#98 — Microsoft's casual Quantum Leap, NodeJS to settle on ESM-only, reserve best hours for dreams and more
Majorana posts gets love for Teams, Zuck bringing cables, Musk groks GPU, Alibaba wanX vids, Crossmint brings Smart Wallets to Solana, Coinbase eyes for India re-entry, create-react-app retires & more
👋🏻 Welcome to the 98th! Apologies for the delays so big, it seems like a skipped week to reset timelines. … and, we also wanted to give you more time to read the last one 👉👈
To make sure we ship the 100th on time, we’ll be pushing a small #99 in the coming Monday!
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. 💪
🤯 Microsoft opens the door to the era of Quantum Computing by introducing the first-ever quantum chipMajorana 1 and built it from “scratch” in a way.
Satya unveiled in his tweet that they have been working for the last 20 years, and have finally created a “kinda” new “state of matter” (yes, like solid, liquid, gas, BEC) unlocked by this new kind of material “topoconductors”, which can create, control, and measure Majorana Fermions (exotic quantum particles that are its own anti-particle).
In essence, this moves Quantum computing decades ahead overnight. Until now Quantum computing was more about tons of error correction because it’s not easy to create or track qubits, but with Majorana 1’s Topological Core architecture, it becomes relatively easier and far more accurate. These little new chips once worked upon more, can actually compute-intensive problems worth millions of our regular bit-based computers
The amount of love Teams get on the internet is still more than Majorana. Humans are weird indeed.
Twitter always reacts to Microsoft officials building cool things.
🌊 It’s official — Zuck goes underwater, with Project Waterworth, building the world’s longest (~50,000 km) subsea cable project with the latest technology.
Men will literally wrap a long cable around the Earth instead of going to therapy.
This is going to cost multi-billion dollars and going to take a few years if not decades to complete, backed by Meta’s aim to bring AI and its capabilities to the masses. (And btw, BSNL is still planning the 5G upgrade)
Even though Apple released this new device, the kind of other stuff that has been happening in tech these days easily overshadows all of it.
This phone has everything one needs to be an Apple user—a mid-range camera, a 60Hz screen, some limited Apple Intelligence, a powerful A18 chip, and a hefty price tag of $599. It will be available for pre-order on February 21st. Although this isn’t cheap, Apple is now competing with flagship Android Devices in terms of price with this cheap thrill. Ah! Also, Apple might have casually killed the Event applications market by introducing its own Event apps.
✨ AGI Digest
🤖 Crazy week of new models and newer agents
📈 Starting this week on a high note, Elon’s XAI released Grok 3, a family of two models — a normal Grok 3 and a smaller Grok 3 mini, both with reasoning capabilities much like DeepSeek R1 and OpenAI’s o1/o3. While it is still in beta (i.e. still in post-training) it is open to use for X Premium and Premium+ users via x.com and grok.com and on their respective mobile apps.
While API access is still a few weeks away, the initial vibe check about the model is very positive, especially for programming and technical understanding! It holds the #1 overall position on LMSys and is the first model to get a 1400+ Elo rating (without style control, though; any way you should evaluate it on your tasks).
Also, following the host of “Deep Research” agents rolling out from every other major lab, xAI followed suit and added a DeepSearch button along with a Think button on the web and app interfaces to have more control over its search area and output length depending on your task. Plus, an unfiltered voice mode may be rolling out soon as well. Overall it’s a well-received model and xAI’s shipping speed is commendable.
🕵️♂️ Perplexity’s been on a small shipping spree as well, quickly launching its own version of “Deep Research” called, well Deep Research. Free users get 5 queries per day while paid users get 500. They also released their fine-tune of DeepSeek R1 on HF which has been post-trained to remove the Chinese censorship while maintainign it smaths and Code performance.
While it gets high scores on some benchmarks, in both our own testing and from what other users posted, it is just not quite there compared to the other similar offerings. @hrishioa has a good tweet on it with detailed comparisons with Gemini’s and OpenAI’s. Grok was not released when this was written but we’d keep it somewhere between Gemini and OAI’s in terms of quality.
🔎 Guess who else got “Deep Research”? Our frienndly neighbourhood embedding and reranking provider — Jina AI with its DeepSearch.
This one’s actually available via API as well as on search.jina.ai but compared to the other offerings mentioned above, we’d say you’re better off with the Perplexity’s free Deep Research over this.
👩🔬 While others are busy building a deeper search agent, Google actually created an AI co-scientist. AI Scientist is a multi-agent AI system built with Gemini 2.0 as a virtual scientific collaborator to help scientists generate novel hypotheses and research proposals. Think of it like a co-pilot but for scientists and researchers.
In its validation test on real-world laboratory tasks, the AI Scientist:
Proposed novel repurposing candidates for acute myeloid leukemia (AML). Subsequent experiments validated these proposals, confirming that the suggested drugs inhibit tumor viability at clinically relevant concentrations in multiple AML cell lines.
Identified epigenetic targets grounded in preclinical evidence with significant anti-fibrotic activity in human hepatic organoids.
Independently proposed that cf-PICIs interact with diverse phage tails to expand their host range (a novel discovery that had not been done before but not yet revealed publicly)
And these are just the validation tests. The future is going to be wild!
Source: AI co-scientist blog post
⚙️ Talking of an AI Scientist, we cannot forget SakanaAI who came up with a primal AI Scientist back in August (we covered it in #73). They are back with an AI CUDA Engineer which can produce highly optimized CUDA kernels, reaching over 10x speedup over common machine learning operations in PyTorch.
While they initially claimed over 100x improvements in some kernels, it was quickly identified as a bug in the evaluation script which benchmarked it higher when it was actually 3x slower (here’s Lucas Bayer of OpenAI taking a jibe at it using o3-mini-high). The leaderboard has been fixed to remove these outliers and the team is working on fixing it. Even without the high 100x speedups, the 10x improvements are already much better than writing them by hand (given that you have good evals to test them)
The AI Cuda Engineer Agentic Framework | Source: Sakana AI CUDA Engineer Blog
The model is based on the StripedHyena 2 architecture and is trained on 9.3T nucleotides—the building blocks that makeup DNA or RNA—at 1M token context length and they released two models — a 7B and a 40B, complete with their weights, training, and inference code and pretraining data. Evo 2 can predict the pathogenic effects of human genome variants across coding and noncoding mutations. In other words, if you have a genetic mutation, Evo 2 has an opinion on whether or not it might cause disease. Evo 2 can also be used for biological design, like for the generation of entire human mitochondrial genomes with coherent synteny and even whole bacterial genomes and eukaryotic chromosomes. This is next-level stuff and it’s only going to expedite the medical research going forward.
🤏 Plus some smaller model news that went under sight
💭 Nous released DeepHermes 3 Preview which unifies both reasoning and normal responses in a single model, controlled by a system prompt. It is based on the Hermes 3 model and the thinking mode improves over the non-reasoning by using more compute during inference to think harder.
👀 Microsoft released V2 of OmniParser, their screen parsing tool, which interprets/converts UI screenshots to a structured format, to improve existing LLM-based UI agents. Over its last version, V2 gets a 60% improvement in latency and a strong score of 39.6 average accuracy on ScreenSpot Pro.
📷 Google followed their last PaliGemma 2 with a new PaliGemma 2 mix modelwhich is capable of performing multiple tasks such as short and long captioning, optical character recognition (OCR), image question answering, object detection, and segmentation. The model comes in sizes of 3B, 10B, and 28B for 224px and 448px resolutions.
🎥 Talking about big companies, Alibaba introduced a new video & image generation model, WanX 2.1. The demos of WanX (yes the name is indeed a little funny) look really promising and they are saying that 2.1 will be open source. China is definitely getting the hang of "how to please developers", if you think carefully about it they have done with every industry in the past too, build fast, make it cheap, and make others look like bad options.
🗣️ Google ain't stopping shipping and Gemini can now leverage all your past conversations with it for crafting new responses. The features are right now only available on Gemini Advanced in English on Gemini 2.0 Flash. Also, all the chats you want to keep are in your control, you can remove or review them in the ”My Activity” Tab of your Google account.
🔐 0x Digest
👛 Crossmint launched Solana Smart Wallets, which are programmable gasless wallets. This is definitely not something new in crypto land, but maybe for Solana tribe.
For unamused Ethereum maxis reading this, yes it’s more or less ERC-4337, but with a state structure that somehow works faster than our Merkle ops.
This is a huge change from the last 2-2.5 years when there was a chaotic relationship between crypto firms and Govt entities that led to many disasters.
So, in case you are a diamond hand 💎🙌 type of person who also seems to receive loads of airdrops and community rewards, it’s time to take a round trip to Hong Kong for “investments”.
After disabling support for UPI in April 2022 (like literally a few days after enabling it), this is the first time Coinbase is openly talking about its interest in the विश्वगुरु.
🛠️ Dev & Design Digest
🪦 React team sunsets the good old “Create React App”, in favor of using frameworks, or using React on your own with custom build tools like Vite, Parcel, or RSBuild.
CRA got many of us started with React, as it was easy to ship applications. But if you have worked on really complex, performance-intensive applications, you might have seen the limitations of the bootstrapped structure of CRA at times. People usually eject or do craco to extend the config without losing everything. With this move, the React team has given us the freedom to use React from scratch as it’s not that hard and has more stable APIs. ‼️ Using this as an opportunity Tanstack team shipped create-tsrouter-app, a drop-in replacement for CRA with Vite and Tanstack Router.
This new plugin leverages the CSSTree AST format to provide linting rules for CSS. With JSON, Markdown, and now CSS, ESLinters are on a roll here. After all, hey the linter business is getting tougher with Rust and Zig guys out in the open.
✨ After Node 22.12.0 (Jod) was shipped with require(esm) enabled by default in December, the grass felt more greener to community and keep that in mind, Joyee Cheung dropped a banger PR to backport require(esm) to Node.js 20.
This is a big change than it looks, backporting it to v20 implies that as soon as Node 18 hits EoL, we can all settle on ESM vs CJS debate and ship ESM-only packages in peace. (no more comparing this debate with spaces vs tabs or vim vs emacs)
🧮 How hard can it get to implement a Calculator app? Going by the works of Hans-J. Boehm who worked on implementing the Android Calculator, was pretty frustratingly hard, as described in this extremely detailed Twitter Thread by @ChadNauseam. If you already hate the iOS Calculator app (hey! mate), you have some more reason to love the Android one.
🧐 Cleo, the infamous Math Stack Exchange account that answered the most difficult integrals with no explanation has finally been doxxed. Turns out, he was a man named Vladimir Reshetnikov who was coming up with various hard integrals to solve and wanted to verify if his solution was correct. So, he used multiple anonymous accounts to ask questions and used Cleo to drop his answer without any explanation so that people would get baited to pay attention and provide an explanation. It’s xkcd #386 all over 😂
⚡ More of TIL: MongoDB has the ESR (Equality, Sort, Range) Rule to make compound indexes more efficient. The idea is that when creating a compound index, the order of keys matters. You should order them so that keys on which you’ll perform equality come before those you’ll sort, and they come before those you’ll check ranges for.
Builders’ Nest 🛠️
📨 postal-mime: an email parsing library that runs in browser environments and serverless functions.
👨💻 Grep by Vercel: Fast code search across 1M public GitHub repositories.
💬 giscus: A comments widget built on GitHub Discussions.
🔎 Complexity: A Chrome plugin that adds advanced model control, code artifacts, and a much larger personalization support to Perplexity
🤲 Ensemble Stars by Seth Godin, on how most organizations don’t celebrate the Ensemble Stars but rather wait for a linchpin to arrive and fix things, and why that’s bad.
💼 Code yourself out of the jobbySwizec, a reminder on how to loosen your grip, let go of your legos, and trust people to get more time to work on yourself and not have the context as your moat.
👑 We Live Like Royalty and Don’t Know It by Charles C. Mann highlights the hidden mechanisms unbeknown to us, that make our lives far better than those of even the royals who lived as close as 200 years ago. I mean, you don’t have to think twice before drinking a glass of water because it might contain a bacteria that might kill you, or worry about a small mosquito biting your kid playing in the park. No amount of riches could have afforded this earlier. (but yes let’s not forget we are at the mercy of certain someone in the Northern part of Korea, who likes to press buttons)
Wisdom Bits 👀
“The best way to complain is to make things better.” — James Murphy
Nibbler Ahad a super busy week mostly at work, the badminton streak is not streaking anymore, and he probably requires a few more weekends to reset timelines. If he ain’t replying to you this week, forgive him.
Nibbler P is back on his workout regime after a little break, wishing his commits and reading sessions were a little more forgiving to him.
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 👇🏻
