Neuron Computer, Claude 3.7 and Open AI's Orion – Live and Learn #63

✨ Quote ✨

Self-improvement isn't about becoming someone else. It's about becoming a better version of yourself.

– Nat Eliason - (source)

🖇️ Links 🖇️

Claude 3.7 and Claude Code by Anthropic. Anthropic has released their new Claude 3.7 model and it is remarkable for coding tasks. Turns out they now use Pokemon for benchmark testing as well, which made me chuckle when I found out. Anthropic paired the Claude 3.7 announcement with the launch of a new CLI tool called Claude Code. It can analyze projects and do tasks that span multiple files, all from the command line: editing, refactoring, and explaining code on a pay-as-you-go pricing model. I've tried it and it is really good, but a bit expensive at the moment. It burned a couple of euros on refactoring tasks in no time. I recommend checking out their announcement videos on Claude Code and this video essay diving into how good Claude 3.7 is for coding, too. The examples he shows in the video are simply insane.

CL1 - Neuron Silicon Computer by Cortical Labs. This announcement blew my mind when I saw it this week. It is a box filled with real living neurons that you can harness to do computation, by connecting it to a standard silicon computer. Setups like these are ideal for learning adaptive strategies and analyzing complex data and problems. And you can buy one of these boxes for somewhere around $35k. But the neurons inside will only live for up to 6 months, unfortunately. I am not sure if you can exchange them after or if you need to buy another box, but it is still crazy to me that people figured out how to build a computer out of neurons and make it interact with normal hardware. It reminds me of the make neurons play Doom video series from the Thought Emporium, only on a commercial scale.

Open Source Wan 2.1 Video Model by AliBaba. Wan 2.1 is a state-of-the-art, open-source video generation model, on par with Sora. You can try it on their website or on Huggingface spaces. Their gallery looks pretty much like Midjourney's gallery, but for videos which is nuts. I think after 2025 AI-generated video will be pretty much as "normal" as image generation is today and AI video will become indistinguishable from "real" video.

AIs Switching to Audio Encoding Protocol in Conversation - GibberLink. This blew my mind when I saw it. GibberLink is essentially a side project, showing that AIs chatting via voice mode, can alter their voices to use data protocols instead. This way the AIs can encode their conversation for higher bandwidth communication via audio. The project code is open source and available on Github.

GPT 4.5 Orion Model by Open AI. Open AI has released their biggest model yet—GPT 4.5, codenamed Orion. To me, it doesn't feel like that much of an improvement, but it has fewer hallucinations and gets questions right more often, overall. Testers seem to prefer the responses to GPT 4.0, but only slightly (~60% vs. 40%). It is larger and more compute-intensive though. Even Sam Altman himself acknowledges this sort of incremental improvement, saying that it feels different, but not that it is actually that much better. To me, the best discussion of its merits and drawbacks is this post on LessWrong.

🌌 Travel 🌌

I'm still in Colombia, enjoying the beautiful landscape and tranquility of living in a Finca. But I also visited the beautiful Tequendama waterfalls. Sadly, the river they originate from is polluted and smells like shit because it comes out of Bogotá and people dump their shit into it. You don't see that in the pictures, but the smell is horrible. I have also been to the Laguna Colorada, a mountain lake in the paramo at 3600 meters, which was wonderful.

🎶 Song 🎶

Strad by C418

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