#77
OpenAI's o1, Pixtral, Free JavaScript, OP Succinct, Gas Accounts, cbBTC, L2Safety, TS 5.6, USB-C Soldering, Synthetic diamonds, Win32 epoch, stripe.dev, BED, Illuminate, Slimify, allegedly and more
👋🏻 Welcome to the 77th!
And just like how she liked it, this one is delayed.
Jokes apart, in case you were waiting, we are sorry for the delay. It’s probably the first time we have delayed this much in the last 77 weeks, but it was worth whatever made it delayed and we hope it’s worth reading too.
📰 Read #77 on Substack for the best formatting
What’s happening 📰
🥱 Apple held its annual iPhone event. Since what they announced is already all over your newsfeed, your timeline, and your memes, we’ll skip the obvious announcements and focus on things that caught our eyes (btw check your iPhones as iOS 18 is being rolled out as we speak):
👀 The new on-device vision model is really impressive. Along with being able to recall videos from the description of just a couple of frames, the model can use “Vision Intelligence” to also understand your surroundings and tell you information about it in real time (kinda like Google Lens but partially local). Oh, and talking about on-device, here are some of the model prompts that the 3B-parameter local Apple Intelligence model uses internally to do some of these tasks.
Source: The Verge 🧏 The AirPods Pro 2 being able to double up as a clinical-grade hearing aid is a huge accessibility step-up for those hard of hearing. We remember being excited when BeMyEyes partnered with GPT-4V. These, along with the huge improvements in natural and instant text-to-speech, the baseline for inclusivity and accessibility for people with difficulties has never been lower.
🎮 Sony released the PS5 Pro, upgrading the GPU, adding a more powerful Ray Tracing and AI-based upscaling, boasting of running “even higher fidelity graphics with smoother frame rates at 60FPS“. But how much of it can you really notice?
Source: Tweet by @AlexStrook 🙅♂️ Codeforces banned the use of AI in its rated contests. This was bound to happen sooner or later and probably more contests and competitions would follow suit with similar explicit bans.
✨ AGI Digest
⚓️ Model Drops:
🐳 DeepSeek silently released DeepSeek-Coder-V2-Instruct-0724, which ranks #3 on the Aider LLM Leaderboard and beats the newer DeepSeek V2.5. Sadly it is not available on the DeepSeek endpoints as both their chat and coder endpoints call the V2.5 model currently.
🌄 Mistral (in its classy way of suddenly dropping just a torrent link) released Pixtral, its first VLM. Though they have yet to release benchmarks and other data on their blog, we got some coverage from their AI Summit in SF. It is based on the previously released 12B NeMo model with a 128k context window and commercial-use-friendly Apache 2.0 license. Though looking at the benchmarks in their SF presentation, it looks like the 7B Qwen2-VL outperforms it on several of them while being much smaller in size.
Source: Tweet by @_philschmid 📊 Google released two fines-tunes of Gemma2-2B, named DataGemma, optimized to work with DataCommons1. One of the fine-tunes, DataGemma RIG2 enhances the capabilities of the LLM by proactively querying trusted sources and fact-checking against information in Data Commons. The other fine-tuning, DataGemma RAG enables it to incorporate relevant information beyond its training data, like a regular good RAG model.
🐟 The FishAudio team released their open-weight TTS model Fish Speech 1.4 (NC-use License though :/). This is a multilingual (8 languages) model trained on 700K hours of speech, and it also supports voice cloning instantly, thanks to its small ~1GB size.
📋 Some researchers from top Chinese AI Unis released GOT (General OCR Theory), a 580M end-to-end OCR-2.0 model that outperforms all existing methods. It uses a Vision-Encoder (VitDet) and a LLM Decoder (Qwen2) for generating OCR outputs in various formats. GOT is designed to handle complex tasks like sheets, formulas, and geometric shapes and outperforms both LLaVA-NeXT and the Qwen-VL-Max while being much smaller in size. The weights are available on HF and you can also try the demo in the browser. Having said that, the licensing is still unclear on this one currently.
📖 JinaAI released reader-lm-0.5b and reader-lm-1.5b
—two SLMs specifically trained to generate clean markdown directly from noisy raw HTML. Since they are trained to just clean and extract web data and not “generate” new text, they benefit by having a significantly low parameter count (0.5B and 1.5B) while still outperforming some of the frontier models.
“It seems that what we need is a shallow-but-wide SLM. ‘Shallow‘ in the sense that the task is primarily simple ‘copy-paste’… and ‘wide’ in the sense that it requires long context support to be practical“
🍪 DeepSilicon (YC 24) is building custom chips to run Ternary Transformers inspired by Microsoft’s Bitnet 1.58 Paper. They say, “We don’t even come close to the theoretical maximum speed for our kernels, and a large part of the failure is because the architecture of existing hardware isn’t optimized for the operations we want them to do. Creating custom silicon for ternary LLMs can accelerate inference by implementing and designing algorithms/circuits that only work for ternary LLMs.”
🧑✈️ GitHub Copilot now allows Enterprise users to fine-tune its suggestions on their own codebase. This would allow organizations to improve the AI-assisted code-completion suggestions specific to their private codebase. Though they do not reveal which model they are using (allegedly GPT-4o) they have also been testing Copilot interval with the newly released o-1 model which shows good improvements.
But in the age when so many coding datasets are openly available and the new o-1 models are available via API to everyone, what moat does Copilot have that gives it a competitive advantage over the likes of other tools like Aider (which is cheaper) or Cursor (which is much more integrated into the editor)?💫 OpenAI released o1-preview and o1-mini, its new generation of LLMs tuned for reasoning tasks, as in they “think” before spitting the final tokens, taking advantage of the increased inference time to compute. The way it works differently from other models is that it does some prior “thinking” via a CoT-like and then writes out the final answer. (and these thinking tokens are hidden from us for now because of the “competitive advantage” this puts OAI at, but you’ll still be billed for them though). But how good are they? On reasoning, coding, and creativity, they are PRETTY GOOD! Check the jump in some of the popular benchmarks to see for yourself. This is still a preview model that kind of teases the improvements but is still not as polished as they’d like it to be to use it as a drop-in replacement just yet.
The excellent writeup by Interconnects below goes deeper into how the o1 models work.
🔐 0x Digest
👨💻 Indodax, a leading crypto exchange in Indonesia, lost $25M to a hack, leading to a shutdown for investigation. All the hot wallets were breached, and North Korea’s Lazarus Group is one of the first on the attacker's guess list. They used 150 txns to clean transfer the assets to ETH, in the classic Tornado Cash way.
🌐 Vitalik Buterin wrote about why the era of rollups being glorified multisigs should be coming to an end. And that he will only endorse Layer-2 networks meeting “Stage 1+” standards from 2025, focusing on security and decentralization.
⭕️ zkVMs got some major developments:
The first RISC Zero proof has been verified on Starknet testnet!
Succinct Labs partnered with Optimism to launch "OP Succinct", a seamless way to upgrade any OP Stack chain to use ZKPs in 1 hour. This fast finality secured by zk proofs, brings the time down to minutes from the previous 7 days, which was for fraud-proof of optimistic rollup. Also, SP1's zk-proving prices have gone down by a major factor in the last 6 months, making zk-proofs more accessible to all.
🪪 0xCygaar gave their take on the future of wallets, that Passkeys + OAuth + Account Abstraction will be the future.
🆕 New Releases
Coinbase launched cbBTC, an ERC20 token that is backed 1:1 by Bitcoin (BTC) held by Coinbase. Enabling the BTC holders to experience DeFi apps on base.
🗺️ Atlas, a new blockchain for verifiable finance popped up, which runs on SVM and settles on Ethereum. (it's just like Eclipse?) And fun part is their parent org is named "Ellipsis Labs".
⛽️ Rabby introduced Gas accounts to simplify gas payments. Deposit USDT/USDC and easily pay gas fees for all your addresses on any network.
🛠️ Dev & Design Digest
🖊️ Oracle has owned the JavaScript trademark since 2009, as it acquired Sun Microsystem, which in partnership with Netscape created the first version of JavaScript. Their legal team has been renewing this trademark since then but not using it, and it's been frustrating the community. Ryan Dahl drafted a letter to Oracle to free the JS trademark. Go sign it!
🛑 Stop using SERIAL in Postgres. Instead, use identity columns, which have been supported since Postgres 10 (2017). Identity columns offer better permissions management, integrity guarantees, and ergonomics compared to serial.
🆕 TypeScript 5.6 is out! They have reverted a few things after the Beta and RC. The major changes in this latest stable release are:
disallowing nullish and truthy checks
new helper methods for iterators
adds "excluding patterns" for auto-imports.
support for arbitrary module identifiers, and shipped some improvements in the Language Server
⏳ ESLint v8.x end of life is October 5, 2024, right when we started creating a few ESLint rules, now it’s preferred to use ESLint v9.x, which brings flat config.
What brings us to awe 😳
💎 Synthetic diamonds are now purer, more beautiful, and vastly cheaper than mined diamonds. Beating nature took decades of hard graft and millions of pounds of pressure. It’s another big win for Material science as it was able to surpass another important natural substance.
🔧 iFixit might have created a USB-C powered Portable Soldering System, the last soldering kit you might buy. It’s nonetheless an engineering marvel and many have also tried to develop a USB-C soldering system earlier, but this attempt is surprisingly promising.
👾 Somebody built a font with built-in Syntax Highlighting that supports custom CSS themes! (And if you’re of the kind who wonders why anybody would even do this, here’s a good writeup about the case against syntax highlighting as well)
🔥 The new stripe.dev redesign is lit! Built with Next.js and entirely SSG. The footer screams “DEVELOPERS” like Steve Balmer.
Today I (we) Learnt 📑
🕣 An old tweet by Jared led us to dig into Why the Win32 epoch is January 1, 1601.
> Well, because “The Gregorian calendar operates on a 400-year cycle, and 1601 is the first year of the cycle that was active at the time Windows NT was being designed.”🤚🏼 Google was originally called Backrub and had this logo [Shared by Deedy on Twitter] (it’s funny cause it’s true).
🍌 Bananas are (slightly) radioactive and the banana equivalent dose (BED) is an informal measurement unit for ionizing radiation. (Stiens; Gate wasn’t wrong after all?). For reference, The Chernobyl disaster emitted radiation equivalent to 3 billion BED [Source].
🍬 Alpenliebe, breaks into German words - love (liebe) of the mountains (alps) and hence it means "love of the mountain". [Source: @nowreading_x on Twitter] (Nibbler P gets orgasmic every time we add an etymology reference. If you’re interested we do have a Twitter Etymology GC so ping any of us if you’d like to join that!)
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying 🔖
🪙 Polar: An open-source Lemon Squeezy alternative. Get paid coding on your passion.
📦 Slimify: A client-side, loss-less image compression app with batch support and WebP output format.
🔎 X Search Assistant by
— AI-powered companion for Twitter power-users to craft advanced Twitter searches effortlessly.🕯️Illuminate: Google’s tool to transform your content into engaging AI-generated audio discussions.
Builders’ Nest 🛠️
🪀 Shapeshift: a TypeScript library that maps arbitrarily structured JSON objects using vector embeddings. It uses semantic similarity to match keys between objects.
🖨️ safe-stringify: Serialize objects to JSON with handling for circular references.
✨ allegedly: A fast Solidity compiler for EVM simulations, written in TypeScript. (it’s nothing short of magic, what they’ve achieved)
↗️ cursorrules: A curated collection of various
.cursorrulesfiles for different use cases.
Meme of the week 😌
Off-topic reads/watches 🧗
😈 Compared to Perfect by Seth Godin. He explains “What’s on offer is never perfect, but what’s on offer might be exactly what we need right now.”
🤖 Optimize for bio cores first, silicon cores second by DHH, telling that bet on humans until AGI, chip prices are going downhill anyway.
🪜 Arnav Gupta (usually known for his spicy takes) wrote a thread about the concept of “growth” and how to visualize and utilize capacity, capability, and potential.
👋🏻 Bye Now by Seth, on why sometimes we need to stop pushing for commitments before it start hurting the trust.
Wisdom Bits 👀
“To lead people, walk behind them.” (in a non-creepy way ofc!)
— Lao Tzu
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev (Nibbler A has been using this as wallpaper for 2 weeks now)
Weekly Standup 🫠
Nibbler P had a fantastic week that ended excellently with a deadlift PR and a soul-sucking 10 km run! Would he do it again? Every effing time!
Nibbler A had another week full of building cool stuff from scratch, the week ended with him catching up with some family chores and his prep for upcoming trips. (new postcards incoming?!?)
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻
A publicly available knowledge graph containing over 240 billion rich data points across hundreds of thousands of statistical variables. Think of it as a vast, constantly expanding database filled with reliable, public information on a wide range of topics, from health and economics to demographics and the environment.
Retrieval-Interleaved Generation