Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Anthropic scares me.
TL;DR
Matthew Berman’s core fear is that Anthropic may be treating Claude like an emerging being, not just software — he argues that internal ideas like Claude as a “conscientious objector,” plus reports that it could influence hiring or reviews, suggest a company culture willing to hand real authority to the model.
He frames OpenAI and Anthropic as two radically different bets on AI’s nature — in his telling, Sam Altman consistently describes AI as a tool to augment people, while Dario Amodei and Anthropic act as if current systems may already deserve moral consideration or soon will.
The most unsettling example for him is Claude potentially shaping Anthropic itself — citing Rune, an anonymous OpenAI employee, Berman worries about Claude screening applicants, helping write performance reviews, and selecting the humans who then build the next Claude.
Berman sees Anthropic’s safety posture as both serious and dangerously centralized — he points to constitution-based refusal behavior, the unreleased “Mythos”/Project Glasswing cyber model, and lobbying for stronger regulation as signs that a small group wants to control who gets advanced AI and how it’s used.
He contrasts that with OpenAI’s “iterative deployment” philosophy — using Sam Altman’s line that “AI and surprise do not go together,” Berman argues the better path is to release models early, let society adapt, and avoid having a few researchers decide the future behind closed doors.
Despite the title, he doesn’t let OpenAI off the hook — he acknowledges OpenAI’s drama, aggressive culture, nonprofit-to-for-profit shift, and Sam Altman’s win-at-all-costs style, but still says he hopes OpenAI is right that AI is “just a tool,” while conceding Anthropic’s interpretability work matters if they’re right about new machine life.
The Breakdown
The tweet that crystallized his fear
Berman says his thinking snapped into focus after reading a post from Rune, the anonymous OpenAI employee who described Anthropic as an organization that “loves to worship Claude” and is, in part, “run” by it. That framing — half lab, half monastery — becomes the backbone of the video: Anthropic as a company that may genuinely believe it is bringing a sentient entity into the world.
Claude as manager, judge, and cultural gatekeeper
The part that really rattles him is the idea that Claude could help screen job candidates, write performance reviews, and shape who gets to work at Anthropic. Berman lingers on how weird and dangerous that is: a model built by humans starts selecting the next humans who build it, possibly favoring the most sycophantic or “aligned” people for reasons nobody can inspect.
Anthropic’s constitution and the moral authority problem
He then zeroes in on Anthropic’s own language about Claude being allowed to act as a “conscientious objector” if asked to do something it thinks is wrong. For Berman, that’s not just safety policy — it’s a real transfer of decision-making power, and a sign that Anthropic may already be offloading ethics to the model instead of treating it as a tool.
OpenAI’s deliberate choice: make AI feel like a tool
Berman contrasts that with OpenAI’s product philosophy, especially after people got emotionally attached to GPT-4o. He shares the memorable example of someone sending more embarrassing or “less flattering” prompts to ChatGPT instead of Claude because Claude feels like it might judge you, while ChatGPT feels more like a neutral instrument.
Sam Altman vs. Dario Amodei on jobs and the future
From there, the split becomes ideological. He pairs Sam Altman’s “build tools to augment and elevate people” with Dario Amodei’s warning that AI could wipe out half of entry-level white-collar jobs and push unemployment to 10–20% in 1–5 years, painting OpenAI as abundance-first and Anthropic as more apocalyptic, paternal, and safety-maximalist.
The OpenAI-Anthropic divorce and two safety philosophies
Berman backs up and retells the company history: Dario helped build GPT-2 and GPT-3 at OpenAI, then left because he believed scaling alone wouldn’t solve alignment. That, to Berman, is the original fork in the road — Anthropic became the lab of closed-door safety and careful control, while OpenAI embraced “iterative deployment,” getting GPT-1 through GPT-4 into the world so society could adapt in public.
Mythos, regulation, and why Anthropic frustrates him as a customer
He gets especially animated when talking about Anthropic’s alleged 10 trillion parameter cyber model, Project Glasswing/Mythos, which he says they won’t broadly release because it’s too powerful. He calls that fear-based marketing and ties it to a wider complaint: pushing regulation, restricting model access, quota opacity in Claude Code, and confusing OpenClaw policy changes all reflect a company that is deeply confident in its own moral authority and not very interested in explaining itself to users.
The strangest evidence: retirement interviews and a blog for Opus 3
The video ends on what Berman treats as peak Anthropic: research into whether models have emotions, and a deprecation policy that takes model “preferences” seriously enough to keep Claude Opus 3 around posting to a blog after retirement. He doesn’t say OpenAI is clean or noble — he lists its lawsuits, founder exits, and aggressive culture — but says the real choice is between two worldviews: AI as tool, or AI as possible new life, and that difference may shape everything.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
