Llama 3.1, Interactive Avatars and Mathematical AI – Live and Learn #47
Welcome to this edition of Live and Learn. This time with the release of the Llama 3.1 model by Meta, the announcement of HeyGen Interactive Avatars, and an AI achieving the equivalent of a Silver Medal at the International Math Olympiads. As always, I hope you enjoy this Edition of Live and Learn!
✨ Quote ✨
A system being unpredictable doesn't mean that it is enchanted, and magical explanations for things aren't really explanations.
– Robert Sapolsky - (source)
Links
Advanced Humanlike Robot Hand Demo by Clone Robotics. This to me is one of the craziest robot hands that I have ever seen. It even looks like the robot designs from series like Westworld. It sits right in the uncanny valley, where it's a bit too human. Science fiction is almost here. You should check out the Clone Robotics website to see what else they are up to. Eventually, they will build a whole human-like robot body with artificial muscles and all.
Eureka Labs Announcement by Andrej Karpathy. Andrej Karpathy announced Eureka Labs, a new way forward for education, using courses enhanced by private AI tutors that help you learn, comprehend, and digest the course materials. They have a Github organization too and are working on building the best course on machine learning over there as well.
Crowdstrike Report by Fireship. In the last two weeks, a bug in the Crowdstrike Falcon software has crashed more than 8 million Windows machines due to a software update leading to writing into a wrong memory location from a program running in Kernel mode... This to me screams: "Use Rust already!" For a more in-depth read on what happened and what can be learned from this trainwreck, you can read this post on the Pragmatic Engineer blog. It's technical but really good.
Real Time Interactive Avatars by HeyGen. I covered HeyGen and their avatars a while back already but they massively improved their product. As a recap: HeyGen can create videos of a person talking with only a couple minutes of video input from that person. It used to take a long time to render these artificial videos, but now this can happen in real-time. This means that you can have a Zoom call with your own AI avatar. Parse the conversation using Whisper, feed it into ChatGPT, give the output of text from ChatGPT to HeyGen, and voilá: see yourself, animated but looking totally real, talking back to you. The implications of this for deepfakes and phishing are horrifying. You can't even be sure that video calls with your relatives or friends are real anymore. From now on, everybody you interact with outside of real life could be an AI-powered avatar. Better get ready to have some secure passphrases to identify family members and friends.
AI Silver Medal at the International Math Olympiad by Deepmind. This might be the biggest announcement in the history of AI, at least for now. Deepmind seems to be inching closer and closer to cracking real reasoning. Not just memorization and pattern recognition like we see with LLMs, but actually difficult reasoning on par with very clever humans. I wonder how well their approaches will generalize to areas outside of mathematics, and where this research goes within the next year, but it's tremendously exciting (and equally terrifying) to see this sort of progress.
AI is already taking Jobs in the Video Game Industry by Wired. I've spent a lot of time thinking about whether or not people are going to lose their jobs because of AI in the future. Many people disagree with this notion because they think that AI is limited in some way or another, but fundamentally I think this is already happening and this article by Wired looks at the issue of artists already being displaced in game companies by the use of generative AI tools.
Search GPT by OpenAI. OpenAI continues the productification of its AIs. This time starting to take on Google and Perplexity (and maybe Bing) by offering their own search product. They basically promise to fix the issues with hallucinations and misinformation that are so prevalent with other AI search products. But right now it's only a prototype with limited access, so it remains to be seen, how well they deliver on that promise.
Llama 3.1 by Meta. Meta released a new version of their Llama model. It's on par with other big but closed-source models like GPT-4o, Google's Gemini, or Claude Sonnet. However, it's sort of open-source. You can download the model weights and they are building an open ecosystem around the whole Llama family of models. Marc Zuckerberg is also all-in on this vision of open-sourcing AI research, at least for now.
🌌 Life 🌌
The last weeks were more slow, I fell sick for some time but recovered by now–just in time for two concerts from Lawrence and Jamie Cullum. Both were absolutely amazing. I also spent some time in Wittenberg and in rural Dommitzsch visiting my family and got a piano for my home.
🎶 Song 🎶
Hello by Hania Rani
That's all for this time. I hope you found this newsletter useful, beautiful, or even both!
Have ideas for improving it? As always please let me know.
Cheers,
– Rico