A week in Generative AI: Energy, Human-Level AI & McDonald's
News for the week ending 14th July 2024
It’s been a quieter GenAI news week, we’re definitely getting into a summer lull, but that allows us to focus on some of the things that don’t usually get as much attention! Again this week, we have some updates on AI energy usage, a change in the governance of OpenAI with both Microsoft and Apple dropping their Board seats. An unofficial McDonalds ad was released that was made just using GenAI tools and Google DeepMind develop a robot that can navigate their offices using Gemini.
Google claims new AI training tech is 13 times faster and 10 times more power efficient
Following on from my slightly deeper dive into AI energy usage last week, this is a third solution that I hadn’t thought of to help bring AI’s huge energy usage under control - software optimisation!
This research from Google shows that they can accelerate model training and increase energy efficiency using something they call joint example selection (JEST). This approach uses a smaller AI model to quality-score the training data in batches and then trains the larger, final model on the highest quality batches. Essentially, AI is being used to help improve model training.
So, to improve on my thoughts last week, we now have three methods for improving AI’s energy usage: generate more carbon-neutral energy, develop less power-hungry AI hardware, and optimise AI software for training and use.
OpenAI Develops System to Track Progress Toward Human-Level AI
It’s interesting that this system hasn’t been shared publicly yet, but from what we do know it was announced at an internal all-hands meeting last week. There are five levels in the system, ranging from conversational interactions at level 1 (where we are now) to level 5 when an AI can do the work of an organisation (🤯).
Level 2 is what OpenAI calls ‘reasoners’ which is when AI can do basic problem-solving tasks as well as a human with a doctorate-level education (and importantly doesn’t have access to any tools). OpenAI believe that they are on the cusp of reaching level 2, which puts down an interesting marker for GPT-5 when that’s released.
Microsoft and Apple drop OpenAI seats amid antitrust scrutiny
Both Microsoft and Apple have given up their ‘observer’ seats on OpenAI’s board, effective immediately, due to increased regulatory scrutiny of Big Tech’s investments in AI start-ups, specifically around antitrust.
I’m not sure how I feel about this - on one hand, we don’t really need Apple and Microsoft observing OpenAI, we need more independent representation on their board from academia and civil society. However, some oversight from third parties is better than nothing and I do think that Microsoft especially have the right to a board seat given their $13bn investment in OpenAI. If regulators are worried about Big Tech’s investment in AI start-ups then they shouldn’t be allowed to invest in them. Threatening antitrust lawsuits isn’t the right way to go about regulating this rapidly evolving technology.
McDonald’s (unofficial) AI Commercial: A Taste of Tomorrow
This is quite a fun little (unofficial) AI ad for McDonalds made by MetaPuppet. They wrote the script themselves but then used MidJourney to generate images and Magnific to upscale them, LumaLabs and Runway to generate video, ElevenLabs to generate the voice over, and created the music with Suno and Udio. They then edited it all together.
MetaPuppet says that it cost them “less than 60 happy meals to make” which puts it somewhere between $300 and $500. This ad is obviously not professional grade and wouldn’t air on prime time TV, but was at least 1,000 cheaper to make than a typical TV ad. We’re not far off the point where the cost/quality equation tips in favour of GenAI commercials for many people, probably in about a year when the technology has improved further and the quality is higher.
Google DeepMind develop a robot that can navigate their offices using Gemini
We’ve seen a few robots powered by Large Language Models over the last 12 months, and this isn’t much different. What is interesting though is how they trained the robot. The Google DeepMind team essentially walked the robot around the office while pointing out different landmarks with speech, exactly how you would show a new employee around the office on their first day. Pretty impressive!
AI Ethics News
Court ruling suggests AI systems may be in the clear as long as they don't make exact copies
Destroying Russian Tanks Is Just The Start For U.S. AI Drone Autopilot
First “Miss AI” contest sparks ire for pushing unrealistic beauty standards
Chinese developers scramble as OpenAI blocks access in China
Summary of NATO's revised Artificial Intelligence (AI) strategy
It’s not just hype. AI could revolutionize diagnosis in medicine
Long Reads
In Good Company Podcast - Dario Amodei
AI Engineer - Exploring Multimodality with OpenAI: Romain Huet
“The future is already here, it’s just not evenly distributed.“
William Gibson