#49

Humanoids, mem safe code, Salty Musk, Genie, 1-bit LLM, Image to Video, iGenAI, Ideogram 1.0, Dencun, 0xMarch, Leap Day, Apple-less, Tellurium, TS infer, coffeestains, GPU List, Centaurs and more

Hola & welcome to the 49th! 👋🏻
This week was as noisy as ever, an extra day last month means extra bullshit over the internet, don’t you worry, we read & filter it all.

📰 Read #49 on Substack for the best formatting.

🎧 Podcast version of this edition is available here → #49 | Recast

📢 Get access to bonus links and discussions with fellow Nibblers

---

📆 The IIITV Coding Club is organizing CodeHers 6.0 - their flagship competitive coding event for girls - on March 3rd (fun fact: both of us were Secretary of this club in our undergrad days) | Check event details on Hackerrank 👩‍💻

---

What’s happening 📰

• 💰 While we were all lost in the AI & Crypto bull market, Figure, a company figuring out how to make humanoid robots, raised $675M at $2.6B (ka-ching). The sac of money is filled by Microsoft, OpenAI, NVIDIA, Jeff Bezos, Intel Capital, and more. In addition, they also inked a deal with OpenAI to develop new models together.

• 💥 Tim Cook hints that Apple is about to do something groundbreaking in Generative AI, and started by killing the “Project Titan”, autonomous Car project (a decade old) and moving (or firing) all employees to GenAI teams.

• 🧑🏻‍⚖️ MemeLord Musk Chan has been threatening to sue OpenAI for a year now, and FINALLY, he did. He is now suing OpenAI officially for breaching the principles of the corporation and putting the profits and commercial interests ahead of the public good and also, he is pissed that OpenAI ain’t open. (Karpathy’s exit was well-timed, no?) [Filing Document]

  • Also, emails from June 24, 2015, sent by Sam to Elon show the same that they wanted to build a truly “Open”-AI.

  • and Peak Domino effect in action, this suing led to a dip of 5% in Worldcoin

  • Oh! And just like usual days 3 more publications sued Microsoft & OpenAI for stealing training data.

✨ AI Digest

• 🌠 Ideogram (you might remember from #23), released 1.0, calling it the most advanced text-to-image model offering SoTA text rendering, unprecedented photorealism, exceptional prompt adherence, and “Magic Prompt” to help with prompting. (it’s scarily better with text & we see that as us not using DALL-E anymore).

• 🪷 Ola launched the promised Beta of Krutrim Chat they claim that it’ll have fewer hallucinations in the Indian context than other models. And that has extensively trained in Indic Data & 10+ languages. But Ola launches have always faced controversies and this time, people filled Twitter with funny & inaccurate responses from Krutrim. The funniest is Krutrim saying “built by OpenAI”, which the company later clarified is due to the data poisoned already by OpenAI in some open-source dataset they use (maybe?!?).

• 🍃 Mistral launched their new multilingual model called the Mistral Large with a 32k context window out-of-the-box and native function-calling support, which makes it a substantial improvement over their previous models (both open and closed weights) and a close competitor to GPT-4. More from this release 👇🏻

  • Le Chat Mistral: A new conversational assistant platform serving Mistral Models where you can try Mistral Large, Mistral Small, or Mistral Next.

  • Mistral 🤝 Microsoft: Now who would have guessed THAT? Microsoft is essentially hedging its position on OpenAI with Mistral’s open AI. This collaboration joins hands between the two for AI infra, scale, and research. [check the meme section below for something related :p]

  • Though this model (and Mistral Small, Medium, and Next) are not open-weight models (yet?). Mistral still is committed to Open Source, as clarified by their CEO on 𝕏. This makes open-mistral-7B and open-mixtral-8x7b the open-weight models and mistral-small-2402, mistral-large-2402 and mistral-medium the closed-weight ones and mistral-next still being under research.

  

  Comparison on MMLU (measuring massive multitask language understanding)

• 🤏 Microsoft Research Asia's General AI Research team released a paper titled The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits where they discuss the extent of taking language models to use only 1.58 bits by restricting each weight to be one of three values among {-1, 0, 1} (log base 2 of 3 is 1.58 hence that bit value). They claim their work provides a Pareto solution to reduce the inference cost (latency, throughput, and energy) of LLMs while maintaining model performance using an “absmean” quantization function.

• 🎶 After achieving safe Image Generation through Firefly, Adobe is now also working on another dimension, "Music", yes!! Their new Project "Music GenAI Control" enables users to generate music from text prompts and edit them with precision using similar instructions.

• 🧞 Google released Genie, a foundation world model trained from Internet videos that can generate an endless variety of playable (action-controllable) worlds from synthetic images, photographs, and even sketches.
  What’s novel is is that it can learn fine-grained controls exclusively from Internet videos which lack any label data. This can be used to create new training environments for developing and testing AI agents, even for real-world tasks.

• 👄 Pika joined hands with ElevenLabs to create Lip Sync for its image-to-video feature and the demo is crazy good (currently in early access for Pro users only)

• 🌸 BigCode released StarCoder2 and The Stack v2 dataset with the StarCoder2-15B model being one of the best in class for coding and maths (though DeepSeekCoder-6.7B is still our favorite as it’s a close match while being more than 2x smaller). The Stack v2 dataset contains a whooping 67.5TB dataset of source code amounting to 900B+ tokens. You can even check if it contains your code in it and request for its removal if you wish to.

🔐 0x Digest

• 💥 Due to high trade volumes in this bull cycle, the Coinbase app crashed & some users saw $0 balances. (still surprised that their India team hasn’t used “Baby ko Base pasand hai” to market their Ethereum L2 chain called Base).

• 📓 MilkRoad’s Guide to Crypto Airdrops 2024 and How to Catch ‘em All.

• 🚀 Ethereum Dencun Network upgrade is scheduled for March 13, 2024, at 13:55 UTC (epoch 269568), and here’s everything you need to know.

• 📆 While February was quite a leap for Crypto in general and multiple DEXes¹ broke all the records, March will be even more happening and fun and to cover that folks from LayerGG made a calendar.

  

  Crypto Calendar - March 2024 by LayerGG

• 💸 BlackRock’s Bitcoin ETF launched 7 weeks (~50 days) back, has reached a milestone of reaching $10B in assets under management. For context, only about 150 out of 3,400 ETFs have achieved these and most of them are more than a decade old.

🛠️ Dev & Design Digest

• 🇺🇸 The Whitehouse advised to use of memory-safe programming languages over non-safe ones, they shared a guide “Back to Building the Blocks”. (*sad JS noises*).

• ⌚️It’s 2024, and Date/Time is still a hard problem in computer science. And Leap years specifically “leap days” make it more tricky.

  Gergely Orosz
  wrote an article “Happy Leap Day!” covering all the software disasters that happened on 29th Feb. USE WELL-TESTED LIBRARIES FOR TIME, DON’T REINVENT IT.

---

What brings us to awe 😳

• ⚡️ SSDs Have Become Ridiculously Fast, Except in the Cloud. Over the last few years, we have come a long way and have seen a huge increase in throughput and capacity in SSDs. We can achieve 13GBps in the State-of-the-art SSDs right now locally. But cloud vendors like Amazon still have EC2 disks capped at 2GBps, maybe because it might be costly or no one needs it.

• 🧾 Netlify just sent me a $104K bill for an S3 simple static site, a user learned the hard way not to host their static sites on Netlify or Vercel, as they don’t have DDoS protection. The person hosted a 3MB sound file & someone DDoSed and consumed 60.7TB in a single day. (telling you to use GitHub as CDN, and be free.)

• 🪟 In his quest to move away from macOS (bro, moved to Samsung S24+, Windows, and VSCode + WSL last week), DHH wrote another post on "Fonts don't have to look awful on Windows", we like that he thinks the same as us on how typography matters more than it does; he also points out that most of macOS’s superiority also comes from high-quality retina display and if you connect your Windows machine with same kinda display, they aren't bad. It is a controversial opinion but we think Apple is just consistently good at packaging general stuff (both software and hardware) into a sleek machine better than its competitors.
  [UPDATE: Apple clarified that they won’t be removing home-screen PWA, so DHH is relaxed and liked his exploration, but won’t be moving away from Mac]

---

Today I(we) Learnt 📑

• 🗿 Tellurium (52) is an interesting & rare (as much as Platinum) element.
  In an experiment in Tokyo, researchers found that when Tellurium crystal is shined with a femtosecond laser, it acts as a “photovoltaic cell” 🔋
  This is exciting as this can be the future of cheaper energy. But the element is notorious and comes with its own set of caveats like “garlic breath to people who hold it with raw hands”. (Sorry Jains in the House)
  Even before this experiment, we all have interacted with Te in one way or another. One of the layers in the re-writable CD/DVDs (nostalgic noises) is made of Te. It was already being used to make cheaper thin-film solar panels.
  Oh! yes, and it ackchyually makes an alloy with Copper to form "YOU", iykyk.

• ✨ The infer keyword makes TypeScript really(?!?) powerful, allowing conditional types and inferring from nested template types. [More in infer]

  

  Source: TS Docs

---

🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.

---

What we have been trying 🔖

• 🖼️ TextArt: A huge collection of ASCII art. Just go there and then copy-pasta.

• ⚙️ GPU List: a place to buy, and sell spare GPU capacity.

• 🎤 Neets.AI: Low-cost AI for high-quality Text-To-Speech (TTS) voice generation, featuring unfiltered LLM, unlimited API access, and premium GPT chat.

---

Builders’ Nest 🛠️

• 🔌 tabgod: execute any javascript on any chromium tabs - cross-tab parallel execution.

• 📦 corepack: Zero-runtime-dependency package acting as a bridge between Node and package managers by the NodeJS team. (install this to not install the installers later, okay?)

• 🏗️ hyperdiv: Build reactive web UIs in Python. This is yet another attempt to write HTML using code.

• ☕️ coffeestains: Add Coffee Stains to your LaTeX documents :p

  

  Legit cleaned my screen twice

---

Meme of the week 😌

Source: @dejavucoder on 𝕏

---

Off-topic reads/watches 🧗

• 🤷🏻‍♂️ You couldn't know by

  Jason Fried
  on “Why you should cut yourself some slack” (neither of us does, but you should).

• 🧹 Stop postponing things by embracing the mess and do not wait for the conditions to be “just perfect” to start things off. Or as Nike says, Just do it. (this is Nibbler P’s mantra & Discord Status btw!)

• 💰 “Why is Centaurs² a better metric than Unicorn to gauge the success of a startup?” by

  Ashu Garg

  $100M in ARR is the best measure of startup success. Here's how founders can get there.

---

Wisdom Bits 👀

“When a person can’t find a deep sense of meaning, they distract themselves with pleasure.”
― Viktor Frankl (notoriously known for attacking people like us 🥹)

---

Wallpaper of the week 🌁

🌌 Link to Wallpaper → wow.nibbles.dev

---

Weekly Standup 🫠

• Nibbler P has been actively trying to close his activity rings the entire week (the first time in several weeks) and is happy with his progress. He also met a baby scorpion in his house and has been taking care of it ever since.

• Nibbler A had another busy but interesting (few airdrops 😛) week, working on some fun stuff, and learning new things every day (about life & computers). He played tons of Chess this week (yay!) and is trying to fold up some pending (just 3 weeks) chores. Trying to get level with "Solo Levelling" episodes.

If you liked what you just read, recommend us to a friend who’d love this too 👇🏻

1. DEX - Decentralized Exchanges, now guess what Centralized Exchanges are called 😛

2. Startups with $100M ARR