Jordan Hale: Okay, picture — no wait, I want to start differently. Did you see the Character.AI thing before all this got going? Like, that's actually where I found the thread.
Ryan Castillo: The valuation number?
Jordan Hale: The valuation number! Character.AI — the companion chatbot startup that Noam Shazeer co-founded with Daniel de Freitas in 2021 — hits a billion dollar valuation by March 2023. No revenue. Zero. And I thought, okay, this is already a wild story. But then Google turns around and pays two point seven billion dollars in August 2024 to essentially buy them back through a licensing deal. And now Shazeer just... leaves anyway.
Ryan Castillo: Right, but — that's the part that doesn't get enough attention. Google wasn't acquiring Character.AI's product. They were acquiring Shazeer. The deal was structured around getting him and de Freitas back inside Google DeepMind. Character.AI the company basically stayed separate.
Jordan Hale: So it was always an acqui-hire dressed up as a licensing deal.
Ryan Castillo: For Shazeer specifically — yeah. And he co-authored 'Attention Is All You Need.' 2017. That's the Transformer paper. That's the architecture everything runs on now.
Jordan Hale: Which makes the June 17th announcement — him leaving Google DeepMind, joining OpenAI as Lead for Architecture Research — feel almost like, I don't know, like the punchline was always coming?
Ryan Castillo: Two point seven billion for twenty-two months. That's the math.
Jordan Hale: And then one day later — June 18th — Dean Ball announces he's going to OpenAI to lead Strategic Futures. You know, I keep trying to convince myself those two things are unrelated and I just... can't.
Ryan Castillo: They're not. The timing is the tell. And there's an IPO coming — that's the frame for both of these moves.
Ryan Castillo: Let's back up to 2021 for a second. Because that's where this actually starts. Shazeer and de Freitas build Meena inside Google. It's good enough to ship. Google says no — safety concerns. And they just... leave.
Jordan Hale: Which, you know, I keep thinking about what that moment feels like. You've built something, you think it's ready, and the institution says 'not yet' — or maybe never. And these are not junior engineers. Shazeer co-wrote the paper that everything runs on.
Ryan Castillo: Right, but — and this is the part I want to pin down — is that a retention failure or a deployment failure? Those are different problems.
Jordan Hale: Wait, say that again.
Ryan Castillo: Retention failure means you didn't pay him enough. Deployment failure means you had the person, you had the work, and you couldn't give him the runway to actually use it. Google's problem was the second one. Meena doesn't get released. Character.AI gets founded. A billion dollar valuation with zero revenue by March 2023 — that number is pure market belief in what Shazeer can build when he's not being held back.
Jordan Hale: Okay but I want to push on that a little, because — I mean, is this actually a Google-specific failure or is this just... Shazeer? Like, the pattern I'm seeing is a person who leaves when he can't ship. That's not about any one institution. That could happen at OpenAI in three years.
Ryan Castillo: That's a fair concern. I don't have a clean answer. But the $2.7 billion acqui-hire in August 2024 — that's Google acknowledging the deployment failure with a check. They're not buying Character.AI's product. They're paying to get Shazeer back inside DeepMind, co-leading Gemini. That's the admission.
Jordan Hale: And it lasts less than two years.
Ryan Castillo: Twenty-two months. So the question isn't was it rational to pay $2.7 billion — maybe it was, given what OpenAI gains by having the co-author of the Transformer architecture running their next-generation model design. The question is whether Google's structure could ever actually let him operate the way he needs to. And the answer, empirically, is no. Twice.
Jordan Hale: Which brings you to OpenAI. And I genuinely don't know — like, is there something structurally different there, or are we just telling ourselves a story because he chose them?
Ryan Castillo: That's the open question. His title is Lead for Architecture Research — that's not a committee role, that's a mandate. Whether OpenAI's culture actually gives architects that kind of autonomy, or whether this is just the next holding pattern... we'll know when we see what ships.
Jordan Hale: Okay but before we even get to Shazeer's new role — can we sit with Dean Ball for a second? Because I keep running this scenario in my head. It's a Tuesday morning in Washington. Some congressional staffer is drafting AI liability language. They need input. They call Strategic Futures at OpenAI. And the person who picks up the phone... is the principal author of the Trump administration's AI Action Plan.
Ryan Castillo: That's not a scandal. That's how policy expertise works — it moves between government and industry constantly.
Jordan Hale: No, I know, I know — but Ball isn't at some generic tech company. He's inside the company most exposed to the rules he helped write. And he's reporting to Jason Kwon, the Chief Strategy Officer. Not an ethics board. Not a safety team. The Chief Strategy Officer.
Ryan Castillo: The optics are bad. I'll give you that.
Jordan Hale: It's not just optics though — like, that's what I'm pushing on. The International Center for Law and Economics put something out June 18th saying talent hoarding claims rest on assumptions, not evidence. Fine. But Ball isn't a hoarded engineer. He's a policy architect sitting inside the regulated entity.
Ryan Castillo: Right, but — the ICLE point isn't about Ball specifically. It's about whether any of this constitutes actual foreclosure. Regulatory capture requires more than proximity. You need to show the regulatory outcome bends.
Jordan Hale: Wei et al., 2024. 'How Do AI Companies Fine-Tune Policy?' Peer-reviewed. They interviewed 17 AI policy experts. Seventeen. And capture risk was flagged — not as hypothetical, as a documented pattern of industry actors gaining extensive influence in US AI governance. That's not an assumption.
Ryan Castillo: I've read it. The finding is influence, not capture. Those aren't the same thing.
Jordan Hale: Okay but — wait, actually — where's the line? Because Strategic Futures is a brand new team. Ball is leading it. It's explicitly focused on the relationship between AI labs, governments, and society. That's not a research unit. That's a positioning unit dressed up as a research unit.
Ryan Castillo: Or it's exactly what it says it is. Frontier AI risks, labor market effects — those are real policy questions OpenAI is genuinely exposed to. You want that analysis inside the building, not contracted out.
Jordan Hale: I don't fully disagree with that. I just think... you know, the IPO context changes everything. Silicon Canals, multiple outlets — they all framed the dual Shazeer-Ball hire as pre-IPO positioning. One foundational researcher, one former White House rule-writer, same quarter. That pairing is not accidental.
Ryan Castillo: No. It's not accidental. But I can't tell you whether that makes it capture or just — good institutional design ahead of a really consequential regulatory moment. And I don't think either of us can.
Ryan Castillo: But let me make the IPO case concrete, because I don't think we've actually done that. An S-1 is a document about risk. Every line is an answer to an investor question. And the two biggest investor questions for OpenAI right now are: can you hold the architectural lead, and can you survive the regulatory environment that's coming. Shazeer answers the first one. Ball answers the second. That's not two hires — that's one prospectus.
Jordan Hale: Written in people.
Ryan Castillo: Exactly. And Shazeer specifically — he's now lost to Google twice. He co-led Gemini as VP of Engineering. That's not a peripheral role. That's the flagship model family. OpenAI doesn't just gain an architect. Google loses the person who knew where the bodies are buried in Gemini's design.
Jordan Hale: Wait — is that actually how model development works? Like, one person's departure changes what a competitor can build?
Ryan Castillo: For foundational architects? Yes. Because it's not about code — the code is committed. It's about the intuitions that didn't make it in. The directions you don't pursue because someone who tried it three years ago says 'that path dead-ends.' Shazeer carries that. You cannot extract it from a Git repo.
Jordan Hale: That's — yeah, okay. That actually reframes what the $2.7 billion was buying. Google wasn't just keeping him away from competitors. They were keeping the institutional memory of where Gemini's ceilings are.
Ryan Castillo: And now OpenAI has it. Which — I mean, look, there's a piece of this that bothers me. Wilfley et al., 2026, arxiv paper on OpenAI's ethical AI discourse. They track how OpenAI uses words like 'safety,' 'ethics,' 'alignment' over time and the finding is... the language has evolved in ways that may signal rhetorical rather than substantive commitment. And you pair that with hiring the author of the AI Action Plan to lead a team called Strategic Futures — that's a branding operation as much as anything.
Jordan Hale: No way. They actually tracked the language drift?
Ryan Castillo: Peer-reviewed. The framing shifts. The operational behavior — actually, the paper's careful not to say the behavior shifts, just the discourse. Which is almost worse, right? Because you can't falsify a discourse move.
Jordan Hale: You know what that makes me think of — and tell me if this is too far afield — it's like, the Tuesday morning scenario I was running earlier? The congressional staffer? Add this layer. They're not just getting Ball on the phone. They're getting Ball, who is embedded in a company whose public safety language has been documented to drift toward the rhetorical. The staffer doesn't know that. The Wilfley paper isn't in their briefing document.
Ryan Castillo: That's the structural risk. Not that Ball is corrupt — I'm not saying that. It's that the information asymmetry is total. He knows what OpenAI needs from the rule. The staffer knows what the rule should theoretically do. Those are not the same conversation.
Jordan Hale: And he reports to the Chief Strategy Officer. Not — I mean, not a safety board, not an external advisory panel. Jason Kwon. Whose job is OpenAI's strategic position.
Ryan Castillo: The number that matters here is two. One technical hire, one policy hire, same quarter, planned IPO. You don't have to believe in bad faith to see what that structure is optimized for.
