Cellular Alchemy, Pi 0.5 and Vibe Marketing – Live and Learn #67

✨ Quote ✨

If we wait until we have certainty about the likelihood of the intelligence explosion, it will by then be too late to prepare. It’s too late to buy home insurance by the time you see smoke creeping under the kitchen door.

– William McAskill - (source)

🖇️ Links 🖇️

Pi 0.5 by Physical Intelligence. This is the second update from the Physical Intelligence company after their Pi 0 announcement last year. Their goal is to build robots that can generalize well across environments, and their new policy, Pi 0.5, is a big step in that direction. Pi 0.5 can generalize to novel environments and perform as good as a model that is trained only for that specific environment. And it only needs around 100 examples to do this. In the words of the authors: "Our recipe can attain effective generalization using relatively accessible amounts of mobile manipulation training data." This IS a big step forward for general-purpose robotics, and the video they show of their robot doing tasks in wildly different homes that it has never seen before is mindblowing. It still looks like a toy, but you can see that they are on the right track, and what they are building might help usher in an era of general-purpose robots that can be used anywhere.

Vibe Marketing by Greg Isenberg. In this video, Greg Isenberg, a startup founder and advisor, gives insight into how AI and automation are changing the landscape of marketing right now. To be honest, I am not the biggest fan of marketing, and much less of AI-driven slop. But this video is still insightful, even if only as an indicator of where things are headed. The amount of high-quality, A/B tested marketing strategies and funnels will explode because everybody will be able to run this sort of thing. What required a team of experts before is now doable with a few tools that automatically browse Reddit and Twitter, analyse users' wants and needs and then produce targeted marketing materials (copy, blog posts, websites, tweets, shorts and what not) to sell products to that specific audience, iteratively improving and keeping only what works.

The Full Stack of Society by Conrad Bastable. This essay is from 2019, but it's one of the best things I have read in a while. It's super long and has the quality of a WaitButWhy article. With this piece, Conrad Bastable tries to lay out a comprehensive framework for the strategies that nations have employed to create wealth for their societies. It's covering everything from Feudalism to Financialism and how all of the building blocks fall in place and explain the current situation of the world. I had multiple "ahhh this is how the world works" moments while reading this, and the amount of links I have left open from this essay is quite astonishing. It's just very thoroughly researched and connects many, many different, wide-ranging ideas, while linking to further reading constantly.

Cellular Alchemy by Zack Chiang. In this essay, Zack Chiang details how science is slowly beginning to unravel how cells differentiate. We know that one stem cell can turn into all the distinct forms of cells we know, because everything from kidney cells to brain cells has to originate from the same egg. Yet, we don't know how this works in detail yet. Overall, the process of finding out how cells differentiate at the moment reminds one more of alchemy than science—hence the title. At the moment, people are trying out lots of different things, seeing what works, instead of understanding how they work. But in the process, they build out the tools and assemble the data necessary to piece together a more complete understanding, much like how alchemy eventually gave rise to the more exacting study of chemistry.

The Urgency of Interpretability by Dario Amodei. With advances in AI accelerating, Dario Amodei urges to invest more work into interpretability research. He believes that it might solve the alignment problem and vastly increase our chances for a positive outcome. We can't stop the bus, but we might be able to steer it. The argument is simple: if we don't understand how AI models work on a fundamental level, it is hopeless to try to verify that they are doing exactly what we want them to do. They might only do as we tell them until they have enough power to pursue their own goals. Interpretability is the path forward for understanding what models really do and maybe even more importantly, why they do it. It is therefore our best chance to unravel what future AI models' goals and potentially hidden motives are. Hence, Amodei argues, interpretability is an integral part to ensure that AI will be to the benefit of humanity and not our downfall.

🌌 Travel 🌌

I've come back from a month of traveling around Bolivia to pick up my bicycle in Colombia. The plan is to cycle down South towards Peru over the coming two and a half months before returning to Berlin, and I am excited for this.

But I miss Bolivia's beautiful landscapes 😊⛰️ The Torotoro National Park, the Isla de Sol, and Lake Titicaca have blown my mind with their caves, canyons, and high mountain beaches, and I am sure there are many more places to discover in this beautiful country. So it's only farewell for now, and I hope to come back here in the future.

🎶 Song 🎶

Freedom by Jon Batiste

Youtube Music | 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