The Biggest Compilation of AI News on the Internet #4

Welcome to Nofil’s News Corner edition #11.

Moving forward, my efforts outside of consulting will be heavily concentrated on this paid newsletter. This newsletter is the reason I’ve been able to afford doing all this research and start my own firm. I appreciate all of the support.

If you’re already a paying subscriber, sign in to view the entire newsletter. If not, you can subscribe here.

• a16z built AI town, a js starter kit based on the Stanford Smallville paper where 25 AI agents live their lives. Link to code [Link] Link to demo [Link]

• OpenAI acquired Global Illumination, a company that made an open-source minecraft clone. There’s talk they could integrate AI agents into it. A multi-agent AI civ sim?? Will be very exciting to see what comes from this [Link]

• Character ai is a silent giant in the AI chatbot race. The statistics are staggering with 9m daily visits and almost 30 minutes per visit. That is absolutely insane [Link]

• Inflections Pi is being built on Microsoft Azure [Link] Azure is going to be a force to be reckoned with in the AI powered future

• OpenAI chief scientist, Ilya Sutskever gave a talk on unsupervised learning. Compression is related to learning is fascinating. Link to youtube [Link] Link to lecture page [Link]

• A fascinating paper on what open-source AI even is. “We find that the terms ‘open’ & ‘open source’ are often more marketing than technical descriptor, and that even the most 'open' systems don't alone democratize AI”. Link to paper [Link]

• New LM from Meta called Shepherd specifically tuned to critique model responses & suggest refinements [Link]

• Someone made a system that takes in a company’s docs and uses a bunch of agents to create a dataset and train a LLaMA 2 for you. So much application for agents [Link]

• Fireworks ai, a new platform to help developers use LLMs [Link]

• A great article exploring how llama.cpp is possible [Link]

• AI reconstructed a song from simply looking at brain activity.. and its very clear. Massive implications of this [Link]

• Fine-tune llama 2 on text-to-SQL makes it a lot better at writing sql using llama index [Link]

• A fine-tuned llama 7b to generate optimal gpt4 prompts [Link]

• Apparently AI has already generated 150 billion images which is equivalent to like 150 years of photography by humans [Link]

• A great breakdown on how to train LLMs to do really good at benchmark tests. You’ll understand how exactly the WizardLM team have done so good using these techniques [Link]

• API support for the Hermes 13b model and its much cheaper than gpt-3.5 [Link]

• Meta released Self-Alignment with Instruction Backtranslation - it can automatically label text with instructions and pick ideal ones for finetuning. No humans required in the RLHF. This was apparently used for llama also [Link]

• NeoGPT - a new way to us AI agents by guiding them in the right direction. Looks very interesting [Link]

• The Unitree H1 is a 90k humanoid robot. $90k is pretty cheap for a humanoid robot if it actually works well [Link]

• A really interesting thread on multi-modality and how bytedance, the company behind tiktok uses it. In future all LLMs will be multimodal eventually [Link]

• Learn about LLMs with semantic search with Andrew Ng [Link]

• Saudi bought 3000 nvidia chips. Saudi and UAE could be big players in this space moving forward [Link]

If you want to read the rest of this newsletter which I spent tens of hours curating, you can subscribe here for $5/m. This is the biggest collection of AI advancements, news and research on the internet. You will get your money’s worth.