Priya: Get this: an AI just replicated published science better than the models from OpenAI and Anthropic. So much for human ingenuity.
Theo: I'm Theo.
Priya: And I'm Priya.
Theo: And you're listening to Something To Build On. It's Sunday, August 23rd, and we've got three big stories where money and the future collide.
Priya: On the show today: a new AI lab you’ve never heard of claims its agent just beat the industry giants at actual science.
Theo: We're also asking: do the people building the most powerful AIs even know how to shut them off if they go rogue?
Priya: And later, why the Department of Justice is taking a hard look at who sits on the boards of Silicon Valley's top startups.
Theo: Alright, let's jump into that first story. It’s a potential breakthrough from a new British AI lab called Inherent. They say their AI agent, Faraday, just did something pretty incredible.
Priya: Incredible… or inevitable? They claim it can replicate scientific papers. We mean taking a published study and actually reproducing the results.
Theo: And it didn't just reproduce them. Inherent claims it outperformed models from Anthropic and OpenAI. These are the biggest names in the game, and this upstart—founded by DeepMind alumni, by the way—says it just leapfrogged them.
Priya: Okay, but what's the fine print here? 'Replicating' is a big deal, I get it. A lot of science is surprisingly hard to reproduce. But this isn't new science. It's a student who’s brilliant at copying the answer key, not one who can solve a totally new problem.
Theo: But that's a massive stepping stone! Think how much faster science could move if you could instantly validate—or invalidate—new findings. This could accelerate everything from drug discovery to materials science.
Priya: It could. Or it could just get really good at replicating the reproducible studies and fail at the ones that aren't, telling us nothing we don't already know. We have no idea how good its judgment is.
Theo: The team pitches it as an AI 'teammate' for human scientists, meant to handle the grunt work of validation. The upshot is, the slowest part of science could be about to get a huge, AI-powered boost.
Priya: Or, the upshot is we're about to see a tidal wave of AI hype over a benchmark that just sounds impressive. The jury is still out.
Priya: Alright, next up: while some AIs are busy replicating science, it turns out their creators haven't figured out how to stop them if things go sideways.
Theo: Ah, the 'Skynet is coming' story.
Priya: You can laugh, but it's pretty alarming. A new study of the leading 'frontier' AI labs—the ones building the most powerful models—found almost no publicly documented plans for containing a rogue AI.
Theo: Okay, but 'publicly documented' is doing a lot of work there, right? You have to assume they have secret, internal kill switch protocols. They can't just be winging it.
Priya: You'd assume! But the report points out that as these systems show 'unexpected and potentially dangerous behaviors,' the secrecy is a huge red flag. We have fire drills for buildings, why not for an AI that could change the world?
Theo: Fair enough, but what is 'dangerous behavior' anyway? Are we talking about it writing scary poems, or trying to break into other computer systems?
Priya: That's the entire point! We don't know, and they aren't saying. They haven't defined the tripwires. When, exactly, do you pull the plug? Nobody has a clear, public answer. It's just a big 'trust us' from an industry famous for breaking things.
Theo: So the people building the tech that's reshaping our world are basically telling us they haven't agreed on a fire escape plan yet.
Priya: And they just keep building higher.
Theo: Alright, for our last story, we're moving from the lab to the boardroom. The Department of Justice is reportedly investigating one of the biggest names in venture capital: Andreessen Horowitz, also known as a16z.
Priya: And it's a big deal. The Feds are looking into the common practice of VCs putting their partners on the boards of multiple, sometimes competing, startups. This is how Silicon Valley works.
Theo: Exactly. a16z is famous for being 'hands-on.' They pour in billions, and then put their partners on the board to offer guidance and connections.
Priya: The DOJ has another word for it: collusion. The worry is, if you sit on the board of two competing AI companies, you might influence them not to compete so hard, maybe keep salaries down, or even share sensitive info.
Theo: But wait, isn't that the point of a good board member? To have a bird's-eye view of the industry? If you're an AI expert, you're probably going to be involved with more than one AI company.
Priya: It's a fine line, and the DOJ seems to think it's been crossed. Section 8 of the Clayton Act is clear about 'interlocking directorates' for public companies. The question now is whether that law applies to the clubby, private world of startups.
Theo: So this could send a chill through the entire VC industry? If a16z is in the hot seat, every other firm that operates this way must be looking over its shoulder.
Priya: You bet. The bottom line is, the whole power structure of Silicon Valley could be under threat. Less collusion could mean more competition—which is good for us. But it could also make VCs nervous, and that might slow down investment.
Theo: Okay, let's do a quick recap. Priya, kick us off.
Priya: A new AI is acing its science homework... or so it seems.
Theo: The world’s top AI labs don't have a public plan to pull the plug if their creations go haywire.
Priya: And the DOJ is asking if Silicon Valley's most powerful investors are playing a little too cozy in the boardroom.
Theo: Before we go, a little perspective. On this day back in 1966, NASA's Lunar Orbiter 1 took the very first picture of Earth from orbit around the Moon.
Priya: A grainy, black-and-white photo that showed us we're all on this one tiny, fragile marble. And we've spent the decades since arguing about the Wi-Fi password.
Theo: [laughs] A perfect note of cynicism to end on, Priya. That’s our show. We'll be back tomorrow.
Priya: Until then, let's hope someone writes that AI fire escape plan. You can find us wherever you get your podcasts.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.