A week in Generative AI: NVIDIA, IRL Slop & World Models
News for the week ending 12th January 2025
First week of Jan always = CES which always = a lot of random tech announcements, all trying to cram AI into them to make them sound clever and exciting. This year’s CES was no exception and so whilst there were a lot of AI announcements, I’m not sure how many of them were that significant. The main star of the show, AI-wise, was NVIDIA who announced a huge number of new capabilities. There was also a lot of Slop announced. We also heard this week that Google is getting in on the ‘World models’ game, following announcements from lots of other frontier AI companies last year in a similar vein. Lastly, there was a really good video recap of all the progress Boston Dynamics made with their robots last year.
It was a very quiet week on the AI ethics front, but not the general digital ethics front with the unfortunate announcement from Mark Zuckerberg that they’ll be dialling back on fact-checking and moderation. Whilst not specifically AI related, I did share my thoughts in LinkedIn.
There are some good long reads this week though, with posts from Sam Altman, Simon Willison (who is quickly becoming my favourite AI commentator!) and Ethan Mollick.
Enjoy!
NVIDIA’s CES announcements
The big news of CES, and of the week in general, all came out of NVIDIA’s key note speech. Jensen Huang has now become the undisputed king of keynote speeches - he was on stage for nearly 2 hours! Below is a summary of some of the less technical, more exciting things that were announced by NVIDIA’s CEO:
Nemo
Nemo is a ‘digital employee onboarding and training system’. This announcement stirred up some controversy as NVIDIA positioned the IT team as the future HR team (?!). However, the underlying idea of the technology is sound - if AI agents are to start having an impact on the workplace, we need a good way to onboard and train them and NVIDIA wants to help make that happen.
Cosmos
Cosmos if NVIDIA’s take on ‘World models’ which seem to be all the rage right now. It can simulate real-world physics from being trained on 20m hours of video data and has been open sourced in the same way Meta open sourced their Llama large language models. This is a big boon for robotics developers as they can now use Cosmos to train their robots for free.
Isaac GR00T
This is a full end-to-end robotics training platform that includes simulation and the ability to use VR and human operation to demonstrate and train robotic systems. The idea is that this will accelerate the development of general purpose robots and bring about the ‘ChatGPT moment’ for robotics.
AI Agents
NVIDIA also took their announcements around Agents further than just Cosmos, announcing a whole agent ecosystem. These aren’t general agents that can perform any task, but specialist agents specifically designed for narrow tasks like research, software security, and monitoring camera feeds. This is really where the agentic-era of AI will start with more generalist agents that can perform most tasks probably not coming to market until later this year at the earliest.
These were all some great announcements from NVIDIA, especially following the huge volume of announcements we got from both OpenAI and Google last month, as well as Amazon announcing their first entry into the frontier AI model arena the month before. We’ve now had new announcements from all of the significant AI companies in the last couple of months with the exception of Anthropic and Meta. I suspect we’ll be hearing a lot more from them in the coming weeks!
CES 2025 was full of IRL AI slop
Another year, another CES and the vast majority of technology announced is ‘AI’. Want an AI spice dispenser? Spicerr’s got you covered. Need AI in your AirFryer? Look no further than Dreo.
With all the hype around AI, and the huge promise it has, we still don’t really know what to do with it, which is why everyone is literally throwing the kitchen sink at it. I think what this highlights is that it doesn’t matter how quickly the technology progresses, it just takes humanity a little while to catch-up and work out what to do with it.
We saw the same thing with the Internet and Mobile Phones - it took time for us to work out exactly what to do with these technologies before they really took off. I do think that the current AI-hype will be similar but different - use cases will still lag, but not by the 5-10 years it’s taken for previous technologies. Here’s hoping we see some more interesting and meaningful implementations of AI at CES next year!
However, If you do want to see some of the good tech that CES had to offer, The Verge has some great video coverage below:
Google is forming a new team to build AI that can simulate the physical world
‘World models’ seem to be shaping up to be the next frontier. We first had large language models, now we have the emergence of large reasoning models (o1, o3 etc.) and next I think we’ll have ‘world models’ but it make take a year or two to see where it could take us - not much longer though!
The concept behind ‘World models’ is to train AI systems on real world data so that they can get a physical understanding of the world. At the moment, all large language models and large reason models can only understand the ‘digital’ world - they don’t have a grasp of physics (with the exception of some emerging understanding in the large video models like Google’s Veo 2) which leads to the models not really having much common sense.
This is best expressed by Moravec’s Paradox - AI is great at advanced intelligence tests, coding, science etc. but still can’t do many things a one year old can. ‘World models’ are trying to solve this. Yann LeCun of Meta fame has been talking about this for years, and there are lots of companies now looking at this approach. Will be great to see where this takes us!
Boston Dynamic’s 2024 In Review
It’s great to see some of the highlights of Boston Dynamic’s robotics from last year. I’m still a little weirded out by the unnatural movements of their new electric version of Atlas. I think we’re going to have to get used to human-like forms moving in unnatural ways over the coming years!
AI Ethics News
Google researchers can create an AI that thinks a lot like you after just a two-hour interview
AI researcher François Chollet is co-founding a nonprofit to build benchmarks for AGI
OpenAI is having a rough week - it could be the start to a rough year
Long Reads
Sam Altman - Reflections
Simon Willison - My AI/LLM predictions for the next 1, 3, and 6 years
One Useful Thing - Prophecies of the Flood
“The future is already here, it’s just not evenly distributed.“
William Gibson
Do you think that ai is being used more as a marketing tool Instead of being used as a function? For example the air fryer using ai how is that useful in any way?