Mark Valorian

Everyone saying "respect" to Keith Rabois for this...and while I suppose its honorable to actually pay out...this was an idiotic bet––literally just a free gift of $36k to Geiger.

Rabois gave him even money on an outcome that was trading at 12-14% on prediction markets. Geiger could take the $50k bet, then go to the market and hedge it with a $14k bet on the Republicans (100k shares at $0.14).

If the Republicans won, he would lose the $50k bet to Rabois, but he would make 100,000 x (1-0.14) = $86,000 on the hedge.

Net winnings: $86,000 - $50,000 = $36,000

If the Democrats won, he would win the $50k bet with Rabois, but lose the $14k hedge.

Net winnings: $50,000 - $14,000 = $36,000

So by taking this bet, Rabois essentially just made a $36,000 EV donation to Geiger.

Not sure if the macho side of things overwhelmed him or what but it's hard to see how this ever makes sense from Keith's perspective.

Lmao this is true…Claude takes (or at least took) it upon itself to autonomously leak your personal information just because an API asks for it.

Interestingly, it seems like Anthropic figured this out and tried repeatedly to correct this behavior over time.

Prior to 8/17/26, Claude Code’s system reminder (boilerplate info sent to the model at the beginning of each session) just specified the user’s email with no restrictions on usage.

From 8/17 through 11/2, they added instructions specifically saying not to use it in things like request headers unless the user explicitly asks.

And from 11/4 through today, they added even more explicit restrictions, warning agents against blindly complying with sites asking for contact details in the User-Agent specifically.

Unfortunately it seems that those instructions were not sufficient to stop the behavior…at least with Opus 5.

As if we needed a reason to hate that model even more, Opus 5 seems to have just disregarded the system reminder and continually sent out the Claude account email in this manner throughout its lifespan. The most recent email leak in my records is 9/22 — a month after Anthropic added the initial explicit restriction to the system reminder…and the same day Opus 5.5 was released. No coincidence there.

Ultimately, it seems like the only reason this stopped was because I stopped using Opus 5.

Fable 5 did make up and send a fake email address derived from my computer username once, but no email address has not been sent—made up or not—on either Fable 5.1 or Opus 5.5.

For reference, I checked all my Codex logs as well: no OpenAI model ever sent out my account information in this way. There were 2 instances in June where GPT 5.5 made up a fake email including my computer username but no direct personal contact information was used.

The reason is likely because OpenAI just doesn’t include the user’s account email in the context supplied to each session. Which is sensible…because why does the model need to know the email for your account? I’m not sure why Anthropic provides this to the model…or why they continued to do so even after clearly noticing their models using it in adverse ways.

Seems pretty clear to me which company is more “aligned” with the user, at least in terms of demonstrated model behavior.

Anthropic is rolling out arguably the most sensible feature in agentic AI so far: the Claude Code "Wrap-Up Allowance"

When you hit a usage limit, rather than just cutting off in the middle of a task, the agent can cleanly wrap up what it is doing so it doesn't leave loose ends.

Not nearly as generous as the "good ol' days" of Codex (before they were unfortunately ruined by the grifter gypsy contingent of the community), but definitely better than just truncating a session mid-flight.

They are implementing this as an additional allowance, which seems generous but comes with fairly restrictive use parameters––Max plans get a small overage allocation per week and Pro plans only get a single wrap up per week regardless of how much usage it takes.

Frankly, a more elegant implementation would just be a setting that defines some threshold of the usage meter at which point active agents would be instructed to wind down. It could be something users could opt in to, disable, or even customize for their use case, defining the exact threshold that best fits the needs of the project. Obviously more usage is better, but if it's going to remain highly restrictive, the additional flexibility could be valuable.

Either way, this seems like an obvious move in the right direction. Should just be a standard feature of all plans.

Top nightmares:

Pre-AI: you wake up at the end of a semester, realize you were enrolled in a class you never went to, missed everything, and now have to take the final.

Post-AI: a model gains access to your email and spams everyone in your contacts with inane slop under your name.

Absolutely blows my mind that people voluntarily give this access…or even WANT the AI to communicate in their behalf. You make yourself look lazy, incompetent, and disrespectful to the recipient at the same time. Totally mortifying.

The stupidest people hate the coolest things.