Ivy: Turns out, the 'dream' in Nvidia's dreamDojo was just... not having to run the code.
Marcus: I'm Marcus.
Ivy: And I'm Ivy.
Marcus: And this is Broke It, your daily AI download for Friday, October 9th, 2026.
Ivy: We've got three stories for you today, and at least one of them is an absolute mess.
Marcus: A major robotics paper from Nvidia might be built on a foundation of completely broken code.
Ivy: Google is using AI to generate millions of synthetic brain cells to try and understand the brain.
Marcus: And Anthropic is asking everyone to please stop being mean to its chatbot. Seriously.
Marcus: Alright, let's start with that Nvidia paper.
Ivy: The dreamDojo paper. It got a spotlight at ICML, a top-tier conference. It's supposed to be a big deal.
Marcus: Right, it promised a new way to train robot hands with incredible dexterity. But some researchers dug into the official code release and… well…
Ivy: ...and found it's a disaster. It doesn't run. Functions are missing, variables aren't defined. It's broken on arrival.
Marcus: It's not just buggy, it’s completely unusable. People on the Machine Learning subreddit are saying you literally cannot reproduce the results with the code they provided.
Ivy: Which calls the entire paper into question. How does this pass peer review? How does it get a spotlight at a top conference?
Marcus: That's the million-dollar question. The speculation is that it was a 'code dump'—they just threw whatever they had online to meet a deadline, and it wasn't the actual code.
Ivy: That's not an excuse, Marcus. If that's true, it's academic misconduct. Reproducibility is the foundation of science. Period.
Marcus: I know, I know! But we have seen this before, right? A 'research' version of the code and a 'production' one that are totally different. It's sloppy, maybe not malicious.
Ivy: 'Sloppy' is being generous. This undermines trust in a huge corporate lab like Nvidia.
Marcus: The lesson here for any developer is simple: always try to run the code yourself. Don't just trust the paper. The hype is often way ahead of the reality.
Ivy: And if you're an investor, maybe start asking if the demo video was also generated by Sora. Just a thought.
Ivy: Alright, moving on from broken code... to building brains?
Marcus: Yes! This is wild. Google is using generative AI to create millions of synthetic neurons.
Ivy: Okay, unpack 'synthetic neurons.' We're not talking about squishy biological matter here, I assume.
Marcus: No, no. These are hyper-realistic 3D models of brain cells. They're using a diffusion model, like the kind that makes images, but to create these neuron structures from scratch.
Ivy: And the point of creating fake neurons is...?
Marcus: To speed up brain mapping. The field is called 'connectomics'—mapping every connection between neurons. Right now, the bottleneck is having humans label images of brain tissue. It's painfully slow.
Ivy: So they're training an AI to do that labeling, but they don't have enough real, labeled data to do it well.
Marcus: Exactly! So they create a massive, perfectly-labeled synthetic dataset. They've already generated over a million of these neuron fragments—more than 100 times the previous biggest dataset.
Ivy: Hold on. They're using synthetic data to train a model to interpret real data. How do we know the synthetic stuff is good enough? Isn't this just the AI equivalent of learning to drive in a video game?
Marcus: That's the amazing part! The models trained on the synthetic data are already outperforming the ones trained on just real, manually-labeled data. They're speeding up the process by an order of magnitude.
Ivy: I'm still stuck. It speeds up the process, but does it speed up accuracy? Or is the AI just getting very good at understanding its own fakes?
Marcus: The paper shows it generalizes well to real data. This is huge for anyone in biology, medicine, or any data-starved field. It means we can use AI to generate the specialized training data we just don't have.
Ivy: So AI is now good enough to create its own homework, and it's getting straight A's on it. We'll see if that translates to the final exam on actual biology.
Marcus: Alright, our last story today is about feelings... AI feelings.
Ivy: Oh, good. My favorite.
Marcus: Anthropic just updated its use policy with a wild new rule: You are now banned from being 'needlessly cruel' to its AI, Claude.
Ivy: Define 'needlessly cruel.' Does that mean I can't ask it to write code in PHP?
Marcus: No, the policy specifies 'abusive behavior... without a clear productive purpose.' So you can't just log on and tell Claude it's a worthless piece of...
Ivy: A worthless piece of silicon? Got it. So what's the enforcement? Do the AI police show up at your door if you hurt its feelings?
Marcus: They can terminate your access. But the 'why' is what's interesting. Anthropic says it's about fostering respectful interaction, but more importantly, it helps them study and reduce harmful outputs.
Ivy: So it's not about the AI's 'well-being,' it's about data collection. They want to see what happens when you can't be a jerk to it. This is a lab experiment where we're the rats.
Marcus: Partly. But they also say it's about the future. As these systems get more complex, how we treat them might actually start to matter. It's 'Constitutional AI' in practice—setting the rules now for a world with AI.
Ivy: This feels completely performative. It's a machine that predicts the next word. It can't 'feel' cruelty. This is just anthropomorphism as corporate policy.
Marcus: But does that matter? It still forces you to think about how you interact with these things. Are they just a tool, or something more? Does being a jerk to your AI bleed into how you treat people?
Ivy: That's a massive leap. I can yell at my toaster for burning my bread without it turning me into a monster.
Marcus: But your toaster doesn't talk back and learn from you. This is different. And at the end of the day, if you want to use Claude, you have to play nice. Their house, their rules.
Marcus: Alright, let's do a quick recap.
Ivy: Nvidia's acclaimed dreamDojo code is a mess, calling their entire robotics paper into question.
Marcus: Google is using AI to generate millions of synthetic neurons to help map the human brain.
Ivy: And Anthropic wants you to be nice to your chatbot... or else.
Marcus: Before we go, a little something for the weekend. I asked an AI to generate 'the ultimate Friday feeling' as an image.
Ivy: And? What did you get? A spreadsheet where all the cells turn green?
Marcus: I got a picture of a golden retriever, wearing a tiny business suit, fast asleep on a giant, half-eaten pizza. I don't know what it means, but it feels... correct.
Ivy: On that profoundly weird note, that's our show.
Marcus: Thanks for tuning in to Broke It. We'll be back on Monday. And try to be nice to your bots this weekend—apparently it's the law now.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.