#70
Tesla FSD, OSS v/s Closed AI, IOI Silver, TS in Node, Rust got LAZY feats, tan to get dimension, Doom on Grafana, Circle STARKs, Canarying, await setTimeout, poku, parseltongue, swappy and more
👋🏻 Welcome to the 70th! (almost 60% of the year is over, fyi)
📰 Read #70 on Substack for the best formatting
🎧 Podcast version of this edition is available here → #70 | Recast
What’s happening 📰
🚘 Tesla rolled out FSD (Full Self-Driving) v12.5 to its cars early last week and it brings a lot of improvements in the hands-free driving. Tesla owners are delightfully surprised with how well it can manoeuvre them to places more confidently and predictably, bringing it closer to achieve the SAE Level 4 of autonomous driving.
🗺️ Apple Maps launches on the web in beta, now allowing devs using Apple’s MapKit JS to link these maps into their web apps as well.
✨ AGI Digest
⚓️ Model and Dataset Drops
🦙 Meta opened the week with a bang by releasing Llama 3.1 8B, 70B and 405B which are trained on a whooping 15T tokens of both real-world and synthetic data. They also extended the context length to 128k with multilingual support and the ability to use tools, all of which bring the 405B the top open-weights model in many domains. It’s already available for inference via API on popular model providers, integrated into the Meta AI assistant in WhatsApp and Instagram, and also available on the web at meta.ai.
Llama 3.1 makes the gap closer b/w open-weights and closed-weights models | Source: Llama-3.1 release blog 🐳 DeepSeek updated its DeekSeek Coder V2 model that brings it to #2 on the Aider Code Editing Benchmark, very close to Claude-3.5-Sonnet, the current coding champion. They also upgraded their API to support JSON mode and function calling and some more changes in beta.
🎐 Right after the Llama 3.1 drop, Mistral dropped Mistral Large 2, a 128B-parameter model with similar context length and multilinguality and comes close to Llama3.1 405B on some coding and and math benchmarks. Though it does not perform as well on the Aider Benchmark, it comes quite close to the 405B and that too after being more than thrice its size. Mistral made it available on la Plateforme and le Chat and released it under the Mistral Research License.
This is something quite unheard of in models till now as many models tend to outright hallucinate when not confident enough in their outputs ⚖️ Stability AI released Stable Video 4D, a diffusion model which transforms a single object video into multiple novel-view videos from different angles/views. It also allows you to specify camera angles, tailoring the output to meet specific creative needs. The best part? They actually released the weights at the time of announcements this time with a Community License.
🍃 The Salesforce AI Research team released MINT-1T, a 1T token (the first of this size) multimodal interleaved dataset consisting of HTML documents, PDFs and ArXiv papers.
📐 The Google DeepMind team leapfrogged in mathematical problem-solving with AlphaProof and AlphaGeometry. These both combined IMO silver-medal-standard AI models, just a point shy of getting a Gold and the team achieved it with a combination of a fine-tuned Gemini to convert natural language problem statements to Lean (a formal language to represent, verify and prove mathematical statements) and then uses the AlphaZero RL algorithm (of the chess, shogi, and Go fame) for solving the equations. Give or take a year and we’ll soon get the Gold!
BTW, solving these problems do not mean it solved all of them faster than humans. While it did complete some of them in a few seconds, one one of them it took 3 days to solve (while IMO is held in two rounds of 4.5 hours each on the same day). 🛟 OpenAI Updates
🧮 Also after the LLama3.1 release, OpenAI announced GPT 4o-mini finetuning for tier 4 and 5 users with the first 2M training tokens a day are free, through Sept 23.
🔎 OpenAI also announced SearchGPT (in waitlist), a LLM powered search and a head-on Perplexity competitor, focused on giving fast and accurate answers with clear and relevant sources. They are still rolling out the access and people who used it say it’s quite fast but still basic in features compared to Perplexity.
🤖 And last week closed with a new chart of where open-weight and closed-weight models stand right now (on MMLU) and the gap was never so close! Ofc with time benchmarks get saturated and MMLU is certainly not the best way to compare models but t still holds some insights on how developed models have become and how fast open-source is catching up. Though, large companies still have a few models up their sleeve and could drop them anytime soon widening the gap.
Source: maximelabonne on Twitter
🔐 0x Digest
🌖 MoonSwap got hacked, as one of their developers installed a phishing app to join a call with scammers who pretended to be a VC (wtf!). And since the botnet has access to MonoSwap-related wallets and contracts, the company has advised to not add liquidity or stake in our farming pools at the moment.
🔪 WazirX hacked but Liminal is to blame? by Mudit Gupta, a deep dive into who to blame finally about the WazirX hack last week.
📉 Wall Street is bearing on ETH ETF due to the lack of a compelling narrative like Bitcoin, competitors like Solana, them not being able to understand it & Vitalkin not showing up on Bloomberg or CNBC explaining what Ethereum is.
🌍 Crypto x Countries
🚗 Ferrari to expand crypto payments for cars to Europe following US launch.
SupportCrypto.org allows U.S. politicians to accept crypto donations on Base.
Russia is considering the official legalization of stablecoins for international transactions to simplify cross-border payments for Russian companies amid ongoing sanctions.
💫 Vitalik writes about Circle STARKs, a variant of stark that operates on a smaller field, leaving less wasted field (as it uses a smaller field 2³¹ over 2²⁵⁶) than regular STARKs. The methods leverage Circle FRI & Circle FFT over regular FRI & FFT in the process, as Vitalik mentions these are relatively less additions to complexity as compared to the efficiency we get.
🛠️ Dev & Design Digest
✨ NodeJS adds support for --experimental-strip-types, which lets you execute TypeScript inside the Node without any 3rd party bundler/transpilers, during the process, it transpiles TS in JS and strips the types. Here’s a video from Wes showcasing the same.
And guess what, this leads to the first piece of Rust entering NodeJS too, to support the TypeScript wrapper, they built a loader amaro, which is a wrapper over @swc/wasm-typescript.📲 Kadi Kraman from the Expo team wrote a detailed guide for React (Web) devs, on Everything they need to know about building their first React Native application.
🔳 CSS Grid has been GA for the last 7 years, but still a lot of us don’t use the Grid area (because we like to flex 🤷🏻♂️) where it could be used. Ahmad is not happy about this fact and wrote a guide on How to leverage CSS Grid Area today.
🦀 Announcing Rust 1.80.0, while already being the choice of JS Bundler devs, it keeps improving and adding more optimizations. This version’s highlights include:
LazyCellandLazyLock, both of these types delay the initialization of their data until first access. LazyLock is thread-safe and LazyCell uses local thread.Automatic checking of
cfgnames and values at compile-time, i.e. checks all reachable#[cfg]matches a list of the expected config names and values.Exclusive ranges in patterns: Rust used to allow only inclusive (
a..=bor..=b) or open (a..) ranges, which were confusing and to be honest might lead to missing points in range. With this version, they’ve enabled exclusive ranges (duh! should’ve been this in the first place). So, now you can writea..bor..b, which iterates from a to element right before b.
What brings us to awe 😳
🔓 Anyone can access deleted and private repository data on GitHub by Truffle Security, where the dive into details and non-intuitive design of GitHub’s fork network and how anyone can access any commit using its hash, as long as one of the forks from the network exists. They’ve thoroughly covered 3 cases in their blog, and GitHub for now has responded that it is by design the way it is, and they’ll make it more clear to the user.
📐 Get the screen width & height without JavaScript, and compute them in pure CSS leveraging
@propertyand trigonometry functions. It’s a neat trick on how we usetan(atan2(h, w))for division in CSS ascalccan’t do that yet.🚧 Someone mapped all the USA Traffic Deaths from 2001 to 2022 and added it to a single portal → Roadway Report
☠️ If you have a laptop or desktop that has an Intel CPU with "vPro" on the sticker there's a chance the management engine in your CPU is just hosting a web server at all times. It's at port 16992.
This opens scopes to vulnerabilities and the thread discusses, how it could be exploited. [Source: Thread by i_lost_my_bagel]
Today I (we) Learnt 📑
⏲️ Node has setTimeout that you can await when imported from
timers/promises.
So, replace all yourawait new Promise(r => setTimeout(r, ms));withawait setTimeout(ms);. Sleepwell 😜🪶 “Canarying” (in software at present) comes from the phrase “canary in the coal mine.” In the early 20th century, miners took caged canary birds with them underground. The bird has a lower tolerance for toxic gasses than humans do, so if the bird stopped chirping or fainted, it was a warning sign to miners that gas was present, and for them to evacuate. [Source: The biggest-ever global outage: lessons for software engineers by
]🔫 You can configure and play Doom on a Grafana Dashboard. And there is a goldmine on templates available at play.grafana.org
🇯🇵 Why Japanese websites are all design very information-dense
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying 🔖
📝 Papermark: the open-source DocSend alternative with built-in analytics and custom domains.
㊗️ translate.photo: Translate Creatives to 75+ languages with 1 click.
🤨 why-is-node-running: Node is running but you don't know why? This tool comes in handy and lists all handles keeping the process running.
🐍 parseltongue: A powerful prompt hacking tool for red teamers, developers, linguists, and latent space explorers.
Builders’ Nest 🛠️
✨ swapy: A framework-agnostic tool that converts any layout into a drag-to-swap one with just a few lines of code.
🧪 poku: A cross-platform test runner that returns the JavaScript essence to testing.
📌 nho: 1KB Web Component Abstraction.
🍪 gpu.cpp: A lightweight library for portable low-level GPU computation using WebGPU.
Meme of the week 😌
Off-topic reads/watches 🧗
⡔ Patterns and Chaos by Seth Godin: As humans, we are hard-wired to look for patterns, but sometimes there are none, what do you do then?
🐳 Unveiling DeepSeek: A more extreme story of Chinese technological idealism (translate a website from Chinese to English): A rare interview of Liang Wenfeng, the founder of DeepSeek AI and his thoughts on Innovation, Open Source Software, and the place of China in the Technological Innovation.
🪜 Doing it step by step by Seth Godin: When you ask an AI to do it step by step, it’s less prone to make mistakes and so are humans.
🧘🏻 Why Fortune Favors the Focused?: Focus doesn’t guarantee your success, but it will certainly increase your odds.
Wisdom Bits 👀
“Success requires first expending ten units of effort to produce one unit of results. Your momentum will then produce ten units of results with each unit of effort.”
— Charles J. Givens
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev
Weekly Standup 🫠
Nibbler A had quite a design, build, and show week, and ended up getting heavier towards the end, where he took a break day (just from the screen, not life) and took 12h nap, while his Mac has an uptime of 48 days, completely leaving Nibble to the next day.
Nibbler P was mostly at home working because of his sprained leg this week. Now that it has healed towards the weekend, he did a quick visit to Jodhpur with his parents and came home really tired, contributing his part to the delay.
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻