Hardware bets, software wins

Advertise | Job board | Community | Investor + Founders

Hey folks, this is Ben’s Bites. Yesterday I tried to send this from our custom domain and I f*cked it up 😅. Can you reply and let me know if this email reached your inbox ok? ❤️

Today we talk about how Microsoft put skin in the AI game and is now reaping the benefits, laying off its AI ethics team, how AI is changing everything and using Meta’s LLaMa on Google Colab.

Let’s get to it.

🤌 Our Picks

Highlights if you've only got 2 minutes

1/

Microsoft spent hundreds of millions on hardware to get ahead in AI

OpenAI really hit the jackpot when they decided to pitch the need for ridiculous amounts of computing power back in 2015. And Microsoft really hit it out of the park when they decided to say yes. Since then, Microsoft have reportedly spent hundreds of millions on thousands of NVIDIA GPUs and new ways to build hardware at a scale never seen before. All rooted in the conviction that all this computing power would one-day yield results in the form of the first truly spectacular AI machines. And here we are in 2023, with ChatGPT shortly followed by Bing AI, an absolute overload of hugely useful tools, and a decent chance that the world is about to change… Jeez, I think I need to chill out a bit.

2/

AI changes everything

We tend to overestimate how long it will take AI to ‘unrecognisably transform the world’. It’s likely to be years, not decades, and there’s a real chance its months. Are developers too close to the tech? Do we understand what we’re building towards? Can we even go slower? Or do we need to adapt faster? This piece is an interesting look at some of the questions swirling around a lot of our minds.

(without paywall here)

3/

Meta’s LLaMa 4-bit works on the free Google Colab

There’s nothing quite like a short tutorial to move something from being quite complicated to a lot less so, and this guy always pulls a rabbit out of the hat with his ones. Over the past few days, we’ve come to realise how big Meta’s LlaMa model might be for democratising the use of powerful language models, and this tutorial takes us through using it within the free version of Google Colab - a Jupyter notebook-style interface on the cloud that allows limited use of GPUs. In bitesize steps, it takes you through setting up your Colab, downloading the Llama model code, and running it on a sample dataset. (In case you haven’t noticed, I still don’t quite know which letters to uppercase in llama…)

4/

Microsoft just laid off one of its responsible AI teams

Microsoft's recent layoff of its entire AI ethics and society team has raised concerns about the potential risks posed by the company's AI products. The move has left Microsoft without a dedicated team to ensure that AI principles are linked to product design, although the company still has an Office of Responsible AI. Employees are worried that the absence of a dedicated team may lead to products not being designed with responsible principles in mind.

🛠️ Cool Tools

Product launches, updates and demos

*quick note for Marcbot (a helpful Telegram AI bot) - First 100 people to send the command /bensbites to the bot will get 300 free messages. (Marcbot won our Whisper API hackathon recently).

• Creator Kit - AI-generated product photos that sell. (link)

• Guardrails AI - An open-source package to add SLAs for LLM outputs. (link)

• Pebblely is now on Shopify AI - Turn boring product images into beautiful creative content. (link)

• Aish - Copilot for the terminal. Generate commands and preview them before running. (link)

• Stanford Alpaca - An instruction-following language model demo for research purposes only. (link)

• Swift GPT - The native macOS app for ChatGPT. (link)

• Trudo AI - No-code platform for building OpenAI apps. (link)

• Mobile GPT - ChatGPT on your mobile device through WhatsApp. (link)

• Dime a Dozen - Validate your business idea in seconds. (link)

• Replai - ChatGPT replies in Twitter, LinkedIn. (link)

• Deepbrain - Create AI-generated videos using basic text instantly. (link)

• Relume Ipsum - Generate website copy quickly and easily using AI, all without leaving Figma. (link)

🤓 Miscellaneous

News, podcasts, videos, blogs etc

• I tried 4 of the best AI productivity tools for a week to make my job easier, but they just made my life harder. (link)

• Discord revises its privacy policy after backlash over AI. (link)

• How AI changes everything. (link, without paywall here)

• Meet the AI expert who says we should stop using AI so much. (link)

• How Microsoft’s bet on Azure unlocked an AI revolution. (link)

• Should we worry more about short-term AI risk or long-term AI risk? (link)

• Microsoft just laid off one of its responsible AI teams. (link)

• Custom AI chatbots are quietly becoming the next big thing in fandom. (link)

• How will generative AI disrupt video platforms? (link)

• OpenAI asks Faculty to help businesses adopt generative AI. (link)

🎓 Learn

How-to’s and resources

• Powerful machine learning library for Node.js that uses Python's scikit-learn under the hood. (link)

👋 Too many links?! I created a database for all links mentioned in these emails. Refer 1 friend using this link and I'll send over the link database.

🔬 Research

Published research papers

• One transformer fits all distributions in multi-modal diffusion at scale. (link)

• Erasing concepts from diffusion models. (link)

• High-throughput generative inference of large language models with a single GPU. (link)

• NeuSE - Neural se(3)-equivariant embedding for consistent spatial understanding with objects. (link)

• Scaling vision-language models with sparse mixture of experts . (link)

• InferFix from Microsoft - End-to-end program repair with LLMs over retrieval-augmented prompts. (link)

• Vision-language models as success detectors from Deepmind. (link)

• Meet in the Middle - A new pre-training paradigm. (link)

• OTOv2 - Version 2 of only train once. Automatic, generic, user-friendly. (link)

• NeRFLiX - High-quality neural view synthesis by learning a degradation-driven inter-viewpoint mixer. (link)

• Stabilising transformer training by preventing attention entropy collapse. (link)

• Synthetic experience replay - A diffusion-based approach to arbitrarily upsample an agent's collected experience. (link)

• Transformer-based world models are happy with 100k interactions. (link)

📰 Unclassifieds

Short, sponsored links 

• Langchain is hosting a session on how to build a chatbot on your documents on 15th March. (link)

• Webinar on applications of generative AI with Replit and NVIDIA on March 15th. (link)

• LLM summit with Pixelbeasts will be happening today. Register here >>

• Gables Search Group is hiring a machine learning engineer. Learn more »

• Diagnostic Robotics is hiring a data scientist. Learn more »

Have a product, service, job, event, newsletter, app, book, movie, tool, or anything you'd like to share with over 45k subscribers? 
Advertise | Job board | Community | Investor + Founders connect

🖼 AI Images of the day

Funny memes and pics from around the web

Send this to 1 AI-curious friend and receive my AI project tracker database! Use this link.