#37
Gemini AI & false marketing, Animate Anyone, shopp.ing, MoE 8x7B, Compact Folders, GTA-6, Apple into ML(X), Meta AI Alliance
๐๐ป Welcome to 37แตสฐ Nibble
๐ง Podcast version of this edition is available here โ #37 | Recast
๐ข Discord refreshed their Mobile App UI, added new icons, and dark mode for OLED and more. would be a good time to join our server ๐๐ป
๐ If you are doing Advent of Code 2023, we have a discord channel for discussions.
Whatโs happening ๐ฐ
๐งถ OpenAI committed to buying $51M worth of chips from Rain.ai (an AI hardware company Sam personally invested in). Also, they delayed the GPT Store launch to early next year. (why you ask? the drama took up the time)
๐ฉฐ Alibaba researchers created a new AI tool for turning static images into dancing videos. It takes in a reference image + a pose sequence and generates a video of the reference image in the given pose using a Diffusion Model. โAnimate Anyoneโ was evaluated on a dataset of clips uploaded by TikTok stars, leading several commentators to predict it would lead to the end of social media influencers.
๐ชฉ Whatโs better than one image-to-animate mode? Two image-to-animate models!! Researchers from NUS and Bytedance (parent of TikTok) released another diffusion-based human image animation called MagicAnimate. This one too takes in a reference image and a DensePose motion sequence to convert the reference image to the pose sequence. We are living in amazing times! (It baffles us to see that both these similar papers were released in the same week. Well, in fact a lot of the release timelines are shortened because of a competitor launch. This is why so many of them are launched with 0 code in their โOSS repoโ.)
๐ฏ Ah! Letโs address the
elephanttwins in the room, yes we are talking about Googleโs new, largest, and most { capable, controversial } multi-modal AI โ Gemini.Plans to release 3 models: Ultra (~GPT-4), Pro (GPT-3.5), and Nano (Android and small devices)
Quickly came in a bad light after demos as folks found out that the demos werenโt shot in real-time or were fake.
Also, the benchmarks were a little shady as they compared CoT (chain of thought, as in prompting and following up multiple times) on Gemini with FSL (few shot learning) on GPT-4, to show they are superior.
No doubt it did beat GPT-4 in some benchmarks.
And, uh! you know how DeepMind had AlphaCode which could do competitive programming, well, now they used a specific version of Gemini, to build AlphaCode 2 and itโs said to crush the Codeforces rounds. (Iโm not as sad as I couldโve been, as I was never good at CP)
Just added this image, coz we love Google Product Sans when used in logos
๐ฎ๐ณ Bengaluru-based AI startup Sarvam.ai raised $41M, and they are building full-stack for AI. Working on Models, Platforms, and Ecosystems. (from their site it seems like they go to the same WeWork as Nibbler A?). Founders are a real start here, they have worked in the past on AI4 Bharat, Aadhar, GST & UPI.
๐๏ธ Hereโs some real-estate investment tip, Google said โIntroducโฆing the new
.ingTLDโ. Get yours at get.ing. FYI, we are already too late or too broke to buy exciting domains. But anyway have a nice shopping day.๐ค Tech Twitter went all โMoE MoEโ after an AI company led by French e/accs with clipart PFPs just dropped another Magnet URL.
Yes, we are talking about Mistral AI, they released a new open-source language model called Mixtral 8x7B 32k, which is a Mixture of Experts model consisting of 8 experts, each with 7B parameters of their own and 55B attention parameters.โจ The benchmark scores match the performance of llama70B & compute the cost of just a 12B model
๐ฐ Also, in other news, Mistral is now close to ~$2B Valuation
โถ๏ธ You can play with Mistral and other OSS models on Vercel Playground
๐ค IBM and Meta launched the AI Alliance in collaboration with over 50 Founding Members and Collaborators globally including AMD, Hugging Face, CERN, Linux Foundation, NASA, Red Hat, Oracle, Stability AI and several others including some top research universities. The alliance has come together to support open innovation and open science in AI. Their goals are to foster an open community which is very different from OpenAIโs Frontier Model Forum which focuses on ensuring the safe and responsible development of AI models.
๐ฉ๏ธ Want GPT-4 but you cannot because the new ChatGPT Plus Subscription is paused? Well, Microsoft will be adding GPT4-Turbo to their Copilot along with Dalle-3 image generation!
What brings us to awe ๐ณ
๐ซฅ How one misplaced
breakstatement caused a $60M loss to AT&T in 1990.
๐ฎ Rockstar Games dropped a trailer of
GPT-6GTA-6 and itโs probably better than most movies I watched in itself (Nibbler P is already saving up for a PS5).๐ธ Google launched a new AI experiment called Instrument Playground, which allows you to generate, play, and compose music inspired by instruments across the globe with the help of Google AI
๐ Remember when Claude 2.1 was having trouble finding information from its 100k context window? Seems like some good prompting fixes it. The team achieved significantly better results on the same evaluation by adding the sentence โHere is the most relevant sentence in the context:โ to the start of Claudeโs response. This was enough to raise Claude 2.1โs score from 27% to 98% on the original evaluation.
Source: Anthropic Blog
Today I(we) Learnt ๐
๐๏ธ You can disable "Explorer: Compact Folders" in the settings to prevent empty folders from behaving like this [Source: @flaviocopesโs Tweet]
โ๏ธ If you are like us and bothered by movements of windows that you have not configured. Hereโs how you can disable Mac auto-changing the order of windows.
๐ค You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying ๐
๐ชด Garden of AI is a new type of assistant that has a better understanding of what you ask and can handle any task you throw at it. It can do things like open a camera and capture a picture.
๐๏ธ DubbingAI: Free real-time AI voice changer. (itโs like RJ Naved as a service)
๐ LLM Visualizer: a free LLM visualizer, explains the concept used and how it happens in the LLM using 3d visualization
๐ข onsites.fyi: Learn from hundreds of real tech interview experiences!
๐ซ Watermark Remover AI - Remove watermarks from your images in an instant.
Buildersโ Nest ๐ ๏ธ
โ๏ธ github-script: Write workflows scripting the GitHub API in JavaScript
๐ซฐ๐ป promises-training: Practice working with promises through a curated collection of interactive challenges.
๐งฐ aws-lite: A faster alternative to
@aws-sdkwith better error reporting.๐ฅ Hotshot XL: State-of-the-art AI text-to-GIF model trained to work alongside Stable Diffusion XL
๐ Vector Database Feature Matrix - Excellent Vector Search comparison sheet by features/APIsโฉ. Link to sheet here.
๐ MLX - An ML framework for machine learning on Apple silicon, released by Apple Research
Meme of the week ๐
Off-topic reads/watches ๐ง
๐ As adults, I think we are just finding ways to make Study Groups so that we are in check by some like-minded folks. All Discord, WhatsApp groups, and some newsletters are just that.
Seth wrote a small piece on "Study Groups" which triggered this thought.โAsk HN: What Side Projects landed you a job? we get asked this a lot. I think the key is to not build to get a job. Itโs really similar to how you give the best interview when you already have an offer in hand.
Wisdom Bits ๐
โExperience only teaches the teachable.โ
โ Huxley
If you liked what you just read, recommend us to a friend whoโd love this too.
Weekly Standup ๐ซก
Nibbler P has been busy juggling tasks this week (just a week more and then heโll be a free bird) and trying to solve AoC. He failed one of them so he is compensating for that on the weekend.
Nibbler A had a busy week at work too, though was able to take some time out to do AoC daily. Had some plans to chill on the weekend, but a bug from the week is still living in his head and heโs trying to squash it. Also, like a good Todo san fan, he caught up with JJK.
