First, in 2022, Sam Bankman-Fried provided $6.5 million to the Center for AI Safety (CAIS), which accounted for nearly all of its $6.7 million of funding revenue that year. Open Philanthropy (backed by Dustin Moskovitz) later contributed over $10 million starting in late 2022.

Holden Karnofsky (Open Philanthropy co-founder/former CEO) is married to Anthropic president Daniela Amodei and joined Anthropic in 2025 to work on responsible scaling. Dustin Moskovitz was an early investor in Anthropic’s 2021 $124M round. Luke Muehlhauser (Open Phil AI officer) sat on Anthropic’s board until 2024. Strong shared EA networks and personnel overlap.

Additionally, in 2022, Connor Leahy co-founded Conjecture, a London-based startup dedicated to scaling AI alignment research. Sam Bankman-Fried participated as an early angel investor alongside other notable tech figures such as Andrej Karpathy, Nat Friedman, and Patrick and John Collison. Leahy served as Conjecture's CEO until March 2026. He subsequently shut down the company after concluding that technical AI alignment would not be solved in time to prevent catastrophic outcomes, shifting his focus to policy advocacy as the US Executive Director of the nonprofit ControlAI.

Connor Leahy has also been running around the podcast circuit and also has ties to Anthropic and Sam Bankman-Fried.

I think all of this AI doom speak is coming straight from Anthropic in order to achieve regulatory capture. Open source and open weight models pose too much of a threat.

(https://xoxohth.com/thread.php?thread_id=5922089&forum_id=2#50255340)

I thought he was 180 and 100% genuine at first, too, but I think he's just a very smooth grifter. Remember, his first startup was literally bankrolled by Sam Bankman-Fried. He also used to work for Anthropic. He's in all of the same EA circles. He's extremely clever, that's for sure.

I'll also point out that Connor does *not* understand LLMs. I just heard him speak on another podcast and he made two completely false statements: First, he claimed that more compute is spent on "reinforcement learning" than anything else. Second, he claimed that LLMs aren't suffering a training data shortage because they're making their own data.

I'll start with the first claim, since it's the easiest to debunk. Reinforcement learning has been around forever; it's just a way to update a model by maximizing the cumulative reward over a sequence of decisions. It's similar to online learning, with the exception that past predictions now actively change the distribution of future data. Connor is probably referring to "reasoning", which has nothing to do with reinforcement learning. "Reasoning" is the chore of breaking a complex, human-supplied task into its individual steps based on chain-of-thought reasoning and external tool calls. It's actually called retrieval augmented generation (RAG), which can be a step in an RL pipeline. Connor is very articulate, so there's no excuse for this fundamental mistake.

Debunking his second claim is also pretty easy, but it requires some work. Model collapse is well-studied (e.g., https://arxiv.org/abs/2305.17493 ) and pretty much guaranteed if you incestually train your models, e.g., use output from gpt-6.1 to train gpt-6.2. It's the machine learning version of mad cow disease. Connor actually debunks the claim himself by admitting that much of the recent gains are because AI companies are literally paying PhD-level researchers to interact with their LLMs. They don't really need to, because they already get most of that data for free because professors and grad students already interact with the models. Additionally, they're throwing even more compute at the models, so the quality continues to increase exponentially. But all exponential growth eventually stops. I think model collapse is perhaps the largest hurdle for AI to overcome.

(https://xoxohth.com/thread.php?thread_id=5922089&forum_id=2#50255947)

It's particularly telling that *every* AI safety "expert" has *deep* ties to people with extremely deep pockets. This is 100% about convincing normies to pester their elected officials to do the heavy lifting to prevent chink LLMs from cutting into Anthropic/OpenAI's profits.

I think AGI/ASI poses an existential threat. So do they. It's just that my idea of an existential threat is most of humanity either dying out or becoming enslaved. Connor's idea of an existential threat is directly tied to Anthropic's stock price. I mean, this would only be persuasive if Anthropic has an IPO imminently planned and only if literally every vocal AI safety advocate had extremely high-level connections to Anthropic/OpenAI.

(https://xoxohth.com/thread.php?thread_id=5922089&forum_id=2#50256019)

>In August 2017, Karnofsky and Daniela Amodei married in an effective-altruism-theme ceremony. Wedding guests were encouraged to donate to causes recommended by Karnofsky’s effective-altruism charity, GiveWell, and to read a 457-page tome by German philosopher Jürgen Habermas beforehand.

>“This is necessary context for understanding our wedding,” the couple wrote on a website for the event.

These people are sick.

(https://xoxohth.com/thread.php?thread_id=5922089&forum_id=2#50256690)