Pioneers Insight Method Research Author
Dario Defends Himself, Datacenter Panic, AI Doomer Trap, Senate Toss-Up
Back to Episodes

Dario Defends Himself, Datacenter Panic, AI Doomer Trap, Senate Toss-Up

Summary

  • Dario Amodei finally broke silence with a two-part essay, and Sacks reads it as a response to the regulatory-capture charge — not Gavin Baker’s “only private company in the world” claim, which Dario did not address. Dario said his messaging had been balanced, while Sacks’s verdict is that Anthropic’s aggressive effort to implement preferred state and federal frameworks “is regulatory capture” even if Dario is sincere. Sacks also points to the 50%-of-entry-level-jobs campaign and a blackmail study he says was prompted “over 200 times”; his conclusion is that Dario’s blindness is “disingenuous or delusional.”
  • Chamath’s macro warning is the tradeable core of the episode: AI doomerism has metastasized into bipartisan data-center pushback (Abbott’s Texas executive order, Shapiro in Pennsylvania, and a GOP memo over the Ohio Senate race) just as rising yields choke risk capital. Frontier labs are “the most at risk” — non-investment-grade companies with balance sheets that need ever more compute they cannot fund — creating “a very precarious situation that we could have frankly avoided.”
  • Sacks predicts an open-source ban is coming via a Trojan-horse “FINRA for AI” — a “DMV for AI” whose equal standards open models cannot comply with in the same way. His acceptable alternative is an MPAA-style body with no government reporting line; “China’s gonna win if we do this.” Chamath’s rebuttal: banning open source would “cripple the American economy” — Coca-Cola and Pfizer would redirect capital abroad and FDI would fall off a cliff, in his Jack Ma-crackdown analogy.
  • Chamath’s open-source thesis rests on harness data: closed models “decay in capability” when wrapped in their own harness, while open models improve when used with other open-source models — “meaningfully more performant and dramatically cheaper.” Unless open source is banned, “the American consumer will win”; the only real debate is “which private American company will win or lose.”
  • Friedberg believes recursive self-improvement is “likely to happen soon,” which makes regulation a “fool’s errand” — RSI needs only “the chips and the power and a connection” and could run from Kazakhstan, Iceland, or the ocean. Better to attract labs into U.S. jurisdiction; Sacks treats RSI as a black swan but wants multiple competing AGIs — “choice has to be our hallmark,” with the worst outcome a “woke Skynet.”
  • The populist backlash is about affordability and trillionaires, not AI safety: “they are sharpening the guillotines,” per JCal, who predicts Waymo bans in New York, Boston, and Washington, D.C. Chamath’s one-degree tune: it is vibes — “we’re gonna go from billionaire… to now trillionaire, and it’s those guys? Those guys are even worse than the last crop.”
  • On the midterms, Polymarket has the Senate at 52% for Democrats — a toss-up — but Sacks cites Patrick Ruffini’s finding of a D+3.7 average polling error across more than 3,000 polls and predicts the GOP ultimately keeps the Senate. His message deck: murder rate lowest in 70 years, overdoses down 20%, refunds up 17%, and eggs down 39% year over year.
  • Friedberg sees a socialism wave culminating in “someone like an AOC” as president by 2028: 53% of under-40 conservatives back government-run grocery stores, the 30-year Treasury yield reached 5.3%, and 63% live paycheck to paycheck. Sacks counters with Cato’s $71–212 trillion ten-year DSA price tag against $182 trillion of total U.S. population net worth — socialism is “a disease that claims to be a cure.”

Deep dive

1. Dario enters the chat — but answers Sacks, not Gavin

  • The setup: Gavin Baker relayed that some people allegedly inside Anthropic believe “they will be the only private company in the world at some point”; Sholto Douglas shot back, “Completely false. Whoever told Gavin this is lying.” Dario’s essay took some responsibility — “we haven’t yet delivered on our big promises to benefit the world. That is totally on us” — and offered the line of the week: “the thing that will work is actually curing cancer, not glitzy marketing.”
  • Sacks’s read on what drew him out: not Gavin’s claim, which Dario did not address, but the regulatory-capture accusation Sacks has pressed “since last year.” His verdict stands: Anthropic has been “extremely aggressive about seeking to implement its preferred regulatory frameworks at both the state and federal level. That is regulatory capture.” Dario may be sincere rather than motivated by pecuniary gain, but “he is seeking to capture the machinery of the state on behalf of his political agenda.”

2. The rap sheet on “balanced” messaging

  • Dario said his messaging was “about equally balanced between risks and benefits.” Sacks: do you expect amnesia? The claim that 50% of entry-level knowledge workers would lose their jobs within one to five years was not podcast riffing — Sacks says it was “engineered and amplified into a full-fledged media campaign,” retweeted by Obama and other Democratic officials.
  • The blackmail study fits the same pattern, in Sacks’s telling: an alignment team prompted the model “over 200 times until they finally get the headline-grabbing result that they wanted,” then Dario, his sister, and other executives “breathlessly pumped” it on 60 Minutes. Sacks says the UK AI Safety Institute also found the study was created under highly pressurized conditions and that the result had been engineered. No founder or CEO, he argues, has done more to promote these fears, making Dario’s failure to see his own role “either… disingenuous or delusional.”
  • The asymmetry Sacks flags: Sam Altman recanted — “I was too negative. I was wrong” — while Dario, now past the one-year mark on his jobs prediction, “has doubled down… He’s an ideologue.” Years into the AI boom, Sacks says, “we only see net job creation.”

3. Chamath’s three-legged stool: doomerism, political pushback, and the funding squeeze

  • The doom narrative has “seeded, and now it’s growing” into mainstream political action: Abbott’s executive order limiting Texas data-center growth, Shapiro doing the same in Pennsylvania, and an Axios-reported GOP memo telling AI executives to “stop rage-baiting people” ahead of a critical Ohio Senate race.
  • The third leg is rising yields — Chamath also said the Treasury Department had just doubled its bond buying — which raises the risk-free rate and constrains risk capital just as data centers are being limited or shut down. Frontier labs sit at the bottom of that cascade: “they’re not investment grade, they’re in the worst position in terms of their balance sheets,” and growth makes it worse because they cannot fund the compute that growth requires.
  • His coda, which Sacks endorsed with “Totally agree”: “we have made some really stupid unforced errors in the United States… we are now in a very precarious situation that we could have frankly avoided.”

4. Friedberg’s steelman and Chamath’s thinking-tokens demand

  • Friedberg steelmans the lab leader: pressure-testing reveals jailbreaks that could enable bioweapons, cyberweapons, and mass manipulation. X’s algorithm is public, so “theoretically an agent could read it, copy it… and figure out a way to manipulate the broad public.” The executive’s problem is that “you can’t get a grip on how to keep it under control… What do you do differently?”
  • Chamath’s condition for taking the fear seriously: “you cannot obfuscate the thinking tokens of a model then.” Claude and OpenAI hide them; open models let you “see it thinking” and observe misalignment in real time — and, he noted, he has not seen the same doomerism from any open-source model company, including NVIDIA. “Make the bloody thing available to other people then.”
  • His rebuttal to industry cross-review: labs will never share every trace and span — “you’re essentially giving away a lot of your recipe” — and if Anthropic is not willing to share thinking tokens even with an API customer, it certainly will not share them with OpenAI in a room. “Not gonna happen.”

5. SRO yes, FINRA no: “a DMV for AI”

  • Friedberg presses Sacks on Demis’s FINRA-style proposal, which Friedberg said Dario, Sam, and Elon had signed onto. Sacks says FINRA is “a pretty sleepy regulatory agency” that founders see as “existing to protect the big banks from disruption.” Demis wrote something high-level, Sacks argues, and “there’s many versions of Demis’ idea.”
  • Sacks’s preferred model is an MPAA-style industry body, created after the Hays Code to forestall heavier government action, not a FINRA-style regulator that reports to government and conducts pre-release testing. Having spoken with Elon, Sacks says Elon supports an MPAA-for-AI, not FINRA.
  • His name for the pre-release regime is “a DMV for AI,” with models queued for approval — “a repackaging of what Dario has always wanted, which is an FDA for AI. Now he’s saying it’s an FAA for AI.” Those agencies take years; wanting FAA-modeled AI regulation is “completely discrediting,” and “China’s gonna win if we do this.”
  • JCal’s examples: China had landed a fully steel rocket while Elon was “still waiting for FAA approval on his steel Starship,” and JCal said biotech investment was moving to China because its revised clinical-trial rules move companies through the process faster.
  • Sacks’s alternative safety stack: liability pressures labs to act responsibly — Meta is being sued “for over a trillion dollars” over alleged addiction and lost a $6 million Los Angeles body-image verdict; Sam had also said frontier releases would be slowed because the “safety stack was not keeping up with the capabilities.” Industry coordination through papers, symposia, conferences, and RFCs can share best practices. The common denominator is openness, versus secret standard-setting that produces “regulatory capture and anticompetitive behavior.”

6. What “China wins” actually means — and the harness insight

  • Friedberg forces the question: if U.S. regulation slows the labs and China’s models leap ahead into RSI, what tangibly happens to the American worker, business, and consumer? Sacks’s counterfactual is to imagine Baidu, Alibaba, and Tencent instead of Google, Microsoft, and Meta — a less vibrant economy, the wealth in China, and a dual-use military disadvantage. “America is not gonna be the number one country in the world economically and militarily if we lose AI.”
  • Chamath’s dissent: unless America bans open source, “the American consumer will win.” His evidence is the harness phenomenon — Codex, Claude Code, Hermes, and OpenClaw: wrap a closed model in its own harness and it “decays in capability”; use an open-source model with another open-source model and it improves. Open technologies are “meaningfully more performant and dramatically cheaper,” and this data “is just going to be a tsunami over the next year.”
  • JCal’s answer: we already know what China winning could look like — cheaper phones, cheaper cars, and a cancer cure sold back to Americans — “and then we’re gonna lose our jobs to them.”

7. The open-source ban, step by step — versus capital flight

  • Sacks’s prediction: “I got bad news for you, Chamath. An open source ban is coming. They’re not gonna call it that.” His proposed mechanism is FINRA Trojan horse → pre-release testing → codification in law → “all these standards need to apply equally to all models.” Referring to what he thinks was Dario’s 2023 Senate testimony, Sacks says open models cannot be centrally monitored, controlled, rolled back, or withdrawn. Equal standards would therefore gradually shut them “out of the marketplace.”
  • Chamath’s rebuttal: force Coca-Cola or Pfizer to run “the slower, more costly, less effective, less valuable model in the United States” and they redirect capital allocation abroad; FDI goes “off of a cliff.” His analogy is Jack Ma’s speech on the eve of the IPO, followed by Xi’s crackdown and government representatives and “golden votes” on company boards. Chinese FDI and the economy “fell off a cliff,” Chamath says, adding, “And then what are you gonna do, take away our passports?”
  • Sacks concedes the consequences, not the possibility: “You’re saying regulators have never done things that are crippling?” Data centers, he says, will be 3% of GDP and are being restricted anyway. Chamath’s counter is that voters are not voting against GDP — “they’re voting against two people that they deeply dislike becoming trillionaires.”

8. Why voters hate AI: guillotines, vibes, and the autonomy flashpoint

  • JCal’s diagnosis: no rank-and-file American cares about AI safety — “the discussion for them is, ‘Why am I not getting rich and everybody else is?’” Dario warned of job losses, Elon said people would not have to work, and Sam ran UBI studies. Those messages reached people with no savings amid persistent affordability problems. “They are sharpening the guillotines,” JCal said, citing the UnitedHealthcare CEO’s killing and the firebombing of Sam Altman’s house.
  • Chamath tunes it by one degree: it is vibes. This crop of “tech oligarchs” is “not aspirational… not Steve Jobs by any stretch of the imagination,” and now “we’re gonna go from billionaire… to now trillionaire, and it’s those guys? Those guys are even worse than the last crop.” Silicon Valley “has completely lost what made it great” — a Harvard-like “credentialing place” rather than a home for weirdos and idealists.
  • JCal’s prediction on autonomy: Zipline and Uber are heading toward 1 million deliveries a day, Waymo is roughly doubling the number of cars on the road every six months, and Tesla is moving toward Cybercab. Driver wages that “at least stayed in the community” could instead flow up to Silicon Valley, so “I will predict it now, they’re gonna ban Waymos in New York, Boston, and Washington, D.C.,” with licenses limiting them elsewhere.
  • The messaging fix: paradoxically, Jensen Huang — “You’re not gonna get replaced by AI. It’s gonna be somebody who works with AI” — and Zuckerberg, with free tools and a training program for fiber and data-center construction jobs, are the industry’s best spokespeople. Sacks: “It’s a very low bar.”

9. Sacks defends Dario — as a sincere dystopian

  • The surprise turn: “I actually think he is very idealistic. I don’t think money is his main motivation.” But the ideal — “machines of loving grace” that apportion society’s resources while humans twiddle their thumbs — “is dystopian, and that’s why a lot of their messaging is dystopian.” If Dario is in the lead to bring about recursive self-improvement, Sacks says, “he should defend this future,” not be thanked for “coming down from Mount Olympus and slumming it with the rest of us on X.”
  • On the money question, JCal said it was his understanding, reportedly documented in press reports, that the seven Anthropic co-founders donated 80% of their shares. He said roughly 10% of Anthropic is in some sort of trust and that each founder owns about 1.8%; at a $2–3 trillion valuation, Friedberg called the trust potentially “the largest PAC in the world,” specifically a Democratic or DSA PAC. Sacks asked, “A trillion dollars to EA causes, what could go wrong?”
  • Why Sacks thinks Dario cannot simply win through transparency: “what he’s selling is fear in order to motivate the government to give him control to bring about this dystopian future that he believes in.”

10. RSI makes regulation a fool’s errand

  • Friedberg’s conviction, hedged as his own view: RSI — models spinning up agents that build better models without humans necessarily architecting each evolution — “is possible. I think it is likely to happen soon.” People he has talked to in the industry think RSI is real. His biology analogy is that models are decomposing into sub-models like genes making proteins that interact in networks; that evolution “is likely going to happen without a human architect,” which is “why I think RSI is probably the future.”
  • The policy consequence: RSI needs only “the chips and the power and a connection” — it could run from Kazakhstan, Iceland, or the ocean — so attempts to ban data centers or create a regulatory body are “a fool’s errand.” Friedberg asks whether it is better to attract labs into the U.S., where they can be monitored and controlled, rather than chase them offshore.
  • Chamath says a new regulatory agency would not be technically sophisticated enough to stop RSI and would instead concentrate power. Sacks treats RSI as a black swan — “very hard to assess the probability” — but wants multiple AGIs as checks: Elon’s “maximally truth-seeking” approach versus Anthropic’s Claude constitution “full of liberal values.” Citing a Terminator television plot in which a benevolent AGI defeats Skynet, he says, “choice has to be our hallmark,” with the worst outcome a “woke Skynet.”

11. The a16z DOJ “investigation” as a one-way PR ratchet

  • According to a Bloomberg report, Andreessen Horowitz is allegedly being investigated by the DOJ for more than a year under Section 8 of the 1914 Clayton Act, concerning interlocking directorates. The reported conflict is Ben Horowitz on Databricks’ board while Martin Casado sits on the board of Five Trains, the entity name rendered in the transcript.
  • Chamath’s speculative read: an aggrieved CEO “reported it to the DOJ and reported it to Bloomberg at the same time.” If a complaint causes an opened file, he said, the firm is “technically under investigation,” but the DOJ does not announce when such a matter closes — “a one-way PR ratchet for an enemy to exact some sort of chaos.”
  • Sacks’s substantive defense: startups pivot constantly, so noncompetitors can become competitors; different partners plus a firewall could address the issue, and these are not filing companies. Chamath’s motive frame is that a16z “is on fire” — Cursor was “an appetizer,” OpenRouter “an amuse-bouche… like 600 or 700 million,” and SpaceX “the main course,” at roughly $15–17 billion.

12. Senate toss-up — but Sacks says don’t overread the polls

  • The board: Polymarket has Democrats at 88% for the House but 52% for the Senate — a toss-up. On February 25, the day before the Iran war started, the GOP was favored 60–41; it had been as high as 70%. Republicans hold a 53–47 Senate edge; 35 seats are up in 2026, 23 of them Republican-held. Alaska, Maine, North Carolina, and Ohio are the races to watch.
  • Sacks cites Patrick Ruffini’s analysis of more than 3,000 polls across four cycles since 2018: an average D+3.7 error, with oversampling of young progressives. He says DSA candidates were sometimes expected to win by a lot but merely eked out wins or lost, as in Wisconsin. His call: Democrats are legitimately favored for the House, but “the Republicans will ultimately keep” the Senate.
  • His post-Labor Day message deck: border encounters at their lowest level in 50 years, the murder rate at its lowest in 70 years and down roughly 20% year over year, overdose deaths down 20%, tax refunds up 17% through the cited tax changes, small-business optimism at its highest level in nearly a year, prescription-drug prices — he said he thought — at their steepest annual decline since 1963, and eggs down 39% year over year in July. “They just have to remind [voters] of all the good things.”

13. The socialism wave and the Girardian scapegoat

  • Friedberg’s numbers, delivered as a warning: 53% of self-identified conservatives under 40 support government-run grocery stores; Republicans’ favorable view of capitalism fell from 72% in 2019 to 61% in July 2026, while the share saying the system is rigged toward the wealthy rose from 30% in 2018 to 42%. With the 30-year Treasury at 5.3% and hitting new 20-year records, and 63% living paycheck to paycheck, “I think someone like an AOC becomes president” — a wave this year and a bigger push in 2028. The asset divide: the top 50% hold $178 trillion of net worth, the bottom 50% just $6 trillion.
  • Sacks’s counter-frame: healthcare has become “almost 300% more expensive,” while computer software fell 67% in real dollars and average hourly wages grew 135%, figures he used to argue that inflation is worst where government involvement is greatest. Competition, not more government intervention, brings prices down; socialism is “a disease that claims to be a cure.”
  • Sacks cited a Cato Institute estimate that the DSA platform would require $71–212 trillion in new federal spending over 10 years, against $182 trillion of total U.S. population net worth and only $6.6 trillion for the Forbes 400. Friedberg games out the sequence: 30-year yields spike to 7–8%, then “a national asset seizure tax”; even 10% of the $23 trillion held by people worth more than $50 million would cover only three months of current fiscal spending.
  • The closing frame: “this is not about mathematical realism. This is about anger towards a class of people that are increasingly odious” (Chamath); the data center is “the temple where the billionaires and trillionaires pray” (Friedberg); through the René Girard lens, billionaires and then trillionaires get scapegoated for a catharsis “that will allow society to reset.”
  • Prescriptions diverge: Chamath — “Pollsters, trash. Media, trash” — wants corporate America to “present the better face of capitalism” and the White House to tell CEOs to “play nice in the sandbox ‘cause you’re about to lose the standard operating procedure that allowed you to exist in the first place.” JCal wants Trump to build the second half of his term around affordable homes and trade schools: “Now do it for homes, Trump. Use that same energy.”