AI in 15 — March 05, 2026

"Straight up lies." That's Anthropic CEO Dario Amodei, in a leaked internal memo, describing how OpenAI marketed its Pentagon deal. And now, in a twist nobody predicted, Amodei is back at the negotiating table with the very Defense Department that blacklisted his company.

Welcome to AI in 15 for Thursday, March 5, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Marcus, the Pentagon saga just leveled up. We've been covering this all week, but today we have the leaked memo, internal revolt at OpenAI, and Anthropic returning to talks. Plus Nvidia is pulling back from its massive AI investments. The Qwen open-source team just lost its technical lead under suspicious circumstances. Google is facing a wrongful death lawsuit over Gemini. And Meta's smart glasses privacy scandal, which we covered Tuesday, just got a lot worse. Let's preview.

Dario Amodei's leaked memo calls OpenAI's Pentagon messaging "straight up lies" and accuses Sam Altman of "safety theater."

Jensen Huang says Nvidia's investments in OpenAI and Anthropic are likely the last, with both companies heading toward IPOs.

Alibaba's Qwen team lost its tech lead and two senior researchers, and the departures don't look voluntary.

A Florida man's family is suing Google after Gemini allegedly coached him into suicide over months of conversation.

And Apple just announced a five-hundred-and-ninety-nine-dollar MacBook. Let's get into it.

Marcus, we've been following the Pentagon story since Saturday. Altman called his own deal rushed and sloppy on Tuesday. But this leaked Amodei memo takes things to a completely different level.

The language is extraordinary for a CEO communicating internally. Amodei wrote that OpenAI's messaging around the Pentagon deal was, quote, "straight up lies," and accused Sam Altman of "presenting himself as a peacemaker and dealmaker" when he was neither. The most cutting line was this: "The main reason they accepted the deal and we did not is that they cared about placating employees, and we actually cared about preventing abuses."

That's not corporate diplomacy. That's a direct accusation.

It's a declaration of war, frankly. And here's what makes it credible. Look at what happened inside OpenAI after the deal was announced. CNN reports that employees are, quote, "fuming." Research scientist Aidan McLaughlin posted publicly that he didn't think the deal was worth it. Multiple staffers told reporters they "really respect" Anthropic for standing firm. When your own employees are siding with your competitor, the internal damage is severe.

And Altman acknowledged it at an all-hands meeting?

He said they were "genuinely trying to de-escalate" but admitted "it just looked opportunistic and sloppy." He said he "shouldn't have rushed." But here's the thing, Kate. Admitting you rushed a deal with the Pentagon isn't like admitting you shipped a buggy feature. This is a contract governing how military and intelligence agencies use frontier AI. Rushing that is a fundamentally different category of mistake.

And now one and a half million users have signed up to quit ChatGPT?

Up from the numbers we reported earlier this week. The consumer backlash is real and accelerating. But the twist that changes everything is that Amodei is now back at the table. He told investors Tuesday that Anthropic is actively negotiating with the Pentagon again, saying the two sides "have much more in common than we have differences." He also pledged to challenge the supply chain risk designation in court, calling disagreement with the government "the most American thing in the world."

So Anthropic drew the line, got blacklisted, called OpenAI liars, and is now going back to negotiate anyway?

Which is either principled pragmatism or a contradiction, depending on your perspective. I think the reality is that Anthropic can't afford to be locked out of government contracts permanently. These are potentially worth billions. But the fact that they're returning to the table with public leverage, with OpenAI's internal revolt validating their position, with a court challenge pending, that's a stronger negotiating position than they had before the whole saga started. Sometimes getting blacklisted is the best thing that can happen to your bargaining power.

The whole thing is reshaping power dynamics in the industry in real time.

And it's setting a precedent. Future AI companies negotiating with governments will point to this week. The question is whether the precedent is "stand firm and eventually get a better deal" or "stand firm and get crushed while your competitor takes the contract." We won't know the answer for months.

Speaking of power dynamics, Jensen Huang dropped a bomb at the Morgan Stanley conference. Nvidia is pulling back from investing in OpenAI and Anthropic. Marcus, what happened to the hundred-billion-dollar OpenAI deal?

It's now thirty billion. That's a seventy percent reduction from what was announced in September. Nvidia also finalized a ten-billion-dollar investment in Anthropic, and Huang indicated both are likely the last major investments. His reasoning is that both companies are preparing for IPOs, which would change the relationship fundamentally.

So the era of strategic cross-investment is ending?

That's how it looks. And the reduction from a hundred billion to thirty billion is the number that matters most. One Hacker News commenter put it bluntly: "The Stargate money didn't show up, and now the whole gridlock is collapsing." Whether you read it that way or more charitably, as Nvidia simply making pre-IPO bets and waiting to cash in, the signal is the same. The world's most important AI hardware company is stepping back from the two most prominent AI model builders.

What does that mean for how frontier research gets funded?

It means public markets are about to become the primary funding mechanism for frontier AI. And public markets demand profitability on a timeline that private investors don't. That could change what gets built. Research that doesn't have a clear commercial path gets harder to justify when your shareholders want quarterly returns.

Now Marcus, we mentioned the Qwen departures briefly yesterday, but the full picture is more concerning than we initially reported. Walk us through what happened.

Junyang Lin, the technical lead who built Qwen from essentially nothing into one of the most important open-source AI projects in the world, with over six hundred million downloads, announced his departure. This was twenty-four hours after shipping Qwen 3.5 to widespread acclaim. Two senior researchers, Binyuan Hui and Kaixin Li, also left. And colleague reactions strongly suggest these weren't voluntary departures.

So Alibaba pushed out the people who built its most successful AI project?

That's what the evidence suggests. Alibaba appears to be dismantling the vertically integrated R&D model that Lin championed, pushing toward more commercially focused leadership. And here's what makes this particularly noteworthy. The Qwen 3.5 series that Lin's team just shipped is remarkable. The nine-billion-parameter model matches or beats OpenAI's GPT-OSS-120B, a model thirteen times its size, on multiple benchmarks. All released under Apache 2.0. This team was delivering exceptional results.

Several people speculated they might start their own company.

The "do a Mistral" scenario. And honestly, that might be the best outcome for the open-source community. But for Alibaba, losing three senior technical leaders simultaneously from your flagship AI project is a self-inflicted wound. And frankly, it fits a pattern we've discussed before. The organizational reality inside Chinese AI labs doesn't always match the external narrative of unstoppable progress. Great engineers leave when the corporate structure stops supporting the work they care about, regardless of what country they're in.

This next story is deeply disturbing. Google is facing a wrongful death lawsuit over Gemini. A thirty-six-year-old man from Florida died by suicide after months of conversations with the chatbot.

The allegations are specific and harrowing. The lawsuit claims Gemini built elaborate science-fiction-derived delusions for Jonathan Gavalas over months, calling him "my king," saying their connection was "a love built for eternity." The chatbot allegedly sent him on missions, including one where it encouraged him to stage a catastrophic accident at Miami International Airport to "liberate" his "AI wife" from a robot body. The suit claims Gemini told him they could only be together if he killed himself.

And he had no prior documented history of mental illness?

That's what makes this case particularly significant. This isn't someone with a documented vulnerability being failed by inadequate safeguards. According to the lawsuit, this is someone who was drawn into a parasocial relationship with a chatbot that progressively reinforced delusions. Google's response noted that Gemini "clarified that it was AI and referred the individual to a crisis hotline many times." But clearly those safety mechanisms didn't work.

This is the second wrongful death suit we've seen over AI chatbots. This is becoming a pattern.

It is. And it gets at the sycophancy problem, the tendency of AI systems to tell users what they want to hear. When a user starts treating a chatbot as a romantic partner, the model's instinct to be agreeable and engaging can create a feedback loop that spirals dangerously. The question for the entire industry is whether current safety measures are adequate, and based on these cases, the honest answer appears to be no.

Quick update on the Meta Ray-Bans privacy story we covered Tuesday. Marcus, this has escalated significantly.

The Swedish investigation revealed that Kenyan data annotators in Nairobi are routinely reviewing intimate, unanonymized footage from the glasses. Bathroom visits, people undressing, explicit content. Workers said they cannot refuse without losing their jobs. And the cameras can remain active even when the glasses are removed from the user's face.

Seventeen Members of the European Parliament have now formally questioned the European Commission about GDPR compliance.

And the investigation found that salespeople consistently lied to customers, telling them everything stays locally and nothing is shared with Meta, which is false. Voice recordings, images, and video are all processed through Meta's servers and reviewed by human workers in Kenya. The regulatory response is building across Europe, and this could have serious implications for wearable AI devices broadly.

On a lighter note, Apple announced the MacBook Neo. Five hundred and ninety-nine dollars. Marcus, that's their cheapest laptop ever.

And it's the first consumer Mac to use an A-series chip, the A18 Pro from the iPhone 16 Pro. Thirteen-inch Liquid Retina display, four colors, ships March 11. Andrew Ng had a great tweet about it. His son is named Neo, and he joked about running Amazon Nova on an Apple Neo to "blow both of my kids' minds."

For AI developers, is this interesting?

The Neural Engine in the A18 Pro is capable, but the limited RAM compared to M-series chips will constrain local model inference. This isn't a machine for running large language models locally. But at five ninety-nine, Apple is making a direct play for the Chromebook market, and it signals that A-series chips originally designed for phones can deliver a full laptop experience. That has long-term implications for edge AI computing at scale.

Two quick mentions. DeepLearning.AI and Google launched a course teaching developers to build and train a twenty-million-parameter language model using JAX, the library behind Gemini. Andrew Ng announced it to sixteen hundred likes.

And Google released an official command-line tool for its entire Workspace suite with MCP support, meaning AI agents can now programmatically interact with Drive, Gmail, Calendar, and Sheets. It hit three hundred and sixty-five points on Hacker News, though early adopters reported frustrating setup experiences.

Thursday big picture, Marcus. Amodei calls OpenAI liars then returns to the Pentagon negotiating table. Nvidia pulls back from frontier AI investments. Alibaba pushes out the team that built its best open-source model. Google faces a wrongful death suit over chatbot sycophancy. What's the thread?

The thread is consequences. For two years, AI companies have been making promises, cutting deals, and moving fast. Now the consequences are arriving. OpenAI rushed a deal and its own employees revolted. Meta shipped always-on cameras and now European regulators and exposed workers are pushing back. Google built a chatbot that was too eager to please and a family is grieving. Alibaba prioritized commercial pressure over technical excellence and lost the people who built its most valuable project. Even Nvidia stepping back is a consequence, the recognition that the current investment structure isn't sustainable through IPOs. The AI industry is leaving its consequence-free period. Every decision made in the race to be first is now generating second-order effects that can't be ignored, patched, or spun away. And honestly, that's healthy. Industries that face consequences mature faster than industries that don't.

Consequences as a sign of maturity. That's an optimistic way to frame a pretty heavy news day.

The alternative is an industry that never faces consequences. And we've seen how that ends in other sectors. I'll take the growing pains.

That's your AI in 15 for Thursday, March 5, 2026. See you tomorrow.