#36
AI chips,DeepMind Gems, debut of real-time LLM API, StackOverflow's AI Search, AWS's AI updates, S3 gets faster, sdxl turbo, DrawThingsAI, un-Stability.ai?, Aitana Lopez, git notes, llamafile and more
ππ» Hola & here comes the 36α΅Κ° Nibble, we had so much content that half of #37 is also pre-written now. π€·π»ββοΈ
π§ Podcast version of this edition is available here β #36 | Recast
π’ Get access to bonus links and discuss it with fellow Nibblers ππ»
π If you are doing Advent of Code 2023, we have a discord channel for discussions.
Whatβs happening π°
π Starting from OpenAI, Sam is back, and the board now has some new members and removed some old (the one who said yes to firing) ones. The Board will include a non-voting observer for Microsoft. Also, ChatGPT turned β1 year oldβ last weekend.
π Google DeepMind has done it again, just like they did with protein structure earlier, now they have found 2.2 million new crystals, including 380,000 stable materials that could power future technologies. It is said to be equivalent to nearly 800 yearsβ worth of knowledge.
π Stackoverflow AI Search is now rolling out slowly to everyone.
It searches among the existing answers and summarizes them, giving you a link to the best answer too. You can ask follow-up questions with it too.π Updates from Amazon's re:Invent
S3 Express One Zone, 10x better perf than regular S3
AWS is into designing chips named Graviton4 and Trainium2 that will make running generative AI easier.
Amazon Q, their AI assistant now pops on your face in the AWS console. Also, it has βsevereβ hallucinations and leaks (tell me a better name for an AI assistant right now, Iβll wait).
ποΈ Stability AI released SDXL Turbo, which achieves SoTA performance with a new distillation technology, enabling single-step image generation with unprecedented quality. You can try the blazing-fast demo at Clipdrop (We have been trying it out through the Draw Things app, which generates good images in ~26 seconds at 4 steps on an 8GB M1).
Also, they seem to be getting unstable (pun intended) on the other end, as there are reports saying investors are asking the CEO to resign and possibly explore a sale over financial/management issues (the cost of OSS is real).π€― Pika, a new image-to-video tool, just launched and it is on par with runway, if not better. Even though most of these tools are limited to around 4-second videos, we love the pace at which image-to-video is pacing forward. Itβs only about to get better from here. Below is what it created for βa boy walking in a forestβ.
π§ Remember Elon Musk's AI chip in the brain startup? Yes, Neuralink, well they raised an additional $43M. Despite their FDA rejection last year, they have continued trials under the exemption. (bro wants thoughts to image/video, prompts are mid)
π€ Perplexity launched two new PPLX models:
pplx-7b-online
andpplx-70b-online
.
These online models are focused on keeping the responses fresh, factually correct, and not hallucinated. You can go try them out at β Perplexity Labs.
They are built on top ofmistral-7b
andllama-70b
and a really good example of RAG in production.π Keras v3.0 is released which is a complete library rewrite (i.e. breaking changes) that JAX, TensorFlow, or PyTorch new large-scale model training and deployment capabilities. (btw, the Keras team is hiring too).
What brings us to awe π³
π€‘ Being an Instagram Influencer is sure a way to become super famous and filthy rich, but itβs hard, right?
Well, Meet Aitana Lopez, a 25-year-old Spanish model created using AI. And they have 141k+ followers, making $10k/month for creators. (βΌοΈ young men out there stop paying money to anyone selling you a few pics, they might be my AI-generated models)β‘οΈ Biome claimed the $25,005 bounty for rebuilding Prettier in Rust. A blazingly fast formatter awaits you all. Also, Biome formatter has now over 96% in terms of compatibility against Prettier!
βοΈ The Q* hypothesis is a nice writeup explaining the Q* algorithm (that created hypes after a leak that OAI built something groundbreaking using it). Though we can only speculate exactly what OAI did until they reveal it, itβs good to be grounded in reality before terming anything as AGI so soon.
πΈ iPhone Camera Computational Photography in action - If the subject is moving when the shutter is pressed, many differing images are captured in that instant. Apple's algorithm stitches the photos together, choosing the best versions for saturation, contrast, detail, and lack of blur.
Today I(we) Learnt π
π git notes
: coolest & most unloved feature of git. You can add notes to any git object (i.e. commit, blob, tree, etc). GitHub one day decided to stop showing git notes in the UI, which by the way you can still see in CLI. This led to notes becoming a hidden feature ingit
ποΈ WhatsApp has a way to βProtect IP address in callβ. This works funnily, instead of P2P calls, your calls relay through WhatsApp. So, itβs a trade-off TBH. (either reveal your IP or let your call get zucced)
π€ You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying π
π QRious: QR Codes that represent your brand and business
π Microsoftβs AI for Beginners: Microsoft's 12-week, 24-lesson curriculum, including basics and all for free.
π₯ LiarLiarAI: Harness the Power of AI to Detect Lies and Heart Rate Fluctuations.
π¨ DrawThingsAI: Run Stable Diffusion locally on iPhone, iPad or Mac (we really really really really love this).
Buildersβ Nest π οΈ
π§π»ββοΈ avatarion: Avatartion is the perfect tool to generate a personalized Notion-style avatar. (It is rare to see a limited option generator work so well)
βοΈ Cosmic Latte: We (in-house) made a Chrome Theme around the color βCosmic Latteβ, itβs the color of the universe. It goes really well, with the Nibble website.
π Understanding Training Loss Patterns - Helps you recognize various training-loss patterns. We can then use techniques to overcome them and finish training (in fact, the entire ML engineering repo is a gem).
πͺ llamafile - Distribute and run LLMs with a single file across different OS without any installation (read the announcement here)
Meme of the week π
Off-topic reads/watches π§
β³ Live with It for a While by Jason Fried, he explains when in doubt about product features, let them be, live them, and let time unfold that is useful or confusing.
π£ The Bitter Lesson by Rich Sutton (shared by
) is that AI's future lies in 1) embracing Mooreβs Law and acknowledging that heavier computation will get easier and cheaper and 2) discovering new knowledge rather than incorporating existing human understanding.π§Ή It's probably time to stop recommending Clean Code a good read on why and how Clean Code guidelines are not context-independent.
Wisdom Bits π
βThe bad news is time flies. The good news is youβre the pilot.β
β Michael Altshuler
If you liked what you just read, recommend us to a friend whoβd love this too ππ»
Weekly Standup π«‘
Nibbler A had the usual full week in BLR. Played some πΈ & lost more games in βοΈ. He is trying to be more active on Nibble HQ Discord this weekend. Also, was trying SDXL locally & created an account on civit.ai to find configs, got scarred for life.
Nibbler P was perpetually on-call towards the end of the week and he feels so very exhausted and impostor-y. He compensated for this feeling by having a little fun solving AoC and AoT and experimenting with Stable Diffusion.