Brad Gerstner: Anthropic's $30B Ramp, Mythos Doomsday, OpenClaw Ankled, Iran War Ceasefire, Israel's Influence

The episode opens with a clip taken from later in the conversation. One panelist asks how many pull requests will actually be pushed to "the core structural Internet" in a hundred days and offers an over/under. The answer given is roughly 10,000, but "it's gonna be a meaningless thing." Another voice replies that stopping someone's browser history from being released to the world may be worth waiting a hundred days for. The exchange then turns into crude jokes.

The host says David Freeburg is out this week and that Brian Gerstner (as the transcript renders the name) is filling in as the "fifth bestie." David Sacks is back after missing the previous week, and Chamath Palihapitiya is also on the panel. Before the news segment there is a stretch of teasing about moderation and politics. One remark refers to being called a "Rokana apologist" for saying the other side had started "doing, like, some logical things."

The banter settles on an internet personality the panel calls Eli (also given as Elisha Long). One participant says they found him through Marc Andreessen's YouTube channel. Asked what is appealing about his approach, a panelist argues that it "actually works at every level of society": the more you want something, the less likely you are to get it, so the real message is to let go and live life. The panelist calls that detachment "really healthy for people." The section ends on a plain summary: build something, stop ruminating, "just do stuff."

The main story is that Anthropic is holding back its newest model, Mythos, on the grounds that it is too dangerous for broad access. Per the company's account as the host relays it, the model autonomously found thousands of vulnerabilities, including bugs in every major operating system and web browser. The examples cited are a 27-year-old vulnerability in OpenBSD, which is used in firewalls and critical infrastructure, and a 16-year-old FFmpeg bug that automated tools missed after 5,000,000 scans. The host also says bugs were found in the Linux kernel.

The show plays parts of a promotional video, which the host introduces with "Here's Dario." The clips make three claims. The model's cyber ability is a by-product of its coding ability. It is "by and large as good as a professional human at identifying bugs." And it can chain flaws together: individually minor vulnerabilities can be combined into exploits built from "three, four, sometimes five vulnerabilities" that together produce "some kind of very sophisticated end outcome."

But as a side effect of being good at code, it's also good at cyber.

The host addresses Brad as an investor in the company and asks whether withholding the model is virtue signaling or a response to something real, and whether giving it to a small group to find bugs before public release is a good move.

The first response argues that Anthropic "deserve[s] a ton of credit." In this view, the company could have released Mythos to pull further ahead of competitors, in the spirit of "move fast and break things." Instead, its own vulnerability testing showed the model would enable offensive hacking and could expose browsers, browser history and credit cards. The speaker stresses that Anthropic acted without the government holding its hand, because doing so is in the long-term interest of the company and the industry.

The concrete step described is Project Glasswing, an AI-driven cyber coalition of about 40 major companies, including Apple, Microsoft, Google, Amazon and JPMorgan. The stated goal is to spend a hundred days using advanced AI to find, fix and harden software vulnerabilities before hackers exploit them.

The same speaker frames this as crossing a threshold. Mythos, and Spud, which they describe as an upcoming OpenAI model and "the first black, well trained model at OpenAI" (as transcribed), mark "the beginning of what I would call AGI models": step-function jumps in intelligence that are "just too smart to be released immediately." They note that nothing requires a finished model to ship immediately to general availability. They predict OpenAI and Google will also adopt sandboxing and defensive alliances. Their view is that the episode shows industry and market forces can be trusted when they coordinate with government, and that Anthropic talked to the government without relying on "top down regulation."

Later, a panelist who is addressed as an investor in both companies adds context. The Frontier Model Forum, formed in 2023, is described as already coordinating Google, OpenAI and Anthropic on anti-distillation work. The panelist says they have pushed back on Anthropic before over possible regulatory capture but sees this case as different. Dario could have declared a threshold crossed and called for a government moratorium, as Elon did in 2023 when he called for a six-month pause over civilizational risk. Instead, he gathered 40 leading companies for a hundred days of sandboxing and hardening and kept moving forward. The panelist says the industry should be praised when it relies on the market "rather than running to the nanny state."

A second participant says Anthropic is very good at two things, "product releases" and "scaring people." They describe a pattern in which Anthropic pairs a new model or model card with a study showing the worst possible implications of the technology. Their main example is the blackmail study from roughly a year earlier.

They rolled out this blackmail study where supposedly the new model could blackmail users.

Their case against that study has two parts. First, they say the model was prompted "over 200 times" to get the desired result, which was "clearly reverse engineered" for headlines. Second, a year later many open source models match the capability of the Anthropic model that was studied, yet they say they have seen no examples of blackmail "in the wild." If the behavior were a likely outcome, they argue, it would have shown up by now. The speaker adds that they are "not saying it's not sincere."

On cyber, however, the same speaker says this case is "more on the legitimate side." Their reasoning is that better coding models find more bugs, which means more vulnerabilities, which means more ability to chain them into exploits. They expect a "onetime period of catching up" over roughly the next six months, during which AI could uncover bugs that have sat dormant for up to twenty years across many systems. That, they say, justifies a pre-release period in which software companies with existing code bases use the tool to patch their own systems.

They argue the window is limited. Chinese open source models such as Kimi K2 are described as about six months behind, and OpenAI is expected to release something similar within weeks. Their advice is that every company, IT department or CSO managing code bases should use the next few months to find dormant bugs and roll out patches. If that happens, they do not expect the "doomsday scenario" Anthropic is portraying, and they suggest the fear may help drive the right behavior. After that comes an arms race between AI for cyber offense and AI for cyber defense, which they describe as "a more normal sort of period."