#65
NVIDIA tanking, Micro-Reactor, BigCodeBench, DeepSeekCoder-V2, Sonnet, Florence-2, Meta Chameleon, zkVM 1.0, Fox's L2, React 19 delay, TS 5.5, NumPy 2.0, 0x Drama, Gaming lay-offs, time.fun and more
👋🏻 Welcome to the 65th!
📰 Read #65 on Substack for the best formatting
🎧 Podcast version of this edition is available here → #65 | Recast
What’s happening 📰
📈 NVIDIA briefly became the world’s most valuable company before tanking down below Microsoft and Apple. This success has backfired them though as their engineers – now holding stock options worth millions – are opting for early retirement, creating a talent drain for the company.
Source: Tweet by @trungtphan ⚛️ Rolls Royce (yes, the luxury car company) unveils Micro-Reactor – a compact, safe, and transportable nuclear power solution that can provide up to 1-10 megawatts of power allowing powering remote civil and industrial locations, as well as space applications. This sounds like straight outta sci-fi novel. What a time to be alive!
✨ AGI Digest
⚓️ New Model and Leaderboard Drops
🌸 The BigCode team announced the BigCodeBench Leaderboard - a robust coding leaderboard featuring complex, user-oriented instructions for each task, including clear functionality descriptions, input/output formats, error handling, and verified interactive examples. The Verdict from this dashboard? GPT-4o still tops the rankings here but the newly released Claude 3.5 Sonnet and the open-sourced DeepSeekCoder-V2 are not far behind (and a better bang for your buck actually).
🐳 DeepSeek is back as the king of open-source coding models with the release of DeepSeekCoder-V2, which excels in coding and math and punches models far above its size. It has a 128K context length and two sizes – a 230B (available with API access and a chat interface on their site) and a 16B (good as a local model and is quite fast to inference thanks to its only 2.4B activated params thanks to its MoE architecture).
📜 Anthropic casually dropped Claude 3.5 Sonnet while teasing the Haiku and Opus from the same family to be announced later. This takes Claude above OpenAI in many rankings, especially with their cheaper pricing that costs $2/M lesser on input tokens than the GPT-4o (same price on outputs). On Aider’s Code editing benchmark, it quickly dethroned the reigning DeepSeekCoder-V2 (which could enjoy only 4 days of the majestic position). But what garnered the most attention was the artifacts – a feature where you can ask Claude to generate docs, code, mermaid diagrams, vector graphics, or even simple games.
Source: Post by @alexalbert__ on Twitter 👁️ Microsoft released Florence-2, a family of SOTA 200M & 800M vision foundation models with multi-task capabilities like captioning, object detection and segmentation, OCR, phrase grounding, and more! Perfect to be deployed in Edge.
💐 Meta released a bunch of new models building on its previous research:
Meta Chameleon: 7B & 34B language-vision models
Meta Multi-Token Prediction LLM
Meta JASCO: text-to-music models
Meta AudioSeal: audio watermarking model
🎊 Product Improvements
🏞️ RunwayML demos Gen-3 Alpha, its latest generation of text2vid and img2vid models and the quality, fidelity, consistency, and motion is right there with Sora. And just like Sora, the access it not opened yet. So take everything with a grain of salt until you see the capabilities for yourself.
📷 Chatbot Arena now supports image uploads so you can challenge GPT-4o, Gemini, Claude, and LLaVA and see which fares better.
🔗 Notion introduced AI Connectors through which when you ask a question, Notion AI will also surface relevant information from your connected apps (such as Drive, Slack, Jira, GitHub, etc.), citing specific sources it referenced. They basically did an RAG over your connected source
🏎️ Character.AI dropped a blog post about optimizing AI Inference on their servers and they have a much much larger load than they get credit for, with a whooping 20k QPS (around 20% of what Google serves)! They share several techniques on how they improve their latency bottleneck – the KV cache and get 20X improvements without regressing quality.
🚨 OpenAI appoints former NSA head Paul Nakasone to the board, Snowden warns to not trust OpenAI’s product anymore calling it “a willful, calculated betrayal of the rights of every person on Earth”.
🔐 0x Digest
✔️ FOX Corp. launches the 1st enterprise L2 chain named “Verify” on Gelato, powered by @0xPolygon CDK, focusing on verifying integrity and control of content.
RiscZero introduced SOTA zkVM 1.0, the world's first and most performant, production-ready zkVM. This makes zk proof available across any chain and cost-effective.
🎬 CertiK (white-hat crypto security firm) found a series of vulnerabilities in Kraken Exchange. Things got interesting when they took 5 days to disclose the vulnerability and withdrew $3M worth of tokens as PoC and didn’t disclose it with vulnerability. Karken didn’t take it well (of course) and got pissed, also reportedly “Threatened individual CertiK employees to repay a mismatched amount of crypto in an unreasonable time even without providing repayment addresses.” Well, never a dull day in crypto. They have returned the funds now, but the drama was inevitable.
🔻 The German Police Department is selling the seized BTC (~50,000 BTC, worth $3B) from movie piracy sites (over the years) and that might cause temporary short-to-medium-term volatility.
🛠️ Dev & Design Digest
🗒️ A detailed 9-question Survey on "What do GenZ software engineers really think?" by Pragmatic Engineer. It goes from what excites them, makes them respect seniors to what triggers them to leave a workplace kind of details.
🆕 TypeScript 5.5 is out with Inferred Type Predicates, Performance and Size Optimizations, and Support for New ECMAScript
SetMethods, Regular Expression Syntax Checking, and more.🆕 NumPy releases Numpy 2.0, eighteen years after the release of the first version, and brings with it a cleaned-up and streamlined API, improved type system and scalar promotion rules, Windows support enhancements, and much more. Since this release breaks backward compatibility please refer to the official numpy 2.0 migration guide and the new ruff linter rules when you upgrade.
⏳ React 19 just got delayed. Why you ask? Well, we shared Wes's tweet in the last edition where he shared some banger features. Except there was one thing none noticed until someone did. There was a change in the way React Suspense works, earlier (in React 18) if you put two components both doing fetch requests inside them, they'll fire in parallel. But in React 19 RC, it was being done sequentially (waterfall).
A lot of libraries like react-three-fiber and react-query leverages how Suspense works to offload async tasks and this change would mean breaking them. This soon reached the React team and they tried to explain that Suspense was never intended to be used on the Client side the way it is used right now. But never the less they are holding the release until they find a better fix than breaking hearts.
What brings us to awe 😳
🗺️ Are there 195 countries? or 282? or 368? or It depends? by MapMen
🧵 The video game industry has set a tragic new record for in-year layoffs, in just the first six months it saw 10,900 layoffs wrt 10,500 in 2023 (whole).
⌚️ Time.fun, a platform that allows you to tokenize and sell your time in minutes by creating a tradable asset exclusively on Base where fans can buy, sell, and redeem your time for ETH. (wow, just wow!)
Today I (we) Learnt 📑
👟 Woodland started as a Blind Copy of Timberland and gained enough trust and traction to kick the original out of the Indian market when it entered.
🤔 Go's
json.Marshalwill marshal empty slices{}asnulland not[]. But at the same timefmt.Printwill mislead you as it'll give you[]for the same array.
Here’s a Go Playground Link to play around with.♟️ Peter Thiel was a professional chess player with a FIDE rating of almost 2200. [Shared by @AkJn99]
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying 🔖
📊 Rows: Excel but with AI as a first-class citizen. Query the VLOOKUP in plain text, pivot the tables, and more.
ℹ️ macOS Icons: Biggest library of free macOS app icons.
💊 Amphetamine: A Powerful keep-awake utility for Mac.
Builders’ Nest 🛠️
🥇 rerank-ts: rerank library for easy reranking of results
☁️ piku: The tiniest PaaS you've ever seen. It allows you to do git push deployments to your servers.
✍️ claudette: A wrapper for Anthropic’s Python SDK that makes Claude's awesome features easier & more powerful for Pythonistas.
💩 shittier: a code formatting tool that aims to make your code look as terrible as possible. (don’t ask us the why)
Meme of the week 😌
Off-topic reads/watches 🧗
🛒 The Coney Island Problem by Seth Godin, raises the question of do we appreciate “The Long Tail” in its real form.
✏️ Non-Fiction Writing Advice by Scott Alexander has several gems for effective nonfiction writing. [Shared by Sankalp]
👋🏻 Humble Exits by Morgan Housel. A note on how it’s rare for people to walk away from success at the peak.
Wisdom Bits 👀
“The garden of the world has no limits, except in your mind.”
— Rumi
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev
Weekly Standup 🫠
Nibbler P experimented with some new LLMs this week and is getting more consistent with his runs. He also tasted some really spicy noodles and now he's addicted.
Nibbler A shipped some stuff at work this week and is onto weekend chores and reading again. He went to play some 🏸 twice this week.
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻