This is a special essay responding to this week’s All-In-Podcast. It addresses how we can all align to approve of the rapid development and deployment of AI, a heated discussion on their podcast.
cc - Chamath Palihapitiya
Source: All-In Podcast: “Dario Defends Himself, Datacenter Panic, AI Doomer Trap, Senate Toss-Up”
The All-In crew debated Dario Amodei, AI regulation, data centers, jobs, China, and recursive self-improvement. Under the hood it is really one argument.
David Sacks sees regulatory capture. Chamath sees political backlash. Friedberg sees a real technical risk that cannot be hand-waved away. Jason sees the thing voters actually care about: jobs, wages, and the fact that most people do not feel richer when Silicon Valley gets richer.
All four are right. But none have nailed the big picture.
Dario’s case, steel-manned, is simple. Frontier labs see capabilities the rest of us do not see. They test models under adversarial conditions. They see cyber, bio, persuasion, autonomy, and agentic failure modes before they show up in consumer demos. Friedberg makes that case well: if you are building the frontier system and you can see how it might be jailbroken or misused, some form of safety process feels morally required.
“By far the most accurate criticism of AI companies, including Anthropic, is that we haven’t yet delivered on our big promises to benefit the world. That is totally on us.” Jason, quoting Dario’s essay
But the moment safety replaces any “big promise”, as it has, the discussion turns to regulation. When that happens, ceding pre-release permission becomes a self-serving power play.
Sacks is right about that. He argues that Anthropic has been aggressive in seeking its preferred regulatory framework, and that this is regulatory capture even if Dario’s motives are sincere. His core charge begins here: a private company should not be able to capture the machinery of the state around its own theory of risk.
“He is seeking to capture the machinery of the state on behalf of his political agenda. It’s just, there’s not much to debate there.” Sacks
The distinction matters. A voluntary industry standard is one thing. A FINRA, FAA, FDA, or DMV for AI is another. Sacks calls the FINRA model a DMV for AI because it would line models up for pre-release testing and slow everything down. Worse, the incumbents would help write the rules. Open models would struggle to comply because they cannot be centrally monitored, recalled, or throttled in the same way as closed services. That is how an open-source ban arrives without being called one.
Chamath is right about the backlash. AI leaders spent years telling the public that AI might take their jobs, destroy society, or require emergency control. Now governors in Texas and Pennsylvania are discovering that voters do not want data centers in their communities. Chamath describes the chain reaction: doomer messaging leads to political pushback, political pushback threatens data centers, and a capital-intensive industry suddenly finds itself fighting the communities it needs.
“You have doomerism and then you now have mainstream political pushback from both sides of the aisle, which I think is extremely dangerous.” Chamath
That should surprise nobody. If you sell the future as a threat, people eventually believe you.
Chamath’s sharper point is about legitimacy. The public is not only angry about energy, water, or zoning. Those are the surfaces. The deeper problem is that the next generation of AI wealth looks likely to make a small number of already powerful people even richer. His “vibes are terrible” point is crude, but it is true. People are not voting only against data centers. They are voting against trillionaires they do not trust.
“I think it’s vibes. And the vibes are terrible.” Chamath
Friedberg is right to keep the hard technical question on the table. If recursive self-improvement is real, then bureaucratic release gates will not stop it. The work will move wherever there are chips, power, capital, and jurisdictional freedom. His RSI argument starts here: once models can help build better models, the process may not fit a six-month release cycle with a regulator standing at the door.
“All that you need to pull this off is... chips and the power and a communication connection.” Friedberg
That does not mean we should surrender to science fiction. It means the U.S. choice is not “regulate or don’t regulate.” It is whether the most important AI systems are built inside a society that can shape them, contest them, sue them, inspect them, compete with them, and benefit from them.
Jason is right about the politics. Ordinary people are not waking up worried about model evals. They are worried about rent, groceries, wages, health care, and whether the next automation wave sends more local income into remote corporate balance sheets. He puts the issue plainly: people who have to work for a living hear “AI will take your job” from the same industry whose leaders are becoming vastly richer.
“Nobody, no regular rank and file American cares about AI safety. That’s not the discussion for them. The discussion for them is why am I not getting rich and everybody else is?” Jason
That is why the data-center backlash is so potent an issue. A Waymo ride may be cheaper and better. A drone delivery may be more efficient. But if the driver’s wage disappears from the neighborhood and the margin flows to the cloud, the politics will turn ugly. Jason’s autonomy example gets to the real fear: the money leaves the community.
This is why “AI will create jobs” is not enough. It may not. And logically Elon is right here, jobs will be automated over some time period and because of that prices will fall, and money will mean less and less.
It may be true that some jobs are being created. It is already true in some companies. Jason says that teams which embrace AI tools are opening new business opportunities and even hiring more people. That is useful evidence. Jensen Huang’s line that you will not be replaced by AI, but by someone using AI, is also useful. Zuckerberg promising free tools is useful. Trade schools for data-center and fiber jobs are useful.
But useful is not the same as sufficient.
The Human Dividend is the missing answer.
My unpublished thought in ‘The Human Dividend’ that we need a Human Wealth Fund needs to be discussed. The idea of giving ownership to every living human is the glue to align us all.
The idea has two parts.
First, intelligence should become cheap, abundant, and widely available. The meter should fall. Everyone should have access to useful intelligence, not only corporations, governments, and rich subscribers. Open source, routing, edge models, cheap inference, better devices, and competition all matter because they push intelligence toward universality. Chamath’s open-model and harness argument is important here: useful intelligence will not only come from the largest closed model. It will come from how models are wrapped, routed, combined, deployed, and made cheap.
Second, the surplus should be broadly owned.
Not nationalized. Not government-run AI. Not a state ministry of intelligence. The companies building AI should build, own, operate, compete, meter, and profit. But if AI is bottling the accumulated intelligence of humanity, then humanity needs an asset claim on the value created in order to welcome the transition to an automated economy. Simple really, people need to benefit economically and in terms of life-style.
That is the Human Wealth Fund idea.
It is the better answer to the argument.s underlying these issues.
To Dario, it says: safety cannot depend on monopoly or secrecy. If frontier labs want public trust, they need open standards, liability, external testing, disclosure where appropriate, and governance that does not quietly outlaw competitors. Sacks is right that industry collaboration can happen through papers, technical blogs, conferences, requests for comment, and open standards. His point is that the process must be contestable.
To Sacks, it says: competition is necessary, but not sufficient. A world of many competing AIs is better than one approved AI, but competition alone will not solve the political economy of displacement and concentration.
To Chamath, it says: the backlash is not irrational. People are reacting to a future in which the same small group appears likely to own the next civilization-scale asset class. Change the ownership structure and you change the politics.
To Jason, it says: jobs matter, wages matter, training matters, but ownership matters more. A higher minimum wage may help. Trade schools for data-center workers may help. Free AI tools may help. But none of those gives every person a share of the upside.
A better version of Norway’s $2.3 trillion sovereign wealth fund. Global, and based on ownership of shares. And iredeemable.
To Friedberg, it says: if AI is as powerful as the frontier labs believe, then the answer cannot be to chase it offshore with theatrical control. We should want the labs, data centers, chips, energy, and talent here. But we should also want the public to have a reason to cheer for that future.
The great mistake would be to turn AI into a morality play between doomers and accelerationists. The real question is simpler.
Who benefits from its success? If the answer is shareholders then everybody should be a shareholder. Gifted now and held until death, with new borns getting a share at birth.
If the beneficiaries are “a handful of frontier labs,” the public will eventually rebel. If the answer is “the state,” innovation will slow and power will centralize in worse hands. If the answer is “nobody, just trust the market,” the politics will not hold.
The Human Dividend is the smart forward looking path: build fast, keep intelligence cheap, preserve competition, reject regulatory capture, and create a broad ownership claim on the wealth AI produces.
That is how AI wins public legitimacy.
Not by frightening people into accepting control.
By making them owners of the future.
You can read the current draft of The Human Dividend here -