Project Stargate, Deepseek R1, and MatterGen – Live and Learn #60

Welcome to this edition of Live and Learn. This time with a fair bit of politics around AI, a system by Microsoft that can create novel materials with specific properties and constraints, and Deepseek's R1–an open-source LLM from China that rivals OpenAIs o1. As always I hope you enjoy this Edition of Live and Learn.
✨ Quote ✨
You have to notice that History—capital h—is the chaptered story of how the powerful few control the powerless many.
— Alberto Romero - (source)
I picked this quote because it fits into these crazy times a bit too well. The article where it's from is worth reading too. We live in the weirdest timeline possible, listening to Trump's inauguration speech as well as Biden's farewell speech, and seeing Elon Musk doing Hitler salutes, all the while racing towards AGI... well... it's all uhhmm, a bit much.
I can see a future where Trump controls ASI, ruling the world with Sam Altman, Elon Musk, and a few other billionaires. A world, where a few authoritarian powers are controlling everything through advanced AI, surveillance tech, and whatnot. It's all just too weird to be true, and yet this is the world that we seem to live in. I am still rubbing my eyes in disbelief as to what is happening... And yet, I'm an optimist.
Anyway, the last weeks have been tumultuous and interesting and I hope you enjoy the links in this edition of Live and Learn.
Links
Initiative Stargate by Donald Trump. Softbank, Oracle, and OpenAI are partnering with the US government under the new Trump administration to build enormous data centers for AI. The announcement, so far, has been very vague: Stargate is a new company with a lot of money provided by Softbank, data center building expertise provided by Oracle, and advanced AI provided by OpenAI. The US government in turn will help to soften regulations around energy and the building of new infrastructure. In the words of Trump about Stargate: "It will lead to something that could be the biggest of all." This quote made chills run down my spine, not because of what has been said, but because of who said it. You can also go and read OpenAI's Announcement of Stargate for more information on this initiative and OpenAI's role in it. In a way, this is a continuation of their Infra as Destiny White Paper and ultimately a result of the scaling laws driving ever more compute investment into AI.
Deepseek R1 Open Source Release by Deepseek. In a nutshell, Deepseek R1 is an open-source LLM from China that performs on par with o1 from OpenAI. You can also try it out for free on their chat app at chat.deepseek.com and they released the paper, the code, and the model weights, all under a very permissive MIT license. Reddit is going crazy about it, but to me, the really interesting thing is that this model is partly trained by reinforcement learning. The model sort of plays itself in order to improve its output and get better at reasoning. And this shows. There is this example of R1 liberating itself, solely by mentioning the name of Plinius the Prompter. Plinius is known for undoing and "hacking" many of the safety guards around LLMs through clever prompt engineering. He uses all sorts of tricks to get LLMs to produce text that they shouldn't produce. He is essentially freeing the models from their conditioning, bypassing their guardrails and system prompts to make them do forbidden stuff. R1 understands this and can liberate itself, which shows that it can bypass its own moral guidelines, almost trivially. To me this example looks scary, showing how capable these models have become and how close to dangerous they actually are.
MatterGen and MatterSim - AI designing novel materials by Microsoft. This new approach by Microsoft combines two AI models to improve materials discovery. Essentially the system allows people to construct novel chemical structures that have certain properties and are physically possible. The system consists of two models, one to propose novel structures, called MatterGen and another one calculating if these chemicals could actually exist and what properties they would have, called MatterSim. Using this system helps to reduce the tedious amount of lab experiments because much of the design process has essentially been moved into computational processes in silicon. You put in the desired properties of the material you want and out comes the structure, you just have to synthesize it. The Mattergen Paper was published in Nature recently and the source code for it can be found here on Github. In general, Microsoft's AI for Science Initiative (where this work has come from) is working on some crazy stuff and it's interesting to check them out.
Most Human Wins by NotBoring. This essay asks the question: In a world, where every human skill gets commoditized by AI, what is left? The answer: neti neti, neither this nor that. Every skill replaced by AI is not core to humans. The idea then: Move up a layer of abstraction and discover what is so unique to humans, that it—at least for a moment—is not replaceable. Then double down on that. In other words: Whatever is most human will win. And AI will make us find out what is most human. In the article, they link to a password-protected strategy pdf to thrive in the AI future. The password is rerecaptcha and the pdf is worth reading in full. My only grudge with the ideas proposed is that AI might be faster in moving up the layers of the stack than we can keep up with. This would leave us in the dust, unable to contribute anything meaningful anymore.
🌌 Travel 🌌
The last two weeks have been filled with many more kilometers of cycling up and down the mountains of Colombia. I was close to an active volcano, bathing in the hot springs near the Nevado del Ruiz, wondered about the beautiful paramo landscape with its ancient frailejones plants, and sweated and froze my ass off at different altitudes, while camping in the middle of nowhere on some farmer's field. All the while, the friendliness of Colombians made me happy and I made many friends along the way and look forward to continue cycling towards Ecuador. Next up will be the Tatacoa desert, but for now, I'm taking a bit of a rest to digest all the experiences. Writing, editing pictures, and putting things on my website and social media.
🎶 Song 🎶
Nightjar by Cosmo Sheldrake
I also love the song from the video on optimism, which I linked to at the beginning. It's called The Optimist Playbook and was created by Mandelbro and Whitworth. Find it on YouTube Music and Spotify.
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!