Ivy: A fifteen percent price hike... on something that already costs a fortune.
Marcus: I'm Marcus.
Ivy: And I'm Ivy.
Marcus: And this is AI Overclocked for Sunday, August 23rd, 2026. Here's what's on our radar.
Marcus: Get ready, because your AI hardware budget is about to explode, courtesy of Nvidia.
Ivy: Then, we'll look at the AI companies that are literally burning books to get training data.
Marcus: And a new AI 'teammate' from some DeepMind alumni claims it's already outperforming the giants.
Marcus: Alright, let's start with the giant in the room: Nvidia. They're jacking up prices again.
Ivy: 'Jacking them up' is putting it lightly. Reports are coming in about price hikes of over fifteen percent on new AI servers.
Marcus: Whoa. Okay, that's steep. We're talking about their next-gen chip systems, right?
Ivy: That's right. The official reason is 'soaring memory costs'—specifically, the high-bandwidth memory these new GPUs devour.
Marcus: I mean, that makes sense! Demand for HBM is through the roof. It's the key ingredient, so of course the price goes up.
Ivy: A key ingredient they basically have a monopoly on. A 15% jump on a server that can already cost a quarter-million dollars? That's... more than significant.
Marcus: A half-million, in some cases. It's a lot of money.
Ivy: It's a fortune. So, what does this mean for the person trying to actually build something?
Marcus: It means your grant money shrinks, your startup's runway gets shorter, and you're probably forced to rely on cloud providers instead of building your own cluster.
Ivy: Who, conveniently, also buy their chips from Nvidia. So the money flows back one way or another.
Marcus: Look, this is the price you pay for being on the absolute cutting edge. You want the most powerful models, you need the most powerful hardware.
Ivy: Or, it's the price of a market where one company holds all the cards. I'm just saying, it concentrates even more power with those who can afford a fifteen percent 'oops' in their budget.
Ivy: Alright, let's move from the cost of hardware to the cost of data. And I mean a... very literal, physical cost.
Marcus: Okay, you have my attention. What are we talking about?
Ivy: Groups are sounding the alarm with the FTC about a disturbing practice: destroying physical books to create AI training data.
Marcus: Wait—destroying them? As in, scanning them and then tossing them out?
Ivy: Worse. They call it 'destructive scanning.' They unbind the book, slice off the spine, and feed the loose pages through a high-speed scanner. The book is gone. It's just a pile of paper.
Marcus: Okay. That just feels wrong. But hasn't the Internet Archive been doing that for years?
Ivy: It has. And now the fear is that AI companies see this as a copyright workaround. Buy a physical copy, 'transform' it by destroying it, and claim fair use.
Marcus: From a purely technical standpoint, I get the efficiency. But we're burning the library to read the books. It's insane.
Ivy: That is exactly what's happening. And it forces a question: what cultural artifacts are we willing to literally sacrifice on the altar of innovation?
Marcus: But you can always buy another copy. And the text is preserved digitally, which is arguably more durable than paper.
Ivy: Tell that to an archivist. Or a historian. Or anyone who gets that a book is more than just text. This isn't progress; it's industrial-scale vandalism for profit.
Marcus: For our last story, there's a new player that thinks it can do science better than the big guys. A UK-based lab called Inherent.
Ivy: Another one. What's the claim this time?
Marcus: So, Inherent was started by ex-DeepMind folks, and they just released an AI agent called Faraday. Its job is to replicate scientific research papers.
Ivy: Okay, that's actually interesting. The replication crisis is a huge problem. An AI that can verify results would be genuinely useful.
Marcus: Exactly! And here's the kicker: they claim Faraday just beat agents from both Anthropic and OpenAI at doing exactly that.
Ivy: Hold on. 'Beat' them how? On what papers? How accurate are we talking? That's a massive claim to make against the two biggest labs in the world.
Marcus: The details are in their paper, but essentially, they took a set of recent machine learning papers and had the AIs try to reproduce the results from scratch, just from the text.
Ivy: Only machine learning papers? So it's an AI checking other AIs' homework.
Marcus: For now. But think bigger! What if you could point this at a cancer research paper and have it verify the findings in a week? Or try a million tiny variations on the experiment?
Ivy: That's the dream, Marcus. The reality could also be flooding science with AI-generated junk that looks real because another AI 'replicated' it. We need independent verification of Faraday itself before we celebrate.
Marcus: Always with the evidence, Ivy. But this is a big step! Replication is the foundation of science, and we're finally building tools to automate it.
Marcus: Alright, let's wrap it up. Your AI servers are about to get more than fifteen percent more expensive.
Ivy: We're shredding books as a potential 'fair use' loophole to train the next models.
Marcus: And a new lab on the block says its agent is already better at replicating science than the ones from OpenAI and Anthropic.
Marcus: Before we go, Ivy, I have to ask: if you had an AI that could perfectly replicate any research, what would you have it check first?
Ivy: Cold fusion. I want a definitive, final, 'yes' or 'no' so we can all finally move on.
Marcus: Going for the big one! I love it. I'd probably start smaller. Like, does putting batteries in the freezer actually do anything?
Marcus: That's our show for August 23rd! We'll be back tomorrow with more of the signal, and hopefully less of the noise.
Ivy: See you then. And Marcus, if you get that battery answer, let me know. For science.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.