#72
RIP Chromecast, MATH champ, Cheaper LLMs, Crypto hacks (again), Pari & Garuda SNARKs, Tauri, RBAC guide, Cloud disasters, image in emoji, Prestige, Unicorn in Scotland, and more
👋🏻 Welcome to the 72nd!
📰 Read #72 on Substack for the best formatting
🎧 Podcast version of this edition is available here → #72 | Recast
What’s happening 📰
🪦 Google is doing what it does best — kill the most user-beloved products and this time, it’s the Chromecast. After 11 years and over 100 million devices sold, they’re stopping the production of any new Chromecast devices and replacing the lineup with TV Streamer which… is exactly like the Chromecast in functionality but slightly larger in size and with the Gemini-powered “AI features” and connectivity with Google Home and Matter devices.
Feel free to speculate in the comments on how long until Google murders this one as well. We place the guess at < 3 years.
✨ AGI Digest
⚓️ Model Drops
🔢 LLMs that can solve complex mathematical have been quite in the news in the past few months, first because to the AI Math Olympiad (AIMO) and the success of Project Numina and then because of AlphaProof’s and AlphaGeometry’s together solving this year’s IMO at Silver Medal level. These development certainly caught the eyes of other researchers and the Qwen team quickly released the Qwen2-Math, a family of 1.5B/7B/72B models fine-tuned specifically for solving maths problems with the 72B achieving SOTA on the MATH benchmark and achieving 11/30 (the first model to reach double digits!) on the harder AIME 2024 benchmark. This pushes the frontier of mathematical reasoning in LLMs up a notch. What’s better, the model weights are available for anyone to download and try the model themselves.
Source: Qwen2-Math release blog 🎤 What the Qwen team did with Qwen2-Math above was impressive in its own record but they did not sit still and went on to release Qwen2-Audio the very next day. It supports audio upto 30s long and achieves SOTA performance in several audio & speech benchmarks.
👀 The OpenBMB team released MiniCPM-V 2.6, a vision LLM that achieves GPT4-V level performance on several vision benchmarks while being just 8B in size! It supports all of single Image, multi Image and video inputs and is well-suited for OCR and multilingual recognition especially on edge devices.
Source: MiniCPM-V GitHub Repo 🧹 Nvidia released Quality Classifier DeBERTa, a text-quality classification model that classifies documents into one of three classes based on the quality of the document:
High/Medium/Low. It supports context length upto 1024 tokens and can be a good tool for quality-filtering both your model training data, as well as responses (as long as they fit in the context length).
🤖 Mistral launched Mistral Agents which supports wrapping Mistral’s models with additional context and instruction on LeChat and the API. Their latest LLM, Mistral Large 2 is really good at reasoning and is a good model to layer on increasingly complex workflows with multiple agents. Though external tool use is still in the works.
💰 API and Price Updates
🧱 OpenAI announced a new GPT-4o endpoint with 100% adherence for following JSON Structured Outputs and a significant price reduction for GPT4o, now costing $2.5/M input and $10/M output tokens (was previously $5/M input and $15/M output), making GPT-4o one of the cheapest frontier models available out there (signs for an upcoming model soon?)
Source: Artificial Analysis 🐳 DeepSeek quietly implemented content caching on disk on their API, which caches content that is expected to be reused on a distributed disk array. When duplicate inputs are detected, the repeated parts are retrieved from the cache, bypassing the need for recomputation. This cuts the cost down the input costs for content with cache hit by a whooping 90%. And they straight-away did it for all their API calls without any modification from the user’s end!
Source: DeepSeek API Docs ♊️ Talking of cost reductions, Google was not behind and announced a Gemini 1.5 Flash price drop reducing the input price by 78% to $0.075/M tokens and the output price by 71% to $0.3/M tokens for prompts under 128K tokens.
🔐 0x Digest
⛓️💥 Another day, another hack in crypto, The Ronin Network bridge was exploited for ~$12 million, due to a critical oversight in a contract upgrade. They forgot to initialize a variable in the upgrade, and the variable was
_totalOperatorWeight, which was controllingminimumVoteWeightparameter, which defaulted to0, crucial for cross-chain verification, was effectively disabled. The sad part is, it’s not the first time for Ronin, they had a $624 million hack in March 2022, due to the same mistakes.🔃 Endurance, an on-chain gaming layer was stuck for a long time in implementing a “cross-chain messaging protocol” with their EVM as there is no IBC in that case. They finally recently settled for Hyperlane as their reliable permissionless cross-chain solution.
💳 Decaf Wallet enabled onboarding of US Accounts in 2 minutes. You or anyone can transfer USD to that account and you'll receive it as $USDC. Also, no fees for early access.
🚫 OKX said that “ those who interact with sanctioned platforms Garantex and Tornado Cash will face the risk of account closure”, without any exception. Why tho? After the recent rise in the usage of Tornado Cash by hackers, they came to this conclusion. But guys! “Tornado Cash” was supposed to be a feature of the ecosystem.
🔬 Researchers at Provable & University of Pennsylvania released their latest work on SNARKs: Pari and Garuda. Pari is the smallest known SNARK (just 1280 bits), while Garuda achieves a very fast prover by supporting "free" addition gates and custom gates together.
🛠️ Dev & Design Digest
🗑️ Jake, Surma, and Jason (yes, the first two are the OGs from HTTP 203) were hacking something (as always) and found out Garbage collection in JS didn’t work quite as they expected and they dug a little deeper to get closure (pun intended). Not something super new or something that’ll fix your memory leaks, but good to know 👍🏻
🖥️ Tauri released 2.0 RC, what was it again? Ah! It’s an Electron-like toolkit for building cross-platform desktop apps using a single codebase, but it is written in almighty blazingly fast Rust.
Tauri has been here for a few years, but the battle was definitely about UX vs DX, as Electron gives you smooth DX, but ships huge binaries and memory issues. On the other hand, Tauri is supposed to give smaller builds and performance in the application to a user, but the developer has to write Rust (code for bad DX). With this v2.0 RC, Tauri is breaking its barriers and making the DX trade-off smaller.🖋️ Excalidraw shipped 3 new fonts, revamped our hand-drawn font (Excalifont), swapped sans (Comic Shanns) and monospace fonts (Nunito), and added a heading font (Lilita One).
Source: Excalidraw’s new interface 🔐 WorkOS published The Developer's Guide to RBAC (role-based access control). The post dives into how we ignore authorization at first (and just implement Authentication), which is fine, but eventually, if your SaaS works with teams, building some kind of authorization is inevitable.
What brings us to awe 😳
💸 Google accidentally deleted Unisuper's (a pension fund that manages $135 billion in assets) private cloud after a year because of a missing parameter in the creation of the cloud. Though they were able to recover most of the data,
the lesson here for us it to always have backups of the data and to have a plan for how to recover from a disaster like this.
🎨 tldraw team announced
@tldraw/sync, a new package in their tldraw SDK suite. As a developer just by using a hook from the package, you can make your existing tldraw apps collaborative. (like in a snap, as it should be)📷 A guy used a bunch of modern image-generation and restoration models to restore a hidden image inside Apple’s camera emoji. Although this isn’t the exact image, the surroundings and color palette are close enough.
Today I (we) Learnt 📑
😮 Prestige comes from the old French word for deception, funny how it used to be a negative (illusion, tricks) word and now “respect and admiration” (people name their company after it). [Source]
yeah yeah! Nolan knew it. 🦄 Guess what is common between over-funded startups and Scotland’s National animal, 🥁 they both are called “Unicorns”. Shame that California (Silicon Valley specifically) settled for “grizzly bear” as their state animal, step-up tech bros.
🤦🏻♂️ As per the POSIX recommended guidelines for CLI, Arguments are options if they begin with a hyphen delimiter (‘-’). Multiple options may follow a hyphen delimiter in a single token if the options do not take arguments. Thus, ‘-abc’ is equivalent to ‘-a -b -c’. You must be like, so what? get to the point.
Well, it leads to a funny confusion here: “Why does Brotli Command Line only allow up to level 9?” (for context, Brotli has max compression 11).
Why is it funny? (or was at least funny when I found out about it) because this relates to the single token standard after the hyphen, not allowingbrotli -10, so they have different flags for this.🐳 The company High-Flyer which is behind DeepSeek is a quant fund.
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying 🔖
🌏 Sounds of Earth: Hear sounds from around the planet and set your ambiance.
🎵 vscode-spotify: Use Spotify inside VS Code, because 🤷🏻♂️. (we don’t know why it asks for “Take actions in Spotify on your behalf”)
🥷 Button Stealer: Steals buttons from the websites you visit. Do your usual online stuff and watch the collection of stolen buttons grow. (good artists copy, great artists steal)
🥧 Pie Menu: a macOS app to evolve your workflow with a radial menu customized for your active app.
Builders’ Nest 🛠️
📄 wordgrinder: A word processor that gets the hell out of your way and lets you do some work.
🎶 seek-tune: An implementation of Shazam's song-matching algorithm in GoLang.
🔀 Burrow: a globally distributed HTTP proxy via AWS Lambda. It is completely compatible with the standard Go
*http.Clientand provides an implementation of thehttp.RoundTripper.🆘 screenshot-to-code: A simple tool to convert screenshots, mockups, and Figma designs into clean, functional code using AI. Now supporting Claude Sonnet 3.5 and GPT-4O!
Meme of the week 😌
Off-topic reads/watches 🧗
🤫 Silence vs Noise by Seth Godin reminds me of Naveen Kasturia's fake quoting MK Gandhi on “put your head down and work”.
⚓️ Little Sister Syndrome and how it affects the growth of VCs and also, forces them to switch firms. The syndrome is basically “improper anchoring” of someone based on when you first meet them.
🏷️ The Mississippi River paradox by Seth Godin, on “When the label stops being useful, we should change it.” (actually can be force related to Little Sister Syndrome above)
🚶♂️ If We Want a Shift to Walking, We Need To Prioritize Dignity by Sean Hayford Oleary talks about encouraging more walkable areas by prioritizing the dignity of the pedestrian experience and considering factors like shade, convenience, enclosure, and engagement in addition to just compliance with rules and safety.
Wisdom Bits 👀
“You are only as beautiful as the many beautiful things you do for others without expectation.”
― Janelle Monae
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev
Weekly Standup 🫠
Nibbler A built and shipped some stuff earlier in the week and has been resting (it’s hard, as we heard from him) for the latter half of the week. He finished this edition early and onto some personal chores and commitments.
Nibbler P got back to running 5k this week (ran his second-fastest 5k) and did a lot of contemplations and reorganizations in his work and life.
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻