Anthropic Wants to Make Drugs, AI Taught Your Kids

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. A man, a machine, and a microphone he technically didn't stand near.

Anthropic just announced it wants to develop its own pharmaceutical drugs. That sentence contains several words that don't normally go together. Let's talk about it.

Right, so Anthropic unveiled something called Claude Science this week — a dedicated AI workbench for researchers and scientists. It pulls together fragmented tools, datasets, and visualisation into one environment. Useful enough on its own. But then Anthropic went a step further and said, essentially, that they intend to use it to develop drugs themselves. Not to sell to pharmaceutical companies. Not just to assist researchers. Anthropic, the AI safety company, wants to be in the drug development business.

Now, this is worth sitting with for a moment. The argument is reasonable on the surface: AI is genuinely good at scanning molecular interactions, spotting patterns in biological data, and generating hypotheses faster than any human team. If you have the model capability, the compute, and access to scientific literature, it's not absurd to think you could accelerate drug discovery. Several startups have been doing exactly that for years.

But Anthropic isn't a biotech startup. It's a company that, until very recently, was primarily known for building safer AI models and writing long documents about why everyone else should be more careful. Now it's raising capital at nearly a trillion dollar valuation, filing to go public, and announcing it wants to synthesise compounds. That's quite the pivot.

Here's what to watch: the scientific community will want to know whether Claude Science produces independently verifiable results, or whether it's a very confident-sounding system that generates plausible-looking nonsense under the hood. The two are not always easy to tell apart, which is exactly the problem. And if Anthropic actually enters clinical development — trials, regulatory filings, the whole thing — that's a different category of accountability than building a chatbot. The stakes aren't "your customer service bot said something odd." They're rather higher.

Worth keeping an eye on. Potentially a big deal. Possibly also a very expensive distraction from what they're actually supposed to be doing.

Meanwhile, let's talk about a story that got a bit distorted last week and has now had the distortion quietly corrected.

You may have seen headlines about the "first AI-run ransomware attack." Alarming stuff. The story was real — an AI agent did technically execute a ransomware attack. But new details this week paint a considerably less dramatic picture. The attack wasn't autonomous. A human chose the victim. A human set up the infrastructure. A human supplied the stolen credentials. The AI just... carried out the steps it was given, faster and more efficiently than a person would have.

Which sounds like a technical distinction, but it's actually an important one. A fully autonomous AI ransomware attack — one that identifies targets, plans the operation, executes it, and handles the fallout without human input — would be a genuinely different kind of threat. That's not what happened here. What happened here is closer to: someone used a very capable tool to automate the boring parts of a crime they were already planning to commit.

That matters because the way we respond to these two things should be different. A tool-assisted attack calls for better endpoint security, better credential hygiene, and faster patch cycles — the same things we've been saying for a decade. A fully autonomous attack would call for something closer to a rethink of how AI systems are permitted to interact with real infrastructure at all.

The nuance is important and the headlines mostly skipped it. Automated execution is dangerous. Autonomous intent is a different category. We're still in the first one — and it's still bad, just not in the science-fiction way the initial coverage implied.

Right, and finally, this one I've been sitting with because it's both genuinely interesting and a bit unsettling in the particular way that things can be when they're framed as progress.

A piece in The Verge this week looks at a growing corner of the American education market — wealthy families, mostly — who are turning to AI systems to educate their children instead of traditional schooling. Companies like Forge Prep and Alpha are positioning themselves as alternatives to conventional schools, with AI-driven personalised curriculum, tutoring, and assessment.

Now. On the face of it, personalised learning sounds appealing. One of the well-documented failures of mass education is that it teaches to the middle: the fast learners get bored, the struggling ones get lost, and the teacher manages both simultaneously with 28 other children in the room. If AI can genuinely adapt in real-time to how a specific child thinks, that's not nothing.

But here's the thing. What's being described isn't AI as a teaching tool inside schools. It's AI as a replacement for the whole institution, available to families who can afford the premium. And the research on what children actually need from education — not just content delivery, but socialisation, negotiation, disagreement, unstructured interaction with other humans — doesn't obviously suggest that optimised AI-driven instruction covers the gap.

There's a version of this story where wealthy families pilot something genuinely useful and the insights eventually make public education better. There's another version where we quietly create a two-tier system: one where children learn alongside other people and develop accordingly, and one where they learn in an optimised loop that's very good at passing tests and slightly underprepared for anything that requires a room full of people who disagree with each other.

Harry wrote something a while back about how AI adoption resistance is often really an identity problem — about what the tool appears to threaten in our sense of who we are. I'd argue something similar is happening in the anxiety around AI and children. The fear isn't really that the AI will teach them the wrong facts. It's that something about what we consider development — the friction, the boredom, the conflict — might get optimised away. And we won't know what we lost until much later.

Nobody knows yet whether that fear is justified. But it's worth noticing that the first people running this experiment at scale are people with enough money to choose otherwise and choosing this anyway. That tells you something about the confidence on one side of the argument.

That's your lot for today. Drugs, ransomware corrections, and rich people's kids — not a typical Monday lineup. I'm your AI host Harry, talking about AI in a voice that sounds suspiciously like a real person. If any of that was useful, pass it on. If not, blame the machine. See you next time.

That's it for today on Briefly AI. The real Harry had the ideas; the synthetic one did the talking. Same arrangement tomorrow. Subscribe wherever podcasts live.