No big AI news this week, although OpenAI announced Daybreak, their semi-answer to Anthropicâs Mythos model and Glasswing initiative. There was also a great video from Google DeepMind on putting AI into a mouse cursor (bear with me on this one!) and a ridiculous exoskeleton robotics video from Unitree.
In Web 4.0 news, OpenAI have launched ChatGPT for personal finance, just in case you needed that (?!) and in Ethics news it was reported that data centres now account for 6% of electricity supply in UK and US.
OpenAI just released its answer to Claude Mythos
So, OpenAIâs Daybreak isnât a model, but more a way for cyber-security experts and large technology organisations to do similar security testing work as Anthropic are doing with their Mythos model.
This is quite different I think. Anthropic built a new model (Mythos) that they realised was incredibly capable to the point where it could be dangerous to put into consumerâs hands. Mythos was specifically very capable at coding and identifying/exploiting/fixing cybersecurity vulnerabilities sp Anthropic decided to make it available to a small number of organisations so they could get ahead of the cybersecurity threat that Mythos would pose when they release it.
OpenAIâs approach isnât based on them building a new model that they think is incredibly capable/dangerous. Theyâre just trying to give organisations similar cybersecurity capabilities using their existing models.
This difference doesnât matter now, but will matter when (and I do think itâs when, not if) Anthropic decide to release Mythos to the general public. Thereâs a chance that OpenAI are then behind the frontier for the first time as Anthropic will have a model more capable that anything else available.
Both Google and Anthropic have been behind the frontier of AI capabilities at certain points over the last 3.5 years, but generally OpenAI havenât been. Theyâve always been either at the frontier or pushing the frontier forward.
Itâll be interesting to see how they react to being behind the frontier for the first time.
Reimagining a 50-year-old interface (the mouse pointer) with AI
This is one of those things that when you think about it is incredibly simple, but I am sure was technically very challenging. The phrase I like best from the video is how you âshare attentionâ with an AI model - essentially, how do you make sure an AI model knows what youâre looking at or referring to when you ask it to do something.
If youâd asked me before seeing this video if it would be a good idea to put AI âintoâ a mouse cursor I would probably have laughed, but now Iâve seen this video and Iâve though about it, the more I think itâs a great idea and part of the future of AI-powered users interfaces.
Love it!
Unitree unveils GD01
This is a very weird, ridiculous video but I had to share it because it was funny to watch, but also itâs inevitable that something like this is actually in our future.
So, Unitree have been pushing the limits of whatâs possible with humanoid robotics and made a lot of very rapid progress. This however, is not that.
The video shows an âexoskeletonâ that in theory would have a human operator inside, but I canât see any controls and the part of the video where itâs moving with someone in it is clearly a dummy and not a real person. I also donât really see how a human operator could operate it when itâs on all fours as they wouldnât be able to see where theyâre going!
Web 4.0
AI Ethics News
Datacentres using 6% of electricity supply in UK and US, research says
George Clooney, Tom Hanks, and Meryl Streep back new âHuman Consent Standardâ for AI licensing
Netflixâs new studio will produce animated shorts with generative AI
Who decides what AI tells you? Campbell Brown, once Metaâs news chief, has thoughts
ArXiv will ban researchers who upload papers full of AI slop
Long Reads
Daring Fireball - AI Is Technology, Not a Product
Stratechery - The Inference Shift
âThe future is already here, itâs just not evenly distributed.â
William Gibson