Psst… if you have a second to spare and want The Nibble to improve, let us know your thoughts by filling out a small questionnaire. We’ll reinforce the good parts and weed out the bad ones. 💪
🌞 What a time to be alive, this summer, YC will be giving grants to college students to work on their own technical projects. They are calling it “Summer Fellows Grants”.
They’ll be handed over $20,000 in cash and $90,000 in compute credits. 📅 Submit your applications by March 28th, 9:00pm PT. Selected folks will be invited to attend AI Startup Schooltoo.
✨ AGI Digest
⚓️ Yes, it’s only Tuesday and we already have so much exciting stuff
👨💻 After weeks of speculation, testing and safety blogs, Anthropic released Claude 3.7 Sonnet and it was worth the wait! This model can respond both with and without reasoning and only gets sweeter than the previous model especially on Coding and Mathematical tasks.
Mogging OpenAI on SWE-bench verified without even thinking
Both on Aider Polyglot Leaderboard and BigCodeBench-Hard, it performs neck-to-neck with the competition without any thinking and though we are yet to see the complete result with thinking on all of the leaderboards, from what we have seen till now, it is going to be right near the top!
Along with 3.7 Sonnet, they also released Claude Code, an agentic coding tool in Terminal (much like Aider on steroids) which directly uses Claude via API to search, read code, edit files, write and run tests, commit and push code to GitHub and even use your command-line tools! It is available as an npm package but is currently on a waitlist.
🐳 DeepSeek is leaving no stone unturned to build the true “Open” AI by doing an “Open Source Week” where they open-source five projects that constitute the building blocks of their online services which they use in production. In their words, “No ivory towers - just pure garage-energy and community-driven innovation.“
On Day 1, they released FlashMLA, an efficient MLA (MultiHeaded Latent Attention) decoding kernel for Nvidia’s Hopper GPUs with BF16/FP16 and Paged Paged KV cache with a block size of 64. Using CUDA 12.6, it can achieve up to 3000 GB/s in memory-bound configuration and 580 TFLOPS in computation-bound configuration on H800 SXM5. The best part is that it was built by a single engineer, Jiashi Li.
On Day 2, they released DeepEP, the first OSS EP (Expert Parallelism) communication library for MoE training and inference. It provides high-throughput and low-latency all-to-all GPU kernels, which are also as known as MoE dispatch, combined with support for FP8 operations.
💭 Qwen too released QwQ-Max-Preview a combined normal-and-thinking model and the most amusing part was that their announcement blog was written by the model itself!
We can predict seeing the bonds between Perplexity and The Browser Company getting a little sour (or sweet, based on how they perceive it), as Perplexity more or less is coming after Dia, or maybe it’s a healthy competition, whatever happens, remember, consumers are the ones who win.
⚒️ Code Editor Updates
💻 Along with 3.7 Sonnet being readily available on both Cursor and GitHub Copilot on Day 0, we have some more interesting AI Code Editor updates.
🎨 Lovable now has a Visual Editsfeature which allows you to directly click any element in the app’s interface and edit it both via prompt and through a Figma-like position and styling fixes.
🏄♂️ Windsurf released its newest wave of updates, including support for MCP, a better tab interface with tab-to-jump to the next relevant line like Cursor, a Turbo mode to let the agent YOLO through command execution, support for images in chat, and support for more models.
⚡ Zed introduces Zeta, an open-source model fine-tuned over Qwen2.5-Coder-7B to support blazing-fast edits. Using Zeta, Zed brings Edit Prediction, a tool that predicts your next move, or more like the next edit and you just have to tab tab tab your way to fixed code. Not to mention, just like anything else Zed builds, this is also blazing fast, thanks to Rust 🦀
🔎 Phind went through a major update to release Phind 2, which now supports web search, code execution, images, diagrams, and other interactive and rich visual outputs along with text… and they also did a fresh paint job for a more aesthetic UI.
📦 And some more small miscellaneous updates
🌳 Google AI Studio has a new message-branching feature where you cancreate a new message branch when there is more than one thing you’d want to ask the model from the current context but at the same time want to keep the history clear.
📒 HuggingFace dropped a massive “The Ultra-Scale Playbook: Training LLMs on GPU Clusters” containing the culminated knowledge of scaling the training of LLMs to thousands of GPUs with theory, practical code, and insights from over 4k experiments!
👀 HuggingFace’s Dataset Pages now has a DataStudio tab where you can ask questions about any dataset and an AI agent would write and run it as an SQL query. Check it live on any dataset.
😵💫 Ian Fisher shared how he is using this new nightmare methodology of programming called “Append-only programming”. In this method, 1. everything goes in a single file, 2. new code is appended at the end and 3. old code can’t be edited.
from the linked post above, bro is all serious 👍
We almost had a panic attack, before he said out loud “It’s a fun challenge”, so yes, we left the farmland and back to keyboards. This really reminds me of memes talking about why men die doing adventures more than women.
🍪 The Cookie Store API, alternative to document.cookie, (nostalgic eh?) is now available on Firefox and Safari (partly).
Both of them haven’t shipped the full API yet and have their own set of concerns in implementing it fully like Chrome and Edge did ages ago. But anyway, this is a cool API and will help us consume cookies more gracefully, so we hope both Apple and Mozilla decide to move the needle enough to get green ticks on MDN.
🛎️ Jared and boys never stop cooking, Bun released v1.2.3 with bunch of improvement as always but the apple of eye being "You can now develop Frontend application with Bun".
What mean? Hold up, we'll explain Bun.serve() got built-in support for hot-reloading frontends. More importantly, you can now do bun *.html and that's all it takes to serve single or multiple HTML files with relevant CSS on a port. What a time to be alive and be a prototyper
🤝 You have read ~50% of Nibble, the following section brings some fun stuff and tools out from the wild.
⬇️ To convert a Nibble to a Podcast, we first have to clean it and convert it into Markdown. Lately, we’ve been using markdown-down to convert any webpage to a clean markdown with images downloaded.
Why are we posting age-old memes in low quality? Well, click here to see an interesting use case of SVG updating this meme for each weekday live.
Off-topic Reads/Watches 🧗
🪭 When to Give Up by DHH, where he talks about how it’s equally important to have the humility to give up as much as it is to have the grit to stick when things start going south.
😸 Repeat Happy Accidentsby Seth, asking us to allow ourselves to do things that might not work, to find things that work.
⚛️ Following the Majorana-1 Quantum chip last week, we got Satya on Dwarkesh’s Podcast talking with deep conviction and charisma about their quantum leap (aka the transistor moment) and what they envision of the world in the coming years.
Wisdom Bits 👀
“Farmers who wait for perfect weather never plant. If they watch every cloud, they never harvest.” — Ecclesiastes1 11:4
Nibbler P has lately been solving prod issues by the day and doing Wordle by night. He has been seeing an uptick in his VO2 max over the last few months and that makes him hopeful that better days are yet to come.
Nibbler Ahad an eventful weekend and sorted things that required his attention during his more than-busy week. His weekend flew by in playing badminton, and chess and meeting few friends.
If you have two minutes, please fill out a feedback form here. Your input will help us improve “The Nibble” for you and other readers!
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻
