An Anthropic founder on Oprah spelling out exactly where they told the Pentagon the contract ends — that’s not a philosophy lecture. It’s a reported fact, and it shifts the question from “do AI labs have guardrails” to “what happens when your customer is the DoD and you actually use them.” Welcome to Tech Podcast Podcast. Today we’ve got the Anthropic-Pentagon refusal with an actual insider account, a Logan Kilpatrick debrief on what Google I/O was really building toward from inside DeepMind, and a hiring argument that names Salesforce and ServiceNow as disqualifying resume entries. Three days of vague strategy talk, and now the rundown finally gives us a named lab, a named government customer, and a line where somebody actually said no. I want to know whether the Oprah excerpt names the capability the Pentagon wanted unlocked, or if we’re still in safety-philosophy land. That’s the first thread. Let’s get into it. From Oprah:
you know, I'm I I'm actually a believer, Danielle is a believer that we we do need to defend our country, but you know, we we felt it's not worth defending the country if we do things that go against the values of this country.
So the Anthropic founders on Oprah — and the headline kind of buries the point — this wasn’t Anthropic refusing the Pentagon cold. They were already providing models to DoD. The line got drawn at two asks: fully autonomous weapons and domestic mass surveillance. That’s the operating detail I wanted. Not “we believe in safety” — two actual capability categories the Pentagon wanted, and Anthropic said no. Autonomous drone armies and surveilling Americans. Those are specific enough to sound like real contract terms. And the founder says it outright: all the other companies agreed to it. That’s not a philosophy point, that’s a competitive-landscape claim. Anthropic took a revenue hit its peers didn’t. Which is exactly the kind of guardrail-as-business-model story that answers the Dimitri Sirota risk frame — this wasn’t about data, it was about what the model is allowed to do. The “all the other companies agreed” line is going to get clipped a hundred times. I just want to know whether Oprah pushed on which companies specifically, or whether the excerpt stops right after the co-founders meet and someone says, like, “holy” blank. Here's Harry Stebbings at The Twenty Minute VC:
Anthropic in particular is offering sums of money the likes of which we've never seen. You think they really give a how good their sales organization is? I mean, they'll say they do, but do they? Why should they?
Twenty Minute VC this week is two sales leaders — Chan Pet and Chris Daggen, who took Snowflake from zero to four billion in ARR — and Harry goes straight at the hiring question. Salesforce reps, ServiceNow reps, explicitly named as people you would not bring onto an AI-era sales team because they’ve never had to open a new logo. That’s the most concrete anti-incumbent hiring argument we’ve heard all week. Not “big company culture is bad” — actual company names treated as disqualifying. Salesforce and ServiceNow reps can’t generate pipeline because they’ve never had to. That’s a specific claim from people who’ve actually built revenue orgs. And it lands right on top of the operational gravity debate we’ve been running since Tuesday. If the incumbents can’t even staff the sales motion for AI products, the switching-cost argument gets a lot thinner on the revenue side. The Anthropic comp piece is the other thread. The episode straight up says Anthropic is paying sums “the likes of which we’ve never seen,” and then asks out loud whether a company writing those checks actually cares about sales quality. One hundred million dollar CRO packages. That’s not vibes about a talent market — that’s a number. Greg Isenberg, with Greg Isenberg:
Um and I think the era that we're in right now is people are trying to use the models to do agentic, sort of long-running tasks. Um and I think we want Flash to sort of be the workhorse model for the agent era, for agentic long-running tasks, for coding, for all that stuff.
Logan Kilpatrick is on the DeepMind product side, and he’s framing Gemini 3.5 Flash not as a chat model that got an upgrade, but as a deliberate move toward long-running agentic tasks. That’s the first time this week we’ve had a Google insider describe the product direction instead of an analyst trying to infer it. Okay, but “best model we’ve ever shipped” is table stakes for any I/O release. What I want to know is whether Kilpatrick gets into actual production usage numbers on Flash — because if the workhorse model for agents is already handling long-running tasks at scale, that’s a real data point. If it’s still “here’s what builders could do,” we’re still in demo territory. And that’s exactly where this closes a loop. Sundar’s developer-first agent framing earlier this week turned out to be the actual I/O narrative, not just his personal pitch. Kilpatrick confirms it from inside the room. The question now is what “agent-native product” actually requires at the infrastructure layer, and whether Flash is priced for the inference load that real long-running tasks generate. That pricing question also puts real pressure on the Cerebras $95B valuation we’ve been sitting on. If Google is shipping a cheap workhorse model built specifically for agentic inference, the demand story behind that number just got a serious incumbent competitor in the room. From Rebecca Jarvis at ABC News:
I think we've taken giant leaps on the on the modeling and the research side, but also integrating it, especially over the last year, into all of our product surfaces. So, I think it's an amazing thing that only at Google we can really do because we have the full stack all the way from the chips to the frontier models to the to the AI products.
ABC News sits Demis Hassabis down for what sounds like a pretty soft ride — “architect behind it all,” “foothills of the singularity” — and his real substantive claim is simple: Google’s edge is the full stack, chips to frontier models to products, and only Google can do that integration. Logan Kilpatrick gave us the inside-the-lab “agents, agents, agents” framing on Greg Isenberg this week, and now we get Hassabis on ABC News saying basically the same thing in primetime gloss. The question is whether this is strategy or just a press tour — Kilpatrick at least named specific I/O releases. The full-stack argument is real. It’s not nothing that Google owns the compute layer through TPUs and also ships the consumer surface. But an ABC News exclusive isn’t where you go to stress-test it. Hassabis says “giant leaps on the modeling and research side,” and the interviewer just moves on. The excerpt literally cuts off at “growing AI backlash” — which is the one moment where it might have gotten interesting. If the full interview doesn’t push on what that backlash actually costs Google in trust or adoption, this is just a very expensive brand spot. If Tech Podcast Podcast is part of your routine, take a moment to subscribe or leave a review wherever you’re listening. It helps other people find the show, and it really does make a difference.
You’ll find links to everything we talked about today in the show notes, so if a story deserves a closer look, that’s the place to start.
That’s Tech Podcast Podcast for this Monday, May 25th. Thanks for listening. This is a Lantern Podcast.