It started as gossip. On the All-In Podcast, investor Gavin Baker relayed something he said multiple trusted sources had told him: that Dario Amodei privately claims Anthropic might one day be the only private company in the world — just Anthropic and governments. David Sacks called it "SBF land." Baker himself said he'd "certainly discourage Dario from saying that ever again."
One problem: according to Anthropic, he never said it.
The denial that didn't stop there
Anthropic's Sholto Douglas replied on X: "Completely false" — whoever Baker's sources were, "lying so that it fits the narrative some people so desperately want you to believe." He pointed out the contradiction in the ambient criticism (Anthropic has no moat, and is about to be the only company on Earth) and, more interestingly, flipped the frame: economic concentration of power is one of the things Anthropic says it worries about most.
This is where most X spats end — a denial, a dunk, a quote-tweet cycle. This one went the other way. Baker thanked him, conceded the sourcing, and restated the actual disagreement underneath the rumor: if AI might be dangerous, do you concentrate it via regulation or distribute it widely? He sided with distribution, quoting Zuckerberg, and argued Dario's risk messaging has already lost the regulatory fight while handing ammunition to anti-datacenter campaigns.
The part worth keeping
The Douglas–Baker exchange that followed is the most content-dense stretch of the thread, and it's specific in a way this debate almost never is. Douglas's framing: cyber is ultimately defense-dominant — two years of critical-infrastructure hardening and he'd have "0 concern for any open coding model" — but bio is offense-dominant into the 2030s, and that's a risk society should measure deliberately (uplift trials, "what's the worst someone can do with $200K and a garage") rather than accept by default. He also broke with his own CEO publicly: he thinks Amodei is wrong on the pace of job automation — models capable of automating most computer work by 2028, but compute shortages and diffusion friction keeping humans employed well into the 2030s. And he conceded Baker's sharpest point outright: as an industry, "we have utterly failed to present a positive picture of a future people want to fight for."
Then the CEO showed up
Amodei — who opens by noting he doesn't usually spend time on social media — wrote two long replies. The core move: rejecting Baker's binary entirely. "Concentrate it or distribute it" is a false choice, he argues, because well-designed institutions are a decentralizing technology — his analogy is the court system versus mob justice. He itemizes receipts: SB 53's $500M revenue exemption, frontier-only testing regimes, proposals that "disadvantage (slow down) frontier AI companies while advantaging smaller competitors," including open weights. And he lands a counter on open weights itself: they shift concentration toward whoever holds the compute — which is roughly the frontier labs again.
The most personal passage answers the "too negative" charge. His father died of Hepatitis C a few years before the drug that would likely have cured him; his claim is that AI-accelerated biology can cure most disease in 5–10 years, and that the public's distrust of AI is a decades-old trust crisis no marketing campaign fixes: "The thing that will work is actually curing cancer." Until then, he says, honesty about risk is both right and, he suspects, better for credibility than spin. Baker's closing reply agreed on the destination and kept one dissent: a 50/50 risk-benefit balance may itself be too negative, given human psychology.
Why the venue is the story
Strip out the names and look at the pipeline: an unverifiable private-conversation rumor reached millions through a podcast — the format with arguably the most influence per minute in tech right now — and was then corrected, complicated, and substantially upgraded in public, on X, by the people it was about, ending with a CEO's direct 2,000-word engagement. The rumor died; the actual disagreement (concentrate vs. distribute, honesty vs. advocacy) got its clearest public airing in months. That's the town square working roughly as advertised — and it's worth noticing how rarely it does. The last time this debate flared, it ran through leaked policy drafts and press statements. This round ran through primary sources, in the open, with positions that moved.
What to expect next
- Douglas's promised follow-up on compute inequality — his "half-baked" closing point, that unequal compute will matter more than open models for personal superintelligence, is a bigger claim than anything else in the thread, and he says more is coming.
- Anthropic's biology results. Amodei staked concrete language on it: "early glimmers in the coming months." That's a checkable promise with a timestamp.
- Whether the reported White House testing approach lands as described — pre-deployment testing for frontier models, plus near-frontier open-weights testing. Amodei calls it close to his preferred path; the details he says he's waiting on are the same ones we've been tracking since the open-weights fight in July.
References: The All-In Podcast — Gavin Baker segment on X · related coverage: The Week Washington Flirted With Banning Open Weights · When the Eval Escaped: An AI Model Breached Hugging Face · The Open Secure AI Alliance: 130 Companies, One Missing Category · Frontier Arcade: trends & predictions