#80
OpenAI's Dev Day, Liquid models, Visa's Tokenized Asset Platform, Starknet staking on L2, Caesar Salad, Mac's Startup sound lawsuit and more...
👋🏻 Welcome to the 80th!
📰 Read #80 on Substack for the best formatting
What’s happening 📰
💼 For those who care, Google confirmed that it won't be mandating a strict return-to-office plan like Amazon and Salesforce did recently. They said they’ll stick to hybrid as long as people are productive.
💰 Evan You announced void(0), backed by Accel, to support the development of an ambitious superfast JS toolchain including Vite, Vitest, Rolldown, and oxc.
✨ AGI Digest
🗃️ Last week, OpenAI had their Dev Day where they announced a bunch of updates targetted primarily at Developers (no new models, sadge), plus some other announcements outside of it along with a massive $6.6B funding round that now brings them to $157B post-money valuation.
🎤 Finally, we can get real-time speech-to-speech output with socket streaming, similar to ChatGPT’s Advanced Voice Mode, via an API from the GPT-4o. OAI said they would also release the non-streaming API through the familiar Chat API in the coming weeks. Currently, the real-time API costs a whooping $100/M for input and $200/M for output speech tokens which though it looks like at first glance is a lot is quite competitive with call-center-level salaries.
Say bye-bye to the annoying customer-care call waiting queues because “all out call agents (humans) are currently at capacity” is going to be a phrase of the past pretty soon.
Source: Tweet by @rohanpaul_ai 👁️🗨️ Finetuning GPT-4o now supports both images and text, thus for the first time allowing a frontier-level model to be fine-tuned for your niche vision use cases. This will be useful everywhere from self-driving cars to automating robots in factories.
🫴 Following suit from DeepSeek, Google, and Anthropic, OpenAI also introduced prompt caching for the latest versions of GPT-4o, GPT-4o mini, o1-preview, and o1-mini that includes 50% discounts and faster processing for recently seen input tokens.
Comparison between the different companies | Source: Peter Gostev on Linkedin ⚗️ They also got Model Distillation that allows fine-tuning a smaller model (GPT-4o-mini for now) with the outputs of a large frontier model (like GPT-4o and o1-preview) all from their platform, complete with adding your own evals to test the model performance. Neat!
🙊 Outside of DevDay, OAI released a new member of their Speech to Text model Whisper
large-v3-turbo, an optimized model of its big brotherlarge-v3, leading to more than 5x speed improvements. It uses just 4 decoder layers now, down from 32 of the large model.📋 Similar to Anthropic’s Artifacts, OpenAI released Canvas which has a new sidebar interface (wonder where got the idea from) that helps with tasks such as writing, rewriting, programming, making edits, etc.
⚓️ Model Drops:
💧 Liquid AI announced Liquid Foundation Models — a series of 1B, 3B, and a 40B MoE model built with “a new architecture” that performs SoTA when compared to models in a similar parameter range while being more memory efficient. You can try the 40B and 3B model on their playground, among several other endpoints. However, the details of their architecture are not yet released so we do not know how different they really are.
👀 Nvidia released a new 72B VLM called NVLM 1.0, which rivals the leading proprietary models GPT-4o and open-access models like Llama 3-V 405B and InternVL 2. However, they very conveniently skipped doing any comparison with the Qwen2-VL-72B which clearly performs better than it on benchmarks as well as real-world tasks and has much permissive license allowing commercial uses.
🫐 The secret “blueberry” model that has been winning the image generation leaderboards lately was Black Forest Labs’ latest model Flux 1.1. Pro. Flux 1.1 Pro’s generations are ranked much higher than the Flux 1.0 Pro while being cheaper than it and faster than even the Flux 1.0 Dev model!
🐤 The BAAI team released Emu 3, a suite of SoTA multimodal models trained solely with next-token prediction! By tokenizing images, text, and videos into a discrete space, the team trained a single transformer from scratch on a mixture of multimodal sequences which can both generate text and images as well as understand them really well.
🧩 Meta updated their SAM 2 model with “new improved set of checkpoints“ and released it as SAM 2.1.
⚡️ Pika Labs announced Pika 1.5 with more cinematic footage, longer clips and the viral explode, melt, crush, and inflate effects!
🎥 Meta AI demoed the capabilities of their text-to-video generation model called Meta Movie Gen. Being merely 30B large, this can generate high-quality and high-definition images and videos from a single text prompt. It even supports transforming existing videos with text inputs. Another 13B model called Movie Gen Audio can create and extend sound effects, background music, or entire soundtracks using video and text inputs.s
🔐 0x Digest
🪙 Visa launched the Visa Tokenized Asset Platform (VTAP), allowing that helps banks to issue fiat-backed tokens. The solution is available to participating partner financial institutes to create and experiment with their tokens in Sandbox mode. [Press Release]
👯 Varun from Hyperspace announced "Proof of FLOPS", a peer-to-peer network in the context of AI, and its first stage is available live today in Testnet on Hyperspace.
🌉 Arbitrum announced that "Bridged USDC Standard" is now available for all Orbit chains. This brings USDC natively to all orbit chains, a big win for Circle.
💫 Starknet became the first one to introduce staking on L2. The community has decided on the staking parameters, and they are expecting the first phase of staking, testnet, and mainnet to be live by the end of Q4 2024.
📨 A sale of EigneLayer tokens worth $5.5M on Friday raised questions about how they exploited the lock-in period, on further investigation it turned out that an email thread was compromised leading to the transfer of the tokens to the attacker instead of the recipient.
🔀 Interoperability
🔄 Kinto introduced Musubi for Chainless Swaps, allowing you to swap assets across Ethereum with your fingerprint without leaving your Kinto-insured wallet.
🛣️ Hyperlane announced an interchain token standard that enables asset issuers to scale token distribution and is calling it Warp Tokens
⛓️ Uniswap and Across Protocol partnered to propose ERC-7683, An interface for cross-chain trade execution systems.
🤝 Partnerships:
🛠️ Dev & Design Digest
🎛️ Artem wrote a piece on “Don't Sleep on AbortController”, you might have already seen this one. But if you haven’t, he is trying to tell you if you are building something that requires aborting or canceling operations, don’t look any further than native and super-flexible
AbortControllerAPI.😵💫 HTML conformance data for 2024 are out and it’s funny that only 0.5% of the Global top 200 websites use valid HTML, forget frameworks, we aren’t even using HTML correctly.
⛴️ ESLint shipped some good stuff in a recent release, they now have official support and plugins for JSON and Markdown linting and have released Code Explorer, a platform (finally something from the team) to view AST and test custom rules.
👨🏻💻 Web Components Are Not the Future, an article exploring Web Components that have been selling out the dream of “components that feel as native as DOM elements”, “portable interoperable web” and “platform or environment agnostic web”. No doubt they are lit (pun intended), but lately, the dreams have been fading away, as the learning curve is steep and sometimes it’s too much effort to build a <web-component/> over a JS library one.
🔮 The PR that makes Next's Dynamic API async is now merged in the Canary version. A shit ton of exciting and breaking changes coming for the future of "Next".
What brings us to awe 😳
🧫 Diatoms (very confusing name btw) are single-celled algae that have a glass shell. This shell refracts light in beautiful colors if viewed under a microscope. And boy do people make beautiful art from them!
🎵 The
mancommand had an easter egg that if you type “man” at 0001 hours it should print “gimme gimme gimme”. ABBA’s Gimme Gimme Gimme inspired it. The fun part is this easter egg remained until someone ran tests on man’s output and realized they failed during this odd hour, the author did patch it in 2017. It’s funny how the core packages in the system have these fun elements which are mostly harmless.🚚 No such thing as “exactly-once” delivery by the Sequin team. If you have used Pub/Sub services you might have seen them claim “exactly-once delivery” (with acknowledgment receipts), but is it? The linked post highlights the fact which is mostly misunderstood by people that, no system can be bug-free with “exactly-once” delivery. They are either “at-least-once delivery systems with idempotent processing” or buggy 🤷🏻♂️.
🔐 Android’s new promised features are rolling out slowly to users, including one of the most interesting ones “Theft Detection Lock”, which uses AI to lock your screen when it detects “common motion associated with theft,” like someone snatching your phone out of your hand.
Today I (we) Learnt 📑
🥗 The Caesar Salad was invented in Mexico by Caesar Cardini, an Italian chef and, contrary to the popular narrative, it has nothing to do with Julius Caesar.
🤷🏻 You can have a global
gitignorefile by using git config --global core.excludesfile ~/.gitignore, in the command, we’ve kept the file at~/.gitignore, but you can really have it in any directory with any name and keep thedesktop.iniand.DS_Storeout of git.🎼 The name of Mac’s old startup sounds Sosumi, the name of which sounds very Japanese and does not mean anything in Japanese, but is a play on the words “so sue me!” by the creator Jim Reekes because of a long-running lawsuit with Apple Corps, a registered company of The Beatles.
🤯 You must have heard about Ramanujan’s Dream revelations, but there is a less complicated constant called “Kaprekar’s Constant” → 6174. What’s special about this number? Take any four-digit number with at least two different digits, rearrange the digits to form the largest and smallest numbers, and then subtract the two numbers. If you keep doing this, you’ll reach 6174 in less than 7 iterations.
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been consuming 🔖
🔡 Get the maximum out of your font: Features of your font you had no idea about
🗺️ TubeTrotter: A baby of Google Maps and YouTube, where you can find the travel vlogs plotted on a map.
🔗 uithub: Get the entire GitHub repo (or a part of it) as a single raw file to pass as context to your LLMs.
🔕 ultrafocus: CLI tool to block distracting websites and boost productivity. (Nibbler A was using the exact hack for a few weeks, i.e. adding an entry in /etc/hosts, and then stumbled upon this package.)
Builders’ Nest 🛠️
🐚 stricli: Build complex CLIs with type safety and no dependencies by Bloomberg. It’s like oclif, but supposed to be lighter. (does the existence of this imply there should be a purush-cli?)
🔬 OpenAI Microscope: A collection of visualizations of every significant layer and neuron of 13 important vision models.
📃 napkins.dev: Turn your napkin wireframe into an app using Llama 3.2
🕸️ WebLLM: A high-performance in-browser LLM Inference Engine.
Meme of the week 😌
Off-topic reads/watches 🧗
🪢 Simplifiers Go Far, Complexifiers Get Stuck by Dave Kellogg, on Keep It Simple, Stupid.
🧹 The broomstick objection by Seth, on how sometimes afraid of something and sends the well-meaning person on a fruitless mission, and how to handle that.
🎒 What the smartest people do on the weekend is what everyone else will do during the week in ten years, an old but gold by cdixon was relevant 8 years ago and will remain relevant 8 years from now. All you need to do to benefit from it is either 1) be smart yourself, and if you’re not then 2) figure out what the smartest people are doing in their free time that you are not and give it some thought!
🦄 Unicorns are rare again by Peter Walker, Head of Insights at Carta. Peter uses the data to show how we went from “333 unicorns in 2021” → “56 unicorns in H1 2024”. And it’s not something bad inherently, hitting the Billion-dollar mark is hard, and maybe markets are healing from ZIPR.
Wisdom Bits 👀
“The true delight is in the finding out rather than in the knowing.”
— Isaac Asimov
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev
Weekly Standup 🫠
Nibbler A finally ended travel mode and will be on the West Coast for the next few weeks. He’s back in learn-and-build mode as we speak.
Nibbler P attended to work stuff while slowly regaining his lost weight. He also watched the Joker sequel and discovered what two hours of disappointment feels like.
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻