Friday Links: Chips, Sausages and Presidential Candidate Clones

Here are the news items that stood out to me this week. I also changed the name of these Friday posts. Smokebox was fun if you’re a train enthusiast, but sometimes it pays just to be to the point :-). “Friday Links” it is..:

• Why NVIDIA’s AI Supremacy is Only Temporary. Peter Warden’s post on NVIDIA is pretty contrarian but is a good reminder that things can change fast. He points out that everyone is currently focused on training neural networks (NVIDIA chips rule this), but that in the long run, it’ll be much more about inference (running the model to make decisions), for which many other chip types also perform very well. Maybe ARM was worth it’s 25% stock price bump at IPO.

• Microsoft offers legal protection for AI Copyright infringement. Derisking the use of its tools makes a lot of sense for Microsoft, but it’s a move that likely surprised most people in the industry.  Microsoft is no doubt making a bet that most claims will fail, or expenses will not be high. However, this is a heck of a bet to take. Competitors may be forced to follow suit and, given the uncertainty of the risks involved, will need deep pockets.
• Chat2024. You can now talk to the clones of each candidate for the 2024 US Presidential Election. Downright scary. I wonder how long it will take for takedown notices to arrive. Still, maybe the clones are better politicians than the candidates themselves.

• Würstchen - Fast Diffusion for Image Generation. Apart from winning the title for best image generator model so far, this showcases another possible leap in model training speed (and inference). The system heavily compresses images before training, so the training cycle happens in a much smaller latent space than previously.  The potential increase in training speed is 10-15x, though there is some accuracy cost.  Paper, Team Video.  Seeing how fast researchers are stripping down training processes to their essentials is stunning.
• Roblox’s AI Chatbot Assistant for World building. This is still hype (launching at the end of the year), but it’s clear to see that adding tools like this to Roblox will be a big win. The Roblox creator demographic is part-time, single-person, or small teams. Text-to-World capabilities allow these types of creators to speed up content creation greatly. There are plenty of startups working on similar text-to-world systems and I’m sure there’ll be multiple winners (even within the Roblox ecosystem). Those targeting Roblox and UEFN will likely have an advantage over those for Unity or Unreal in that game development cycle times are so much faster.

As a final item, Stability.AI just released an AI audio platform. This is notable for two reasons: 1) this is another big player in a space that’s already quite crowded, 2) the service is a web app, not a released or open-sourced model.

Wishing you a great weekend!