#47
Sora, Gemini 1.5, Navarna, V-JEPA, LLRT, YC's RFS, Citeria, iWork, ccTLDs, SP1, EIP-7623, Pinokio, hurl, freenginx, Fix You, Zoxide and more.
Welcome to the 47th!, Silver (Ag) edition 😛
📰 Read #47 on Substack for the best formatting.
🎧 Podcast version of this edition is available here → #47 | Recast
📢 Get access to bonus links and discussions with fellow Nibblers
📆 Feb 21 - Deadline for YC Summer 2024 batch. Also, YC has their latest Request for Startups (RFS).
What’s happening 📰
😌 Amazon’s LLRT (Low Latency Runtime) is an experimental, lightweight JavaScript runtime designed for serverless envs (read Lambdas). The world is bullish on making JS go brrrr…
🥽 Zuck went all out and did a comparison of Meta and Vision Pro all by himself. Ofc later he concluded, “Quest 3 is the better product”.
✨ AI Digest
䷉ Google released Gemini 1.5, their new multimodal model built using the MoE architecture that they claim to perform on par with the Ultra 1.0 model while being a smaller model. It comes with a context length of 128k right now and it would gradually expand to a whooping 1M token context length! What’s even more impressive is that they conducted experiments with up to 10M context length (~7M words) with surprisingly good accuracy. The API is available now on their AI Studio and Vertex AI.
The context length of this capacity is insane! Though we are yet to see it on real-world data (if they ever release the 10M context length someday) 📹 But Sama was not sitting still and soon after Google’s announcements, OpenAI announced Sora, their text-to-image model that can create video up to a minute long in length, and boy it absolutely nails the prompt adherence as well as the temporal and spatial accuracies (numerous startups died right here). Sam even took prompt requests from users on 𝕏 and generated videos on them showing how it can understand even complex prompts.
We were seeing the hallucinating video of Will Smith eating Spaghetti just 10 months back and now we are here!!!!!!
The OpenAI team is currently doing safety tests on Sora before releasing it.🪜 Stability AI released Stable Cascade, their new text-to-image generation model which they claim to have both better prompt alignment and aesthetic quality than their previous models (now that we have reached quite a good level of realistic level in t2i generations, we are seeing more folks focusing on adherence and alignment these days). It is based on a new Würstchen architecture which uses two-stage compression using two diffusion models achieving a much larger compression ratio while maintaining quality, improving text in image quality, and improving adaptability. As with the previous diffusion models, we are curious to see what the community does with this because that's where we see the most impressive fine tunes and use cases coming out.1
the images are getting riyal, guys 🪄 Magic is working to create a perfect AI-powered software engineer who can be your co-worker and not just a co-pilot. And they raised $145M from Nat & Alphabet. (also, they boast about having thousands of GPUs)
👶🏻 Meta’s V-JEPA (Video Joint Embedding Predictive Architecture), is a non-generative model that learns by predicting missing or masked parts of a video in an abstract representation space. Yann LeCun wants AI models to learn like how a baby does. He thinks this approach will make it easier for models to build a “world model” with less data. Zuck even has a demo video of him playing guitar and V-JEPA filling in the gaps.
📝 ChatGPT is rolling out a feature to remember things you discuss with it (including when you were rude to it 😈) to make your chats more helpful. Currently, it is limited to a few users.
🌱 Cohere introduced Aya, a state-of-the-art model and dataset, pushing multilingual AI for 101 languages.
🪷 Navarna v0.1 is a novel SFT + DPO finetuned by TokenBender on @NousResearch's OpenHermes2.5 (Mistral v0.1). It is built to be good in Hindi/English chat with sentence retrieval (RAG) tasks capability inbuilt in Hindi.
🤝 Slack AI, a paid add-on for Enterprise subscriptions, is essentially a set of features that uses AI to help users find information more quickly and easily. Some AI-powered apps that you can use with Slack are Notion (view doc summaries with links), Perplexity (subscribe to hot topics), and more.
🏷️ Meta will start tagging images as AI-generated (from other companies too) on FB, Instagram, and threads (No way! my Insta model startup ain’t gonna scale).
✨ Apple released a paper “Keyframer: Empowering Animation Design using LLMs”. Uh! and they also bought the iWork.ai domain, maybe this indicates an incoming AI-powered workspace by Apple?
🔐 0x Digest
🪙 Polish Town Mińsk Mazowiecki launches own StableCoin “MinsCoin”. Few businesses will start accepting payments using the coin once launched, while this is an experiment, others might follow later. Also, people will be able to earn tokens for participating in community initiatives later.
🤡 Uniswap’s creator Hayden pointed out a UX scam coming along, based on the dropdown for addresses showing suggestions to pick ENS starting with the characters/address.
Source: Twitter 💰EIP-7623 is proposed to “Increase call data cost to decrease the maximum block size”. With an increasing number of rollups posting data on the Ethereum chain, the average call data size has increased. This move will in a way force people to move to blob for DA. The goal is to lower the maximum block size to ~0.55 MB.
🔬 Succint Labs did a Valentine’s release and announced SP1, A performant, OSS, zkVM that verifies the execution of arbitrary Rust (or any LLVM-compiled language) programs. As per their docs, it also beats RISC-0 benchmarks.
🛠️ Dev & Design Digest
🤖 Android 15 Developer Preview was announced by Google.
🕊️ Maxim Dounin, one of the core contributors at nginx (read “Engine X”), left it last week announcing freenginx.org. Why you ask? → Well, F5 acquired NGINX for $670M back in 2019. And just like any other acquired OSS project, the non-technical folks wanted to market nginx at the cost of some compromises on how the project used to be maintained, which led to Maxim leaving F5’s side and starting an OSS fork (again!!).
What brings us to awe 😳
📊 The Zerodha team wrote about how they process (generate, sign & email) 1.5+ million PDFs in 25 minutes, and more importantly how this process used to take 7-8 hours before they redesigned their batch job handling this task.
Major takeaways are:🧳 Big Tech jobs have lost their glamour, before the 2022 economic slowdown, FAANG was the place to be if you were in tech, you got hefty pay, lavish perks, and an ideal work-life balance. But in the last 2 years, with layoffs, a major decrease in pay, and overall all nullification of energy in BigTech. The FAANG and big tech, in general, are not as pleasing any more, the grindset and FAFO folks are building their shit and some even want to exit the industry.
Today I(we) Learnt 📑
≄ There are two kinds of TLDs (Top Level Domains) → gTLDs (generic, managed by ICANN) & ccTLDs (country-code, managed by countries, really risky bets).
You see the problem is that ccTLDs are very risky and you might have seen tweets about it. Also, now it makes sense why OpenAI took openai.com and not open.ai.
And yeah let’s not talk about ENS for some time, it ain’t real. 🙂
[Source: Not all TLDs are Created Equal]🤯 badmephisto is karpathy: Do you know how to solve a Rubik’s Cube? If you do and have learned it online, chances are that you’d be familiar with the badmephisto’s website or his YouTube tutorials. Turns out he is none other than Andrej Karpathy!
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying 🔖
↪ As weird as it may sound, we are trying a CLI tool called Zoxide, as an alternative to
cdcommand. Yes! And it’s worth it once youzenough (it trains on your cd).
If you need some convincing of this absurd idea, you can watch this explaining which command is picked when conflicts arise & how frecency is used.
an image showing how z could find the correct folder to cd into given two elements from its path, in order.
Builders’ Nest 🛠️
🚫 noTunes: A simple macOS application that will prevent iTunes or Apple Music from launching (ofc you need an app to prevent that)
🧪 hurl: run and test HTTP requests with plain text.
🤥 Pinokio: Install, Run & Control AI apps on Your Computer easily (including the newly released Stable Cascade)
Meme of the week 😌
Off-topic reads/watches 🧗
🌏 Coldplay's “Fix You” for Mother Earth: Coldplay is on its Music Of The Spheres Tour and in just the first 12 months, they produced 47% less CO2 than their last tour. From using plant-based wristbands to kinetic dance floors that create energy for the concert to planting a tree for every ticket sold (making it 5M in total), we wish more artists learn from this and do something just like this 🎶.
📖 On Shortification of “Learning” and Teach Yourself Programming in Ten Years: Two of the greatest AI scientists of our generation have spoken out on why you should not go for quick “snacks” when learning something out and instead embrace the hard work and sweating sessions.
Wisdom Bits 👀
“In a free market for attention, someone is always racing to the bottom.”
― Seth Godin
Wallpaper of the week 🌁
🌌 Link to Wallpaper → wow.nibbles.dev
Weekly Standup 🫠
Nibbler P continued his usual One Piece readings and stretched some more muscles, resulting in more muscle pain (smh). His day is now usually split between learning new things at work and clearing up a backlog of 700+ unread mails.
Nibbler A is a madman who finds new ways to make his week busier every time. Apart from that he spent this week without
muchany physical activity. Had a fun time setting up new gadgets and reading a lot of shit from the internet.
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻
BTW, the Stability folks figured out they weren't making any money and hence slapped a non-commercial license on this and the only way to use this commercially now is to pay them so I guess all the finetunes would also carry this license hmph.
Amazing knowledge doses sers, Keep 'em coming!
awooooo