E38: 5/3/24
MSFT's multi-$1B renewable energy investments, Amazon Q finally in GA, news outlets piling onto OpenAI lawsuit, criticism on overfitting LLMs towards benchmark evaluations, and more...
Those of you who know me well know I nerd out over utility-scale renewable energy systems, so this is a fun week for me. See headline #1. :)
Enjoy!
PNW AI News:
Microsoft just announced development of 10.5 gigawatts of renewable energy capacity (costing $11.5B-$17B), to offset its datacenter energy consumption from AI usage and march towards "net neutrality", financed by Brookfield Asset Management
This is the largest-ever corporate renewable energy purchase agreement IN HISTORY. Equivalent to 3.3K utility-scale GE wind turbines
Amazon unveils new features in GA preview for their Amazon Q product, which now “allows employees to easily and quickly create generative AI-powered apps based on their company data in natural language, without requiring any prior coding experience”
Amazon Q had a turbulent unveil in Dec 2023, with severe hallucinations and leaks of confidential data in public preview
Financial analysts review the CapEx spikes born by the cloud providers as MSFT, GCP, and AWS all bet big on AI demand with multi-$B datacenter and chip innovation investments
Microsoft bans US police departments from using Azure OpenAI Enterprise services for facial recognition, amid fears of hallucination
Comes a week after Axon announces a new product using GPT-4 to summarize audio from police body worn cameras (BWC), like Truleo's product
Sanctuary AI announced a new partnership with Microsoft to advance AI development for humanoid robots utilizing Azure infrastructure
Key AI Product/Research Updates:
Rumors swirled earlier this week about a mysterious "gpt2-chatbot" LLM that outperformed GPT-4, Mixtral 7x8B, and LLaMa-3 on several benchmarks
Pundits thought this was OpenAI testing GPT-4.5 in stealth, but Sam Altman denied this claim
Tesla’s next-gen Dojo AI training tile is in production, according to supplier Taiwan Semiconductor Manufacturing Company Limited (TSMC)
Helping on-EV AI training to finetune self-driving models by continuously injecting simulated training data
New research from Scale AI just uncovered significant ‘overfitting’ of certain LLMs on popular AI benchmarks (Mistral is the worst offender), raising questions about the validity of current leaderboards and evaluation practices
Caleb John @ Pongo has been saying this for the last 12 months! "AI Benchmarks are a marketing mechanism"
Gradient AI introduces Llama-3 8B-Instruct-Gradient-1048k, boasting an impressive capability to handle context lengths of up to 1M tokens
Myshell-ai just released OpenVoice V2, an updated version of its text-to-speech model with a revised training approach, and now supports multiple languages natively, including English, Spanish, French, Chinese, Japanese, and Korean
The viral Rabbit R1, positioned as an AI device to replace smartphones, has been scrutinized for its utilization of the Android Open Source Project (AOSP)
Effectively just an AI wrapper Android app… boooo
A new AI algorithm identified over 27K previously unnoticed asteroids from existing telescope images, enhancing the ability to monitor near-Earth objects
Google just introduced a new Chrome shortcut that lets users initiate Gemini conversations directly from the browser's address bar by tagging '@gemini'
Sam's Club is rolling out new AI technology across nearly 600 stores to speed up checkouts by verifying purchases without manual receipt checks
OpenAI introduces the "Memory" feature for ChatGPT Plus users, allowing the model to retain information across chats
The Defense Advanced Research Projects Agency (DARPA) continues to advance its autonomous AI tank program through the Robotic Autonomy in Complex Environments with Resiliency (RACER) initiative
Key AI Business/Investment Updates:
Eight major U.S. newspapers just filed a joint lawsuit against OpenAI and Microsoft, accusing the tech giants of illegally using their copyrighted articles to train AI chatbots like ChatGPT and Copilot
Publishers jumping onboard NYT's pioneering suit. Is this the end of Common Crawl?
Common Crawl is allegedly the source of 60% of GPT-3, 67% of LLaMa-2 training data. LLaMa-2 pre-training data breakdown below:
OpenAI signed licensing deal with the Financial Times (FT) for training AI models on latest media tie-ups
Ukraine’s Ministry of Foreign Affairs just introduced ‘Victoriya Shi’, an AI-generated avatar, as the department’s digital spokesperson — a first in the world of diplomacy
Sam Altman of OpenAI spoke at a Stanford seminar, saying that GPT-4 is "mildly embarrassing at best", and that it's the "worst model" we will ever use
He doesn't care if OpenAI burns $500M or $50B/yr as along as it continues course towards AGI
CoreWeave secured $1.1B in new funding from Nvidia and others at a whopping $19B post-money valuation (3X its last val) as the company expands its specialized AI cloud infrastructure
Lamini AI raised a $25M Series A from AMD Ventures, NEA, Andrew Ng, Figma CEO Dylan Field, Dropbox CEO Drew Houston, and OpenAI Co-founder Andrej Karpathy to help enterprises deploy generative AI tech
PNW AI/ML Fundings:
Advocat (Seattle, WA) raised $2.8M in Seed funding from Refashiond Ventures
A legal software platform designed to create researched legal documents from attorney conversations using AI - TFTD $4.1M
CEO: Pradnya Desh
Acumino (Seattle, WA) raised $4.2M in Seed funding from undisclosed investors according to an SEC filing
A computer vision training platform designed to enable robot makers and robot users to perform tasks autonomously - TFTD $4.2M
CEO: Patrick Jarvis
PNW AI/Startup Events:
Open Source AI Hackathon at the MSFT Reactor
May 5th: sign up here
IBM: Build Trust with AI
May 7th: sign up here
Seattle Data, AI & Security Meet-up at the MSFT Reactor
May 7th: sign up here
Seattle's AI Tinkerers
May 20th: sign up here
The Future of Robotics in Manufacturing
May 28th: sign up here