Ivy: A researcher just told Congress that within two years, a bad actor with a laptop and a PhD-level chemistry background could use AI to design a novel, targeted pathogen.
Marcus: And that is absolutely where we have to start today. I'm Marcus.
Ivy: And I'm Ivy.
Marcus: It's Saturday, September 12th, 2026, and you're listening to AI Overboard.
Ivy: On the show today, that line between science fiction and the daily news? It's basically gone.
Marcus: We're starting with that resignation: a senior researcher at a top AI lab just quit, issuing a stark public warning about AI-powered bioweapons.
Ivy: Then, over in the open-source world, a developer says they've finally trained a model to stop sounding like a chirpy, corporate AI assistant.
Marcus: And the battle over intellectual property is getting hotter. Some of the world's most brilliant mathematicians are escalating their feud with OpenAI.
Ivy: So, about that resignation. The AP is reporting that Doris Jones, a leading biosecurity researcher at Anthropic, has quit.
Marcus: And she didn't just leave quietly. She went straight to the press, warning that Anthropic and others aren't taking the risk of AI bioweapons seriously enough.
Ivy: Her claim is terrifyingly specific. She says their own internal red-teaming showed that current models, like Claude, can dramatically lower the barrier to creating biological threats.
Marcus: Lower the barrier how, exactly? Are we talking about it just spitting out a Wikipedia article on anthrax?
Ivy: No. We're talking about giving a researcher with some lab skills—but not world-class expertise—the equivalent of a team of brilliant consultants to help them design novel pathogens that evade existing treatments.
Marcus: Okay, that is... significantly worse. But Anthropic's whole thing is safety! They were founded on being the responsible AI company.
Ivy: And that's her whole point. She says their public commitment to safety doesn't match what's happening internally—that the pressure to compete with OpenAI and Google is forcing them to cut corners.
Marcus: So she's saying they're just lying?
Ivy: She's saying they're suffering from 'motivated reasoning.' They want to believe their safety measures work, so they don't test them rigorously enough. It's the classic case of grading your own homework.
Marcus: This is the core dilemma, isn't it? You can't know how to defend against these capabilities without... building them. You have to build the scary thing to learn how to stop the scary thing.
Ivy: And what happens when your 'defense' research becomes a tutorial for someone else's offense? That's the tightrope. Jones clearly believes they've fallen off.
Marcus: For everyone listening, this just poured a gallon of rocket fuel on the AI regulation debate. This isn't theoretical anymore.
Ivy: It means the people building the most powerful tools in history are openly disagreeing on whether they're actually building weapons of mass destruction. Sleep well.
Marcus: Okay, let's pivot from existential risk to something a little more... manageable: making AI sound less like a robot.
Ivy: A noble goal. What's the breakthrough this time?
Marcus: So, a user on the LocalLlama subreddit posted a fine-tune of a huge Qwen model. They call it 'Human-Like Chat,' and the whole point is to finally kill that 'As a large language model…' persona for good.
Ivy: And how did they accomplish this feat of personality engineering?
Marcus: With something called a LoRA. It's a clever way to tweak a massive model without retraining the whole thing. They just fed it a custom dataset of what they call 'natural' human chat and trained it to copy that style.
Ivy: Define 'natural human chat.' Is it just, like, using more slang and not capitalizing sentences?
Marcus: Basically! It's less formal, more direct. It won't start its answer with a five-sentence disclaimer about not being a doctor. It just answers the question. The examples are pretty convincing.
Ivy: So the secret to making AI more human is to make it less helpful and polite?
Marcus: Less obsequious. It's about removing the layers of corporate safety-ism that make these models so tedious to interact with. Getting to the point.
Ivy: I'll believe it when I see it. I suspect the 'humanlike' dataset is just a different kind of performance. Instead of a helpful assistant, you get a 'chronically online' redditor. Which might be worse.
Marcus: Hey, it's progress! The bigger picture is that the power of fine-tuning is now in everyone's hands. You don't have to accept the personality Big AI gives you; you can shape it yourself.
Ivy: And let's talk about a group that is really not happy with AI right now: mathematicians.
Marcus: Right, this isn't entirely new, but TechCrunch reports it's escalating. A new open letter signed by twenty-five very prominent mathematicians is directly challenging OpenAI and other labs.
Ivy: And when we say prominent, we mean Fields Medal winners, heads of university departments... this is not a fringe group. They argue that AI companies are threatening their field by training on published work without consent.
Marcus: Which is the same argument artists and writers have been making. But is math different? Aren't proofs just universal truths, not creative expressions?
Ivy: That's a huge oversimplification. The path to the proof is a creative act. They're worried models will ingest their unique, painstakingly crafted proofs and then spit out slightly altered versions, passing them off as new discoveries.
Marcus: Okay, but isn't that how human mathematicians learn? By studying the work of others and building on it?
Ivy: Humans are expected to cite their sources and show genuine understanding. These models are black boxes that launder attribution. It devalues the original work and pollutes the field with derivative, uncredited 'discoveries.' It's academic plagiarism at industrial scale.
Marcus: So what do they want? For OpenAI to stop training on math papers?
Ivy: They want recognition, attribution, and a seat at the table. They're proposing ways to license their work and ensure AI-generated math is clearly labeled. They are drawing a line.
Marcus: This fight over AI and intellectual property has reached the most abstract fields of human knowledge. The outcome here could set a precedent for every academic discipline.
Ivy: Exactly. If you can't protect a mathematical proof, what can you protect?
Marcus: Alright, let's try and wrap our heads around all that.
Ivy: A top AI safety researcher quit, warning that AI-designed bioweapons are a near-term threat.
Marcus: An open-source developer trained a model to talk more like a person and less like a corporate PR department.
Ivy: And two dozen of the world's top mathematicians are now accusing AI labs of threatening their life's work.
Marcus: Before we go, Ivy, speaking of mathematicians and slightly-off AI...
Ivy: You saw the screenshot of the AI theorem checker that was supposed to verify new proofs?
Marcus: Yes! A professor fed it a complex new geometry proof, and it replied: 'Looks good! Have you considered adding more emojis?'
Ivy: The future of mathematics is secure.
Marcus: That's our show for Saturday, September 12th. Thanks for tuning in.
Ivy: We'll be back tomorrow. Until then, maybe hold off on inventing any new plagues. I'm Ivy.
Marcus: And I'm Marcus. See you next time.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.