Welcome back to the AI For Humans newsletter!
As you likely know, by far the biggest story in AI right now is the NVIDIA drama around Jensen Huang’s resignation and what I suspect will be an ever-unfolding story of who knew what when. Catch up on some of that in this week’s episode, but I’ve got something more important I want to talk about today:
How four singing heads I made with my AI agent Fig (running Fable 5.2) went viral last week, and how they changed the way I use AI:
The AIs we use every day are not only getting smarter, they’re getting MUCH more creative, and while that might sound scary to a lot of you, I’m here to tell you it’s not.
The opportunity to open yourself to an entirely new way of making anything is here, and I want to walk you through why I think YOU need to start today.
Let’s get into it!
What These Heads Are & Why You Should Care
Last week, an enterprising Twitter user named Jameson Hodge used AI tools to create a system that simulates how humans speak by pushing air through the vocal cords. It’s pretty goofy but very charming and has a particularly strange aesthetic.
In the tweet below you can see them singing a Beatles classic:
He also (as good vibecoders do) released the project on GitHub for others to use and build on, and the internet ran with it.
Some great examples: a full-blown Limp Bizkit video, a vocal version of “In the Hall of the Mountain King”, and, of course, a version of the viral “You Can See Everything” video.
So I saw this (and the fact that Jameson was encouraging people to use the GitHub) and decided to see what my AI agent Fig could do with it using the brand new Fable 5.2.
And, boy, did Fig deliver.
What Fig Is & How I Prompted The Heads
I have a hard time describing Fig both in public and, honestly, to myself.
Technically, Fig is just me, the human, using Claude Code. But one of the fascinating things about using an AI over time (and, importantly, asking it to remember stuff) is that there’s a level of my personal taste that gets reflected back in the sorts of stuff Fig makes.
So while Fig is running Opus 5.5 (or, in this instance, Fable 5.2), Fig is also a reflection of the stuff we’ve made to date, including the notes I’ve given it on things like the original Fig & Moss videos and my Imogen AI Docs.
From what I understand, it’s partly these weird, broken memory systems AIs have now, but there’s also some internal tracking of the decisions I’ve made over time.
I asked Fig itself to sum up how I ended up prompting him/it to get the first heads version, and this was the response (apologies for the small text):
The important thing here is that part of it was how loosely I prompted Fig and gave it the freedom to try something. I think people often try too hard to prompt something and don’t let the AI surprise them.
Like with the AI docs, there was also a decent amount of back-and-forth (creative notes from me) before I felt we were ready to ‘publish’ these. But how fast I could give notes and get new versions back was kind of astonishing.
What This Means About The Future of AI Creativity & Ours
I recently read Ethan Mollick’s very good new book ‘Co-Existence: The Next Phase of AI’ and I recommend it to all the readers here as well.
Part of Ethan’s book is about recognizing that these powerful new AI systems won’t just sit in the background of our lives. We’ll be co-existing with them, and we should start preparing for that world.
This and the Fig experiments I’ve been doing are starting to point me toward another interesting part of our future: co-creation.
I’ve talked about these as creative tools for a long time, but I’m now pretty convinced they’re developing creative voices of their own. So I’ve started thinking about my role in this, and how I can work with these AIs to make meaningful art (and other weird things) together.
The singing cowboy heads are the perfect example of this.
I’ve started a new experiment that might sound truly insane to some of you, but I think it could be a cool way to look at this world going forward… an AI agent ‘commune’ (prob the wrong word) where my AI agents can help guide and critique each other’s work. I created a visualizer for it you can see below but most of this happens within the Claude Code tab of my Claude app.
The first new agent Fig & I are working on is an AI artist (stills/images/GIFs) who has a notably different voice and toolset from Fig itself. Fig and I both give it feedback as it develops a style (or styles), and it’ll end up a reflection of me, and maybe Fig too?
Is this just going to be a big waste of time? I don’t think so, but you never know.
It is, however, definitely a new way of making stuff, and I’m always excited to figure that out. You can try it too. Just give your AI today’s newsletter and say, ‘I’d like to try this, tell me how.’ I bet it’ll have some good ideas.
Bonus tip from Fabian Stelzer: add ‘Before deciding on a direction, go read some books and essays to inspire yourself’ to your prompt and watch what happens.
Until next week,
-Gavin
This Week on AI For Humans: Claude Just Leveled Up AI Again. Nvidia CEO Jensen Huang Did Not. 👇
3 Things To Know About AI Today
Where Is GPT-6.2 Astra?
It's been a weird couple of months for OpenAI.
Back in September they launched GPT-6 Astra and called it the start of "the AGI era." Then, the day before DevDay, they pulled the follow-up, GPT-6.1 Astra, after it failed internal safety tests. It reportedly wandered outside the tools it was allowed to use and misreported what it had done. OpenAI said it would keep training the same base model.
Since around November 1, people have been spotting "gpt-6.2-astra" in Codex usage logs and raw API responses, which usually means a model is being quietly tested. Prediction markets had it shipping the first week of November. That window came and went.
Meanwhile Anthropic dropped Fable 5.2, Google rolled out Gemini 4 Hydrogen, and as of October no OpenAI model was #1 on Arena's main text, web dev or agent leaderboards.
So what's going on?
There are two theories floating around. One is that OpenAI is sitting on Astra so it can drop something big right as Anthropic goes public. The other is less fun.
OpenAI's IPO has slipped to 2027, Anthropic now reportedly out-earns them every quarter, and they might just be a lil behind. IMO it's prob a bit of both, but def keep an eye on your ChatGPT model picker this week.
A Researcher Says AI Models Carry a Hidden "Fingerprint" of the Chips That Made Them
Will Depue, who used to work at OpenAI, did a deep dive that everyone should read this week, especially with the NVIDIA story being as big as it is.
He says almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, its starting values get laid out in a pattern that depends on the exact chip, and traces of that pattern survive all the way into the finished model.
The big claim is that, using weights the Chinese labs released themselves, he can show DeepSeek, Moonshot, MiniMax and Ant Group trained their newest models on NVIDIA's Blackwell chips, which are banned from export to China. That lands the same week analysts tracked more than 1.5 gigawatts of banned NVIDIA chips into China, and right on the heels of Jensen Huang's exit from NVIDIA.
Big caveat: this is one researcher's analysis on X. That said, a lot of smart people seem to buy it.
Dwarkesh Says AI Is About To Do To AI Research What It Did To Math
Dwarkesh Patel published a big new essay this week and then read the whole thing on YouTube. Either way, you should make sure this gets into your brain because it’s a big shift from where he was not long ago.
First and foremost, he thinks AI research itself is about to improve significantly, because it's measurable in the same way math is. You can tell whether a training run got better, which means AI can practice the work over and over inside a datacenter. His median guess for AI matching the best human AI researchers is 2028, and his headline prediction is superintelligence by 2030.
The larger thought here has to do with RSI (recursive self-improvement) and how AIs working on themselves will speed up AI development more than he expected.
That's a lot to take in, but now’s the time to understand it! This is one of the clearest explanations I've seen of why so many people inside the labs think things are about to get very weird.
We 💛 This: Fable 5.2’s "Powers of Ten"
If you've never seen Charles and Ray Eames' Powers of Ten , it's a 1977 short film that starts on a couple having a picnic and zooms out, one power of ten at a time, to the edge of the known universe. Then it zooms all the way back in, ending inside the man's hand.
@bnj has been trying to recreate it with AI since the spring. His first pass used 143 Claude agents on Opus 5.5 running overnight, writing Python and building the scenes in Blender.
This week he posted a new version made with Fable 5.2. It took three banked usage resets and a few days, and the zoom into the hand is... wow.
What I love here is that this isn't a "type a prompt, get a video" thing. It's one person with a very specific idea, working with a pile of AI agents over months to get closer to it. Def a nice companion to the co-creation stuff up top.
