GPT-5, AI Gold at IMO and Genie 3 – Live and Learn #73

Welcome to this edition of Live and Learn. This time, OpenAI unveiled their newest and most powerful model yet, while Google has released the successor to their Genie models, and Meta gave a peek into their ambitions for the age of superintelligence. As always, I hope you enjoy this edition of Live and Learn.
✨ Quote ✨
How many of my preferences are genuinely mine, as opposed to some elaborate psychological strategy to feel socially safe? Even when I like things that are ‘low status’ I feel cool for seeming ‘immune to status’, which… is high status. So who am I?
– Aella - (source)
I find this funny, because in a way, including this quote (or anything else I write in this newsletter) is by definition a status signal, even though I would like to be immune to status. But yet, here I am, signaling: "Look, I have good taste, here's something nice to make you think (and hopefully learn something new)".
Realizing this and writing it down for you to read makes me chuckle, in a sort of recursive meta way, because again—waves hands in the air: status signals.
Links
GPT-5 by OpenAI. OpenAI just dropped their latest model: GPT-5. From now on, there won't be any more confusion as to which model to use; GPT-5 will do the choosing for you, with an internal router deciding which model will best answer your question. That way, OpenAI combined the strengths of previous models into one. You just use GPT-5 and trust that it gives you the best output possible, sometimes guiding it to "think hard about this" or "use reasoning" in your prompt to help it select a stronger model. So far, opinions seem to be mixed. I've read people claiming that it's just a cost-cutting move by OpenAI, basically charging you for a premium model while serving you their less expensive ones. Simon Willison likes it: "My verdict: it’s just good at stuff." and I have seen similar sentiments echoed elsewhere from Theo or Ethan Mollick: "GPT-5 just does stuff, often extraordinary stuff, sometimes weird stuff, sometimes very AI stuff, on its own. And that is what makes it so interesting." Overall, though, it is not the groundbreaking AGI moment people on Twitter have been hoping for, "just" a pretty solid model + UX improvements.
Genie 3 by Google DeepMind. Genie 3 is the latest improvement in Google's series of world models, and it starts to look like Midjourney back in 2023, not for images but for fully interactable, and consistent 3D worlds. It feels like a reality simulator, an infinite game, where anything you can type can become a controllable, photorealistic 3D video game in real time. Where LLMs are simulators of languages, Genie models are simulators of worlds. I can totally see a future where we start "traveling" into these weird realms—where you "explore" the dream of a Genie model to find something interesting, something valuable, worth sharing. You might sense by my use of italics that I am very excited about this. And true, to me, this, much more than anything else that has happened in the past few weeks, seems like the big milestone to remember.
Personal Superintelligence by Meta. Mark Zuckerberg's Meta has been making the rounds in the AI news hype cycle lately with its crazy cash offerings. Hundreds of Millions of Dollars of compensation for top AI talent seems to be the new hiring norm, apparently. This essay describes their vision and why they hire like this, how they imagine a world where everybody will have access to a personal superintelligence that helps them live their day-to-day lives and make decisions. It leaves out that you can make a lot of money in brand deals if everybody "improves" their lives using your AI, but that is to be expected from a former Facebook. Somehow, all those millions invested in GPUs and AI talent have to earn their ROI.
AI Gold Medal at IMO by Google Deepmind. The IMO is a competition for very good mathematicians, and getting a gold medal there is really, really hard. But AI is now up to the task, showing just how far reasoning models have come in the last few years. Deepmind partnered with the IMO, participating in the competition with a model of their Gemini family, and got an officially verified gold medal, which showed that their AI can solve very advanced problems. OpenAI's reasoning models also achieved a gold medal performance with a small team pushing for it, but their solutions were only verified internally.
🌌 Travel 🌌
I went back to Berlin for a couple of weeks and enjoyed the slightly rainy summer there. But I miss my bicycle and the grind of going up and down mountains every day, haha. Now I am in Istanbul, traveling with a good friend of mine around Turkey for a couple of weeks before returning to Germany. Istanbul is a wild city, very vibrant: a lot of people, culture, and history. I haven't seen all that much of it yet, but I somehow thoroughly enjoy the vibe that this place has.
🎶 Song 🎶
The Dane by Nils Frahm
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
Subscribe to Live and Learn 🌱
Join the Live and Learn Newsletter to receive updates on what happens in the world of AI and technology every two weeks on Sunday!