Ep. 214: Musk v. OpenAI Round 2, Coinbase AI Layoffs, AI “Soft Nationalization" & xAI + SpaceX

A Saturday recording and a week too chaotic to skip

Paul and Mike open by joking about recording on Saturday morning because Paul is spending Monday doing a 100-hole golf marathon for the Orange Effect Foundation. The detour is light and human, but it sets up the mood: they weren’t about to skip a week because the AI news cycle has become too wild to pause.

Musk v. OpenAI gets weirder than anyone expected

The second week of the Oakland trial delivered what Paul basically calls a script no Hollywood writer would dare submit. The biggest reveals: Musk reportedly texted Greg Brockman days before trial about settlement, Tesla in 2017 explored pulling Sam Altman and even Demis Hassabis into a Tesla AI lab, and Siobhan Zilis appeared less like a neutral board member than an undeclared Musk conduit inside OpenAI.

Paul is especially floored by the evidence trail showing just how seriously Tesla once considered making Altman the face of a Tesla AI launch at NeurIPS. Then the Mira Murati deposition adds another layer: under questioning, she says Altman wasn’t always honest, undermined her as CTO, and played executives against one another — giving hard-edged confirmation to rumors OpenAI had long tried to wave off.

Coinbase lays off 700 people, but the real story is the future org chart

Brian Armstrong’s memo says Coinbase is cutting about 14% of staff while rebuilding around AI, and Paul thinks people are missing the point if they reduce this to “AI washing.” Yes, crypto markets may be weak and maybe Coinbase overhired, he says, but the deeper claims still stand: AI is compressing timelines, enabling non-technical teams to ship code, and making bloated org charts look indefensible.

The vision Armstrong lays out — no more than five layers below the CEO, “player-coach” leaders, and AI-native employees managing fleets of agents — feels to Paul like a real preview of where VC- and PE-backed companies are headed. His punchline is simple: even if the layoffs aren’t entirely caused by AI, the memo gives every executive permission to start redesigning their company as if AI changes everything.

The White House tests the waters on model review

The administration’s rumored executive-order discussions around reviewing AI models before release sounded at first like a huge shift, especially when Kevin Hassett compared it to FDA drug approval. Then the messaging changed almost immediately, with Susie Wiles and others emphasizing partnership over regulation, which Mike reads as a sign the White House is feeling for a middle ground and getting blowback in real time.

Paul’s concern isn’t that catastrophic-risk oversight is inherently wrong — he actually says some intervention is probably inevitable. It’s that in today’s political climate, there’s no reason to trust such a process to stay objective, especially when even more straightforward science is already getting politicized. He points to Dean Ball’s essay arguing for intermediary institutions that can sit between the state and the frontier labs as a way to avoid outright nationalization while still managing serious risk.

Their own survey shows people expect job losses — just not their own

Previewing SmarterX’s new State of AI for Business report, Mike highlights the stat that made them stop: 71% of respondents think AI will eliminate more jobs than it creates in the next three years, while only 13% expect net job creation. But when asked about their own role, only 21% say they’re seriously concerned.

Paul’s read is that this may reflect the sample as much as denial: these are AI-forward people who read newsletters, attend events, and actively try to stay ahead. So the vibe is less “nothing will happen to me” and more “I think I’m doing what I need to survive while others may not be.”

Anthropic, SpaceX, and the strange logic of compute politics

One of the week’s most surreal turns: Anthropic signed up to use all of SpaceX’s Colossus 1 compute capacity in Memphis, while Musk posted that xAI would be dissolved into SpaceX and rebranded as SpaceX AI. Mike notes how absurd that sounds given Musk was publicly calling Anthropic “evil” not long ago, and Paul basically laughs at how quickly a few billion dollars can make ideological fights disappear.

His theory is that this is a strategic concession: rather than trying to win the frontier model race head-on, Musk may be leaning into SpaceX’s infrastructure strength and becoming a compute provider. If that’s right, it’s less a betrayal of xAI than a pivot toward the part of the stack where he can actually dominate.

Recursive self-improvement stops sounding like sci-fi

Jack Clark’s essay lands hard because he’s not an outsider guessing — he’s an Anthropic co-founder watching the internal trend lines. His estimate that there’s a roughly 60% chance of end-to-end AI R&D by 2028 means models that don’t just assist researchers, but fully run the loop of designing, testing, and improving the next model generation.

Paul explains it in plain language: today, humans still set the agenda and manage the loop, but once AI can do that itself, model timelines could collapse from years to months, weeks, or even days. The image he leaves listeners with is memorable and appropriately dizzying: understanding this exponential is like staring at the stars — you can see it’s there, but you can’t actually comprehend the scale.

The labs want the services layer too

Anthropic and OpenAI each launched PE-backed enterprise deployment vehicles within hours of one another, and Paul immediately connects it to an old SaaS truth: for every dollar of software, there can be many more dollars in services. These ventures aren’t just about selling more model access; they’re about embedding inside portfolio companies and redesigning how work gets done.

He’s blunt about what that means. Private equity isn’t funding these efforts to add headcount — it’s trying to squeeze more output from fewer people, raise revenue per employee, and figure out what the AI-native org chart really looks like. Stripe’s new “forward deployed AI accelerator marketing” role fits the same pattern: the company wants people embedded with teams to transform workflows until AI becomes the default operating mode.