What are the key takeaways from this Hard Fork episode?

Key insights include: **OpenAI Pentagon Damage Control:** Sam Altman publicly admitted the Pentagon deal announcement was rushed, sloppy, and opportunistic, then amended contract language to explicitly prohibit using OpenAI tools for domestic surveillance of US persons. However, government procurement experts note that releasing only "relevant portions" of the contract makes independent verification impossible, leaving employee and public trust largely unrestored despite the AMA and blog post efforts.; **AI Talent Leverage:** OpenAI's original technical staff — those with three-plus years building frontier models — retain significant leverage because they hold irreplaceable knowledge for training GPT-6 and beyond. Leaders cannot afford mass defections from this group. The resignation of post-training VP Max Schwartzer, who publicly cited respect for Anthropic's values upon departure, signals that contract amendments have not satisfied this critical employee segment.; **Anthropic's Dual Reality:** Anthropic simultaneously faces an existential Pentagon legal battle and extraordinary revenue growth, reaching a projected $20 billion annualized run rate in mid-2025, up from $1 billion at the start of the year — a 20x increase. The growth is driven primarily by Claude Code adoption in enterprise. The US State Department has already switched from Claude to GPT-4.1, a model now considered several generations behind current capabilities.

How long is this episode of Hard Fork?

This episode is 65 minutes long. SignalCast provides an AI-generated summary so you can get the key insights in about 3 minutes.

Hard Fork

OpenAI's Fog of War + Betting on Iran + Hard Fork Review of Slop

March 6, 2026

65 min episode · 3 min read

Episode

65 min

Read time

3 min

Topics

Investing, Fundraising & VC, Design & UX

AI-Generated Summary

Published Mar 7, 2026

Key Takeaways

✓OpenAI Pentagon Damage Control: Sam Altman publicly admitted the Pentagon deal announcement was rushed, sloppy, and opportunistic, then amended contract language to explicitly prohibit using OpenAI tools for domestic surveillance of US persons. However, government procurement experts note that releasing only "relevant portions" of the contract makes independent verification impossible, leaving employee and public trust largely unrestored despite the AMA and blog post efforts.
✓AI Talent Leverage: OpenAI's original technical staff — those with three-plus years building frontier models — retain significant leverage because they hold irreplaceable knowledge for training GPT-6 and beyond. Leaders cannot afford mass defections from this group. The resignation of post-training VP Max Schwartzer, who publicly cited respect for Anthropic's values upon departure, signals that contract amendments have not satisfied this critical employee segment.
✓Anthropic's Dual Reality: Anthropic simultaneously faces an existential Pentagon legal battle and extraordinary revenue growth, reaching a projected $20 billion annualized run rate in mid-2025, up from $1 billion at the start of the year — a 20x increase. The growth is driven primarily by Claude Code adoption in enterprise. The US State Department has already switched from Claude to GPT-4.1, a model now considered several generations behind current capabilities.
✓Prediction Market Insider Trading Risk: More than 150 accounts placed bets of at least $1,000 correctly predicting a US airstrike on Iran within one day of the strike occurring. Israel has already arrested individuals using classified military information to trade on Polymarket. The CFTC, the primary regulator for platforms like Kalshi, lacks sufficient enforcement staff to investigate the volume of potentially insider-informed trades occurring daily across these platforms.
✓Soft Nationalization Pattern: The current Pentagon pressure on AI companies — demanding model behavior changes, threatening Defense Production Act invocation, and pulling agencies away from non-compliant vendors — represents an early template for gradual government control of AI. This differs from World War II-style asset seizure; instead it operates through contract clauses, supply chain designations, and regulatory pressure, with full nationalization remaining a plausible longer-term outcome as model capabilities increase.

What It Covers

OpenAI navigates Pentagon contract fallout as VP of Research Max Schwartzer resigns and employees publicly condemn the deal. Prediction markets face scrutiny after 150+ accounts correctly bet on US strikes against Iran. Guest Arijeta Lajka reports that 40% of YouTube Kids recommended videos in a single 15-minute session were AI-generated slop.

Key Questions Answered

•OpenAI Pentagon Damage Control: Sam Altman publicly admitted the Pentagon deal announcement was rushed, sloppy, and opportunistic, then amended contract language to explicitly prohibit using OpenAI tools for domestic surveillance of US persons. However, government procurement experts note that releasing only "relevant portions" of the contract makes independent verification impossible, leaving employee and public trust largely unrestored despite the AMA and blog post efforts.
•AI Talent Leverage: OpenAI's original technical staff — those with three-plus years building frontier models — retain significant leverage because they hold irreplaceable knowledge for training GPT-6 and beyond. Leaders cannot afford mass defections from this group. The resignation of post-training VP Max Schwartzer, who publicly cited respect for Anthropic's values upon departure, signals that contract amendments have not satisfied this critical employee segment.
•Anthropic's Dual Reality: Anthropic simultaneously faces an existential Pentagon legal battle and extraordinary revenue growth, reaching a projected $20 billion annualized run rate in mid-2025, up from $1 billion at the start of the year — a 20x increase. The growth is driven primarily by Claude Code adoption in enterprise. The US State Department has already switched from Claude to GPT-4.1, a model now considered several generations behind current capabilities.
•Prediction Market Insider Trading Risk: More than 150 accounts placed bets of at least $1,000 correctly predicting a US airstrike on Iran within one day of the strike occurring. Israel has already arrested individuals using classified military information to trade on Polymarket. The CFTC, the primary regulator for platforms like Kalshi, lacks sufficient enforcement staff to investigate the volume of potentially insider-informed trades occurring daily across these platforms.
•Soft Nationalization Pattern: The current Pentagon pressure on AI companies — demanding model behavior changes, threatening Defense Production Act invocation, and pulling agencies away from non-compliant vendors — represents an early template for gradual government control of AI. This differs from World War II-style asset seizure; instead it operates through contract clauses, supply chain designations, and regulatory pressure, with full nationalization remaining a plausible longer-term outcome as model capabilities increase.
•AI Slop Saturation in Children's Content: Arijeta Lajka's New York Times investigation found that in a single 15-minute YouTube scrolling session starting from approved channels like Cocomelon, over 40% of recommended videos were AI-generated. YouTube's current policy only requires creators to label "realistic-looking" synthetic content, leaving animated AI slop unlabeled. No filtering tool exists for parents on YouTube Kids, and YouTube Shorts time-limit controls announced in January 2025 represent the only near-term mitigation option.

Notable Moment

Arijeta Lajka revealed that after she published her story linking to a YouTube channel showing a beloved children's character with its stomach cut open, YouTube only removed the video after she flagged it directly — despite the content being live and actively recommended to children beforehand, illustrating how reactive rather than proactive YouTube's moderation remains.

Know someone who'd find this useful?

Episode Transcript

Casey, something horrible has happened. What's that? My wife has fallen in love with an AI. Which AI? So for background, you know, my wife is not, like, a very, techy person. She's not an early adopter. She works in city government. She's like a regular adopter. Like, when it gets good, she's interested. But don't talk to her before then. Exactly. And it just got good enough for her to take an interest. And so yesterday, she discovered ClubCode and started using it to do some stuff at work. And now it's all she wants to talk about. She went, like, out at a party last night, and she's, like, I just can't stop thinking about my coding. She had a dream about vibe coding last night. Really? And maybe all of this is sort of some, like, karmic, you know, revenge for the the Bing Sydney episode, but I do feel like I'm getting a taste of my own medicine here. I mean, the to me, this seems like your dream come true. Like, what does any man want more than his wife taking an interest in his hobby? This is, like, could be one of the best things that's happened to you in a long time. It's true. It's true. I think we've always, benefited in our marriage from the fact that Are you speaking about our marriage or your marriage to your wife? My marriage. Most of my wife. That we're sort of, like, interested in different things, and we can kind of cross pollinate. And so I guess I wanna ask you as someone who's in a relationship with, someone who works in AI, like, how do you stop talking about it? Do you have, like, set hours where you're like, we're not gonna talk about AI for this next hour? Here's what I'll say. If I ever figure that out, you'll be the first to know. Okay. Thank you. In my house, the few respites we get from AI during the week, I would say, would involve, Friday night episodes of RuPaul's Drag Race. That's a good solid one hour of not talking about AI. Okay. And, yeah. Outside of that, we're we're really monitoring the situation, Kevin. We are fully locked in. I'm Kevin Roose, a tech columnist at the New York Times. I'm Casey Newton from Platformer. And this is Hard Fork. This week, OpenAI scrambles to contain the fallout from its deal with the Pentagon. Then how prediction markets have become one of the most controversial parts of The US attack on Iran. And finally, it's a hard fork review of Slap for children. They'll never guess what Peppa Pig is doing now. Well, Casey, we are now in week two of this incredible high stakes drama that's been playing out between the Pentagon and America's leading AI companies. There's been a lot going on. We now have more clarity on why the deal between Anthropic and the Pentagon fell apart. …

Get the full transcript (12,567 words) + summary by email — free

One-time email with the complete transcript and AI summary of this episode. No account needed.

One email, no spam. We’ll also show you what SignalCast does.

Browse all Hard Fork transcripts →

You just read a 3-minute summary of a 62-minute episode.

Get Hard Fork summarized like this every Monday — plus up to 2 more podcasts, free.

Pick Your Podcasts — Free

Books, tools, and gear mentioned in this episode

SignalCast may earn commission on purchases via these links.

Tools

Kalshi
“Israel has already arrested individuals using classified military information to trade on Polymarket. The CFTC, the primary regulator for platforms like Kalshi, lacks sufficient enforcement staff to investigate the volume of potentially insider-informed trades occurring daily across these platforms.”
YouTube Kids
by Google
“Guest Arijeta Lajka reports that 40% of YouTube Kids recommended videos in a single 15-minute session were AI-generated slop. No filtering tool exists for parents on YouTube Kids.”
Polymarket
“Israel has already arrested individuals using classified military information to trade on Polymarket. The CFTC, the primary regulator for platforms like Kalshi, lacks sufficient enforcement staff to investigate the volume of potentially insider-informed trades occurring daily across these platforms.”