Phi models, then it is that these models are trained majorly on synthetically-created datasets and the family often falls short of performance on real-world tasks, so take these claims with a grain of salt and always do your due diligence with your own eval sets.
📟 Cohere4AI released Command R7B, a 7B model — the smallest one in their R series which is deemed to be a good model for RAG use cases. The model is both available on HuggingFace and available via an API at the Cohere Platform.
Source: Cohere Blog
🧺 Misc Updates
🧑💻 Cognition Lab’s Devin is now Generally Available for $500/month (not expensive considering that we are replacing entire software engineers over here). And though it did have its own share of hiccups, this is the worst it would ever be. A year from now, this may very well be integrated all around every workplace you come across. Brace yourselves!
💨 Hugigngface released TGI v3.0 which is faster than ever, processing 3x more tokens than vLLM and 13x faster processing on long prompts. The way they have achieved this is by removing all config flags and instead evaluating the hardware and model, to select automatic values to give the best performance. Plus, they also introduced new kernels, optimized prefix caching, and improved VRAM efficiency. Ohh btw, vLLM has joined the PyTorch Ecosystem meanwhile.
Source: TGI Blog 🤺 LMSys has a new WebDev Arena to compare the web coding capabilities of various frontier models. And unsurprisingly, Claude 3.5 Sonnet does come on top and that too by a wideeeeee margin.
Source: @lmarena_ai on Twitter 🪡 Midjourney has introduced Patchwork, an infinite-scrolling world-building canvas with which you can create characters, events, places, and factions — creating anything from your imagination all of it in the theme of the current patchwork. You need to see it to understand how wonderful it is.
🔐 0x Digest
💸 Since none of us like paying taxes on our gains (realized or unrealized, it doesn’t matter), Nick made a list of 22 countries with no capital gains tax on crypto.
💰USDT is now officially recognized in the Abu Dhabi Global Market (ADGM), and as much as I want to say “yay! global expansion for crypto ftw ;)”. Let’s be honest here, Tether is mostly controlled, it’s pegged against USD and has reserve backing it, not sure if it’s a big win for “acktual” cryptocurrencies or decentralization in general. But.. hey, you can now swap and spend, yay!
💳 Moonwell Finance and Cypher Capital partner to launch a new crypto card, Moonwell Card. This is yet another attempt to make it easy to use crypto to buy groceries, and coffee, basically making it mainstream and usable. A few key things are it has no annual fee and allows auto-top-up from Google or Apple Pay.
⛔ Microsoft shareholders voted “no” to hold Bitcoin in reserve. The main reasons cited for rejection were “too much volatile” (ufff) and too much of it driven by FOMO from riding the next wave.
🇯🇵 Satoshi Hamada, an MP in Japan proposed for Strategic Bitcoin Reserve like Argentina and Russia. (the name is purely coincidental, right?)
📦 Stackr published “Your definitive guide to zkVMs”. They took 8 real-world algorithms to benchmark 6+ different zkVMs/zkWasms. (shameless plug Nibbler A here.)
🛠️ Dev & Design Digest
🕶️ If you have tried the Vision Pro demo or have heard from people about it, it’s a dope cyberpunk-themed gadget. But we bet you didn’t see this one coming, In a not-so-rare partnership, Google, Qualcomm & Samsung are collaborating on a new AI-powered OS, Android XR a way to put Android VR headsets on your face. Samsung built a device to showcase this and codenamed it “Project Moohan”. To remind you, this isn’t the first time Google touched this space, Google Glasses came and crashed a decade ago.
Victoria from The Verge tried it, wrote her experience here, and ended her experience with “I felt like Tony Stark with Gemini as my Jarvis”.
For Consumers → The first devices will be available in 2025.
For devs → Google has already launched the Developer Preview of Android XR SDK.
🎨 A decade-old rant on how compilers of hot new languages handle async and sync code. “What color is your function?” by Bob. He concludes it with Java ain’t bad, Go is great as it breaks the notion of async and sync and leaves it up to users.
🖋️ A frontend developer, when bored of fixing build systems and drilling props into children's components usually
goes to therapystarts drawing SVGs by hand. You see, SVGs are beautiful (yes I skipped that therapy session too), and drawing them is hard (for most of us). Akshay Gupta wrote an article Mastering SVG Arcs, breaking down the move and arc attributes ind(drawn) attribute ofpath, which allows us to draw curves as we like them. (shut up!).🚥 If you were busy writing
useStateslop (like us), while the web folks were making the noise about Signals. Corbin has you covered with their really intuitive article “What are Signals?”, where they explain Signals API by implementing it function by function for you. The Signals API is cleaner as compared to our usualuseMessyHookWithDepArrayand that’s what the noise was all about. One interesting bit that Signal has that makes it different is auto-tracking, which allows you to do things on the effect of a signal without a dep array.
What brings us to awe 😳
🗺️ How does Google Maps even exist? To give you an idea of complexity, Maps processes 25 Petabytes3 of data daily, uses Graph DB with over 1 trillion edges, and maintains 99.99% uptime with a global response time of under 100ms. And yes it uses every other software optimizations you might have heard of or used. With all this, building a maps application that can load vector tiles as fast as you scroll is peak engineering.
🚏 A map that plots the orientation of streets of 100 cities in the world. Some of the big cities worldwide are more disoriented than you might think. Also, Delhi and Mumbai are surprisingly well-oriented than you could’ve guessed. The chart is not intuitive, but you have to feel it. If you are feeling like you have to dig into it here’s the source data provided by Deedy that he used to plot this.
Source: Tweet by @deedydas
Today I (we) Learnt 📑
🫗 When you drop molten glass in cold water (yes, just like ice plunge) the glass rapidly cools and solidifies (as it should, duh?), but what happens next is interesting, it forms a tadpole-shaped (sorry this was only PG-13 reference) glass drop that is formed. It is called Prince Rupert's Drop. It is not just like any other object though, it has a very strong head but a very delicate tail (like male muscle and ego?), such that the head can survive bullets, and bend the hydraulic press BUT if you press the tail a bit, it's EXPLODES the whole thing. We went down this rabbit hole, here's a good video to learn about it.
😵💫 We all come across situations, conversations, TV shows (CID?), or Movies where the whole thing could’ve ended sooner if not everyone involved in that was not an idiot, right? Well, this thing is so common that it has the name Idiot's Plot. Not only that there is a second-order idiot plot when a narrative "in which not merely the principals, but everybody in the whole society has to be a grade-A idiot, or the story couldn't happen".
▲ We dived into a rabbit hole that started from someone asking about Decentralized DNS on HN (a year-old post), turns out DNS-based Authentication of Named Entities (DANE) solves this and proposes a way to authenticate the TLS connection without the need for centralized Certificate Authorities (CA). Then what’s the catch? who is stopping us, well most of these CAs are controlled/operated by big tech giants like Google or Microsoft (just like DRMs). The major browsers haven’t implemented DNSSEC support and it’s kind shame. But never the less, someone on the thread pointed out that it’s not as easy as it sounds as suggested by Zooko's Triangle.
three traits of a network protocol
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying 🔖
📦 npmpackage.info: Discover detailed information about npm packages.
🤝 this to that: a tool that guides you on how to glue item a (this) to item b (that).
💻 lima: Linux virtual machines, with a focus on running containers. (been writing and debugging some C/C++ lately and realized you can’t have
gdbon Mac)🖼️ TextBehindImage: a dead-simple app for new images with text behind the main subject. (like iOS putting a digital clock behind objects in wallpaper)
Builders’ Nest 🛠️
🛵 scooter: an interactive find-and-replace terminal UI app. ofc it’s written in 🦀
📨 himalaya: a blazingly fast (yes, 🦀) CLI to manage emails.
💽 limbo: a work-in-progress, in-process OLTP database management system, compatible with SQLite.
🤖 react-executor: Asynchronous task execution and state management for React.
Meme of the week 😌
Off-topic reads/watches 🧗
😫 Complaints by Seth Godin talks about why complaints are good (can't press this enough) but there is a fine line, where the complainers become whiners and there's where shit hits the fan. So, don't be a whiner — complain, but with dignity.
💪 James, CEO of Posthog talks about how Product Management is broken and Engineers can fix it. He talks about *gasps* Production Engineers.
🏋️♀️ Dalton & Michael from YC talk about “How to make the most out of your 20s”. Some important bits are Embracing hardcore challenges early, delaying gratification, and seeking risks.
📝 How to choose a startup by Dan Hock. The article covers great points telling it how good old metrics like retention, unit economics, compensation and culture should be accessed and used to determine if you should join the startup. And please stop applying lessons from big companies too early in the life of a startup.
Wisdom Bits 👀
“How beautiful it is to get up and go do something”
— Kurt Vonnegut
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev.
Weekly Standup 🫠
Nibbler P is back at work — shipping features, trying to meet his Goodreads reading goal, and maintaining his workout streak at the gym and the running track. He has been taking time to watch some new movies and anime meanwhile.
Nibbler A is catching up with the OSTEP Book Club reading (covering the last few missed weeks) and enjoying AoC 2024, and rabbit-hole maxxing. He has been consistent with 🏸 & 5 am club after so long. (yay!)
If you have 2 minutes, you can fill out a feedback form here. This would help us make The Nibble even better for you and other readers out there!
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻
qubits are hard to make (super low temperature and conditions required). They are more hardware-heavy, unlike our regular bits, which are just voltage changes in the conductor at room temperature.
Hartmut Neven is the director/CEO of Google's Quantum AI Lab.
1,000 terabytes (TB), 1,000,000,000,000,000 bytes, and 1 million gigabytes (GB)