#71
GemmaScope, FLUX.1, TS5.6, Million lint, WebGPU on Native, Nintendo Kawaii, Terra exploit, Ghost's ParEVM, Daylight, Orbit, URI Protocols, SQLite > fs?, Smartphone Pinky, rrweb, The last bit and more
👋🏻 Welcome to the 71st!
📰 Read #71 on Substack for the best formatting
🎧 Podcast version of this edition is available here → #71 | Recast
What’s happening 📰
🔻 INTC 0.00%↑ reported a loss of $1.6B for Q2 2024, which led to their stockholders not being happy and led to stock price going down too. Now they’ve come up with a plan to save $10 billion in costs by laying off over 15,000 employees and stopping ‘non-essential work’.
🤝 Google Cloud now has a dedicated cluster of Nvidia GPUs for Y Combinator startups. As if there were not already enough AI startups in the last few YC Batches (finally, those machines will be put to use? jk).
✨ AGI Digest
⚓️ Model Drops
🤏 The Google-Deepmind team released new members of the Gemma family — Gemma2-2b, Shield Gemma and Gemma Scope, and also upgraded Gemini 1.5 Pro to a new experimental checkpoint in the Google AI Studio which apparently outperforms GPT-4o on many lmsys benchmarks (but not coding though!):
The Gemma2-2b has excellent text formatting for a model its size and even beats the Mixtral8x7b on the lmsys arena that measures average human preference on input-output pairs b/w different models.
ShieldGemma is a family of AI safety classifier models, designed to detect harmful content in the input and outputs of models with high accuracy. Like Gemma, it comes in sizes 2B, 9B and 27B.
Gemma Scope is a suite of freely available SAEs (sparse autoencoders) to take a peek into Gemma’s activations and understand what features the model looks at when processing a particular input sequence. Many research labs are focusing on mechanistic interpretability and seems like we’ll be seeing much more interesting work in this going forward. Open up them LLMs!
🏞️ Black Forest Labs released FLUX.1, a SoTA text-to-image diffusion model that absolutely mogs the likes of SDXL, SD-3, and Midjourney in real-looking images. It comes in three sizes — schnell (meaning ‘quick’ in German), dev, and pro, all with excellent coherent image generation with very apt typography. You can try them right now on fal.ai or replicate APIs.
Source: @bindureddy on Twitter 🧵 OpenAI has a new endpoint for GPT-4o with Long Outputs upto 64k tokens but with increased costs of $6/M input tokens and $18/M tokens (a $1 increase in input and $3 in outputs). Interesting to note is that the 4o-mini offered the 64k output length at launch and its cheap pricing already accounts for the increased computation needed to process the longer token sequence lengths.
🖨️ Stability AI released Stable Fast 3D, an OSS model to generate 3D assets quickly, from a single image input. And boy, it is fast with taking just 0.5 seconds per 3D asset generation on a GPU with 7GB VRAM!
🎙️aiOla released Whisper-Medusa, a speech-to-text model built on top of OpenAI’s Whisper model with multiple decoding heads claiming to be operating up to 50% faster with no loss in performance. They released both the model weights and code for devs to try it out.
🎏 The LLM Hallucination Index: RAG Special published by Galileo compares and evaluates top LLM models testing them to adhere to given context for RAG tasks on both short and long contexts. Google’s Gemini models consistently bag the most affordable model ranks while Claude 1.5 Sonnet is the overall winner in terms of accuracy on all context lengths.
🧠 While we wait to get access to OAI’s SearchGPT, the Chinese LLM company InternLM team released MindSearch — an open-sourced AI search engine framework, claiming performance comparable to Perplexity Pro. This app first decomposes your search query into atomic sub-questions as nodes in a graph and then progressively extends the graph based on the search result, trying to mimic how humans generally search and discover stuff when browsing.
👯♀️ GitHub launches GitHub Models making exploring various AI models (LLMs for now) accessible natively within GitHub. You can also test and compare different models in the model playground and spin up a GH Codespaces environment to try things out on Azure AI easily. Feels kinda like a mix of OAI Playground + HF Deploy button in one place.
🔐 0x Digest
🥊 The Terra blockchain was halted Wednesday after an exploit drained ~$4 million worth of tokens. That included 60M ASTRO tokens, driving its prices down to drain.
👷🏻♂️Ghost team shipped par.tryghost.xyz, an app that lets you check how parallelizable any Ethereum block is by visualizing transaction dependencies, and it updates in real-time. This answers your question of “How parallelizable is the EVM?”.
🪙 As if Mert was already not carrying Solana, his company Helius Labs became the #1 validator on Solana, having a ~3.4% stake.
☀️ DePIN company Daylight raised $9M in Series A led by a16z. They are building a decentralized protocol that lets users connect their energy devices, such as thermostats, batteries, electric vehicles, and solar inverters, to its app and earn rewards. They pool together capacity from the flexibility of connected devices and make that available to energy companies.
⛓️💥 Arbitrum will expand Orbit support for ecosystems other than Ethereum. The proposal has 83% support so far from the community. They mentioned that they’ve received requests from developers wanting to deploy Orbit on non-Ethereum chains.
🛠️ Dev & Design Digest
🆕 TypeScript 5.6 Beta is here and as always brings exciting syntactic sugar and checks to prevent us from burning our fingers with almighty JS.
Here are some major ones, we liked (& understood):Disallowed Nullish and Truthy Checks just made it harder for us to write a
if blockthat always runs.The
--noCheckoption is available to set you free from types, raw dogging to production.Allow
--buildwith Intermediate Errors, there are no more early exits of build on the first error, rather it’ll now throw intermediate errors and output files will be generated on a best-effort basis.
✨ Million Lint by MillionJS released 1.0.0-rc.84 which adds in cool debugging features like:
Get parents/children on hover (yes hover over the component in VSCode and see its relationships in React DOM)
Locate render triggers (setState & interactions)
Next.js --turbo support
Record & replay the DOM
📱 After AppJSConf, William and the team were wondering what the shortest path to a public technical preview of React Native WebGPU would be. And they just showcased a viable demo accessing WebGPU on ReactNative.
📄 The team from the good old arXiv launched alphaXiv, an open discussion forum for arXiv papers where researchers and authors can comment and discuss the work. It fits right in with the current arXiv by just replace arxiv.org with alphaxiv.org!
🦖 Ryan Dahl pressed again where it was hurting, and talked about What they missed when they decided to do “HTTP Import” in Deno why they came up with JSR, and why the hell we have package.json and node_modules back in town. Hinting towards Deno 2 in September. (not sure, the man had my attention for the third time and ended up shilling Deno 2, I hope this time it works.)
What brings us to awe 😳
👾 Some folks (to be precise @Wesk and Yveltal) in the BitBuilt community have designed and are building a keychain Wii named Nintendo Kawaii. It’s 60x60x16mm in size, super cute, and will be an engineering marvel once built.
Source: Linked forum above 💽 SQLite is 35% faster for reading and writing small blobs. Also, it uses less disk space because it packs blobs more tightly than the filesystem.
😭 By just replacing is-number dep with a single line of code in to-regex-range, we’ll save traffic as much as 440GB weekly. (they merged it last week)
📣 The ChatGPT Advanced Voice Mode (that’s rolling out to a select ChatGPT Plus users) was asked to count as fast as it could do 50 and it stopped to catch its breath like a human would! [Video taken from the linked tweet]
Today I (we) Learnt 📑
⛳️ Explanation of the Olympic Flag by Sketchplanations.
Source: Sketchplanation 🌐 Protocols in URI, what to infer from them, and how to register one? by ThioJoe. It’s fascinating how much OS uses it in bulk. (think of them as deep links only)
💰 Loom paid $150k after negotiation and a smart lucky deal to acquire their domain “loom.com” in 2018. (real estate of the web)
🤷🏻♂️ People like us who use smartphones a lot, usually end up having dents in the pinky finger of the hand we hold the phone with. They are called “Smartphone Pinky”. We just checked and we do have it, too. 😭
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying 🔖
⛳️ FlagMatch: There are 236 countries in the world. How many can you match with their flags?
📢 ElevenLabs Reader App: Listen to anything on the go with the best-sounding voices from ElevenLabs (Nibbler P has been using this a lot to listen to some light stuff when doing other chores around the house)
🪶 Prompt Poet: Streamlines and simplifies prompt design for both developers and non-technical users with a low code approach.
🔪 CyberChef: The Cyber Swiss Army Knife - a web app for encryption, encoding, compression, and data analysis.
Builders’ Nest 🛠️
🖥️ mactop: Apple Silicon monitor top written in pure Golang.
▶️ rrweb: a tool for recording and replaying users' interactions on the web. (yes, LogRocket uses this)
🔦 torchchat: Run PyTorch LLMs locally on servers, desktops, and mobile
🥞 stack-pr: A tool for working with stacked PRs on GitHub.
Meme of the week 😌
Off-topic reads/watches 🧗
🚪 It seems… by Seth on how this five-letter modifier opens the door to discussion.
👨🏫 Randy Pausch’s last lecture at Carnegie Mellon University titled "Really Achieving Your Childhood Dreams" encapsulates his reflections on life, his legacy, and the importance of pursuing one's dreams (The last ‘headfakes‘ from in the video really got us all emotional).
😌 Some gentle reminders from the 6-year-old & and top regret of the dying.
ࠁ The last bit by Seth on the importance of the last 1% (has nothing to do with this being the last bit in this Nibble)
Wisdom Bits 👀
“And once the storm is over, you won't remember how you made it through, how you managed to survive.
You won't even be sure whether the storm is really over. But one thing is certain.
When you come out of the storm, you won't be the same person who walked in. That's what this storm's all about.”
― Haruki Murakami
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev
Weekly Standup 🫠
Nibbler P had a super busy but fun week with hitting some deadlines at work to going explore mode around the streets of Jodhpur to run some errands.
Nibbler A had quite a week debugging, building, and discussing software. and ended up finishing some chores over the weekend.
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻