Brain2Music, Biosecurity Risks, and FlyWire: A Connectome of a Fruit Fly Brain ā Live and Learn #23
Welcome to this edition of Live and Learn. This time with a paper decoding neural stimuli into music, some resources that dive into the risks that AI and other recent advances will bring to the world of biosecurity, a complete connectome of a fruit fly brain, and more. As always, I hope you enjoy this edition of Live and Learn.
āØ Quote āØ
The health of the eye seems to demand a horizon. We are never tired, so long as we can see far enough.
ā Ralph Waldo Emerson
Links
AudioCraft by Meta. Meta open-sourced their work for generative audio synthesis. The code is available on GitHub and the results sound really good. What people are going to build based on this will be interesting because tools like this have the potential to change music production workflows forever. And this will make creating captivating music more accessible than ever, which means we will have more awesome music, which is wonderful.
FlyWire by Princeton. People built a complete neuron-by-neuron map for a fruit-fly brain. Even the brain of a simple organism like this is incredibly complicated and has millions of neural connections. In a community effort, they have now mapped all of them into a single connectome. Which is insane. The data is publicly available and people do all kinds of interesting things with it. There is also this Twitter thread keeping track of the stuff that has been going on.
Brain2Music by Google Research. Machine Learning Algorithms become better and better at decoding neural patterns into the stimuli that have originally caused them. Previously this was mostly image reconstruction from FMRI data, but this paper shows that the same techniques can be applied to decoding audio stimuli just as well. The accuracy of the decoded patterns is still somewhat low, but that algorithmic models can even capture anything from brain waves at all is magical.
Decoding Biosecurity Defense by Hummingbird Ventures. With the rise of AI we get better tools and methods for generating and manipulating DNA. This is wonderful but at the same time poses serious risks. Because it also becomes easier to create truly terrifying bioweapons. Think Covid but much more violent, designed to kill many more people. New AI-powered tools make creating such biological weapons and viral strains easier than ever before. And with companies offering to synthesize DNA for you, it is also easier than ever to get your hands on the materials necessary. This Notion document details the risks involved and how we could defend against bad actors before something really terrible happens. It also gives an overview of companies active in this space. If you find stuff like this interesting, there is also a presentation on YT and this Google Docs, which are worth checking out.
RT-2 by DeepMind. Google's robotics division is moving closer and closer to making true general-purpose robots a reality. The paper builds on their previous work with SayCan (based on the RT-1 architecture) and improves upon it even more. This leads to robots that can learn novel tasks on their own, without being explicitly shown. These robots also have basic reasoning capabilities just like ChatGPT, which they can use to solve novel tasks in ways they haven't seen before. All of this is pretty rad and I wonder when robots like this, will become reliable enough to be commercially successful.
š Midjourney š
š¶ Song š¶
The Great Gig in the Sky by Pink Floyd
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,
ā Rico