Meta is Winning, AlphaChip, NotebookLM, and More – Live and Learn #52

✨ Quote ✨

After thousands of years of compounding scientific discovery and technological progress, we have figured out how to melt sand, add some impurities, arrange it with astonishing precision at extraordinarily tiny scale into computer chips, run energy through it, and end up with systems capable of creating increasingly capable artificial intelligence.

– Sam Altman - (source)

✍️ Post ✍️

The Best Yellow.

I tested all the Yellow acrylics colors that I could find on Amazon to find the best one. The results were surprising and the color that I liked the most was not the one that I expected.

🖇️ Links 🖇️

Orion AR Glasses by Meta. The biggest announcement of the Meta Connect Keynote was the Orion AR Glasses. Even though they are "only" a prohibitively expensive and non-available research prototype they show where things are heading => Compute Devices you can wear on your head, all the time, that look just like normal glasses. The form factor is insane and it's interesting to see other people use them. You can check out the testing video by TheVerge as well as the official Meta Orion Announcement Video. There's also a great podcast with Marc Zuckerberg by Cleo Abram, where they talk more about all of this and all the other stuff that Meta is doing.

Meta Connect Keynote. Meta had their Meta Connect Event recently and it was packed with announcements and reveals. All the technology they are building is crazy and we are in for wild technological marvels, such as the Orion AR Glasses, the Quest 3S, and Llama 3.2. The video is worth watching in full, but you can also read the official Meta Blog Post with the announcements.

AlphaChip by Google Deepmind. Google has been working on using AI for chip design for a while now. Their original paper on this method was released already way back in 2021 but they made much more progress since then, adding an addendum to the paper recently and also open-sourcing the code. Essentially, with AlphaChip they have a reinforcement learning system capable of creating super-human chip designs. Yes. We have AI that can design chips to run itself, better than humans can now. Nvidia has similar systems that they use and to me, this signals the first steps of a recursive self-improvement loop. Where AI creates better AI until a point where the whole thing takes off in an intelligence explosion.

Agent Announcements at Dev Day by OpenAI. This video blew my mind. AI is good enough now that it can make real-time phone calls trying to achieve a certain objective. Let me repeat this. AI can try to achieve goals by talking on the phone to real humans. It's uncanny and I wonder when this will be good enough that 0-person companies will become a thing. One year or two maybe? I think we will see a future version of o1 that can spin up a bunch of "fake" human agents to call people, negotiate, buy and sell stuff in the real world, coordinate for freelance gigs, and in general make decisions to achieve the higher level goal of running "the company". The future is gonna be wild. For a list of full announcements of what happened at the Dev Day see the official OpenAI blog post.

MovieGen Video Model by Meta. Meta has been outdoing itself recently. They just dropped a text-to-video model that looks better than Sora by OpenAI. The quality is insane, there are almost no artifacts anymore and the video length looks promising too. You can edit videos with it and it even produces sound... Wtf. And they even published a proper 92 page paper on the research (unlike other organizations which only do technical reports these days). All of this is huge. You can even try the model online at their Meta AI playground. They are also aware of the impact that creating this sort of technology can have on content creators. The world of content creation is changing rapidly and I think we will soon see automatically generated video content that is on par with or even better than human-made content. This is scary but exciting.

NotebookLM by Google. Staying in the world of content creation, Google recently released a new tool: NotebookLM. You can put in PDFs and other documents, links to YT videos or websites, and then can ask your source material questions about the contents within. This alone would be an epic RAG application but it gets better. NotebookLM can produce 10-20 minute long podcasts from the source material, with two AI-generated voices that discuss the content. The voices are based on their SoundStorm Model and it's insane how lifelike they sound. It's just like listening to a podcast with two people discussing the stuff you pasted in. You should 100% try it with a website that you like, or maybe even with some stuff of your own. It's free. One of the best showcases of what this can do is the one from Simon Willison.

Fruit Fly Brain Map by FlyWire. Recently, a team of researchers has mapped the entire brain of a fruit fly. Doing this yields what is known as a connectome, an overview of each connection for every neuron. Connectomes are giant piles of data. Fruit Fly brains are still relatively simple with "only" 140.000 neurons, yet their connectome still consists of over 54.5 million connections. The complexity of brains, even simple ones, such as that of a fruitfly astonishes me every time that I stop to think about it. I already mentioned previous work of the FlyWire team in Newsletter 23. Back then, they "only" had a neuron-by-neuron map. Since then they have come a long way–a full connectome is a whole different level of work and I am curious to see what's next. Maybe even more complicated animal brains or sub-synapse detail? Let's see.

🌌 Traveling 🌌

I was hiking a lot over the last week along the South Western Atlantic coast of Portugal. There is a beautiful long distance trek there called the Fisherman's Trail or Rota Vicentina. I walked about 80km of it spread over 3 days. The landscapes were amazing and I clicked a lot of pictures–here is a selection.

🎶 Song 🎶

Nothing by Bruno Major

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