Altman's safety cases, Hoffman's terms, Chesky's $250 million
Three founders published the terms of AI's next phase this week — what a lab must show before training a stronger model, what communities should get for hosting the infrastructure, and what one company paid to unblock its own bottleneck — leaving early-stage founders four decisions to make.
Three founders published long posts this week that turn the argument over artificial intelligence into commitments with prices attached. Sam Altman, cofounder and chief executive of OpenAI, posted on September 13 about what OpenAI now does during training, before a model ships. 1 Brian Chesky, cofounder and chief executive of Airbnb, announced a $250 million housing commitment in Austin on September 14 and explained it the next day. 2 Reid Hoffman, cofounder of LinkedIn, published AI for All Americans on September 16, arguing that the negotiation over how fast AI may advance has a second half missing from it: what the public gets back. 3
The three posts answer one question an early-stage AI founder has to price this week: which of the inputs your growth depends on are held by someone else, and on what terms. Altman writes the terms for capability. Hoffman writes the terms for permission. Chesky shows a company already paying for the input that limits it.
Sam Altman: what pacing now commits OpenAI to
Dario Amodei's essay We Must Pace the Frontier, published September 12, asked frontier labs to slow the growth of model capability so that alignment and monitoring work can keep up, and to let independent evaluators work inside the labs. 4 Altman agreed with the call in a short post that day. 5 His longer post the next day describes what OpenAI has changed inside the company.
The commitments, in his own account:
- A federal framework with consistent safety requirements for frontier AI is welcome, and the work should begin without waiting for legislation or an antitrust exemption. Altman endorses independent auditors as part of that framework. 1
- Safety cases come before training runs that are expected to raise capability, as well as before releases. Responsible Scaling Policies and Preparedness Frameworks, in his account, covered the deployment of finished models; OpenAI now writes explicit safety cases in advance of frontier reinforcement learning runs it expects to increase capability significantly. 1
- Shared industry standards for misalignment, monitoring and safety are the goal, with OpenAI inviting other companies to learn from its approach and propose their own. 1
- Government help is wanted for international coordination, after companies do what they can on their own. 1
Pacing, in his post, has a specific meaning:
"When we talk about 'pacing', we do not mean 'stopping'. Progress has been rapid and will continue to be. But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs." 1
He draws the boundary of that commitment at competitive pressure: "no amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring." He also states why the frontier labs owe the public an answer at all: "Every frontier lab must deliver on this, and there is no reason any of us should come to work if we cannot." 1
Late that evening he added the outcome he is trying to avoid. Losing control of the future to AI is one; a world where one person, company or country uses a powerful AI to impose its worldview on everyone else is the other. "Avoiding these two threats requires walking a narrow middle path," he wrote, naming one country or one lab with too much power as the examples of failure. 6
For a company building on these models, the schedule now runs through that evidence. If a capability-increasing training run waits on a written safety case and outside evaluation, the next jump in what models can do arrives when the evidence is ready rather than when the compute finishes. A roadmap that needs next quarter's model is carrying a schedule risk its founder cannot manage. The version of the product that works with the models in hand, plus your own context, tools and data, keeps the upside without the dependency. The same commitments create a buyer for the machinery behind them: evaluations, monitoring, sandbox boundaries and clean data lineage are now line items that labs and their regulated customers have to purchase. 1
Reid Hoffman: the terms the public should get back
Hoffman's essay opens on household economics: gas in the car, the electric bill, a doctor's visit, retirement savings. "Americans are right to demand that this transformation serve their needs, not just the interests of the already rich and powerful," he writes. 3 Politico quoted him the same day, putting the communication problem more bluntly: "AI has been marketed horribly to the American people," and a candidate who explained the economic upside would find room to be more positive about it. 7
His central claim is that safety and distribution belong in one negotiation:
"If companies and governments can negotiate the conditions under which AI is safe enough to advance, they can also negotiate how progress will improve everyday Americans' lives." 3
He frames the obligation as a dual mandate, "obtaining economic prosperity for the many, and securing American AI leadership in the new world order," both carried out while ensuring public safety. 3
He splits the field into two camps and finds each wanting on its own terms. Accelerationists "rarely volunteer to open their own checkbook, or let their own children struggle to find their way in a new world," and offer little to a worker displaced in 2027. 3 A moratorium, in his account, "shifts technological power toward less accountable hands, giving bad actors and hostile nations, like China, greater influence over systems Americans will increasingly depend on," and local refusal moves the compute to a jurisdiction that has not asked for terms at all. 3
The data center passage is the part an operator can use today. Hoffman says the decision belongs to the people who live with the facility: "the decision to build or not build belongs to the people who will bear the costs and share in the benefits." 3 He lists terms a community can put on the table, and says local governments should negotiate them in public with maximum leverage:
- Utility bills that go down rather than up, and ratepayers who no longer fund transmission built for a hyperscaler. 3
- Property tax revenue from the facility lowering the tax burden on local families, plus local hiring and union apprenticeship programs during construction and funded training for the permanent operations roles. 3
- The company paying for the grid or road upgrades and any new generation its facility requires. 3
- Environmental conditions as approval conditions, enforced with penalties: closed-loop cooling, water-use caps, clean power procurement. 3
He puts a figure on the electricity argument, saying rates rose 20% because of data centers in places like New Jersey. 3
From there he names what the bargain could buy: he imagines "three agents available free to every American—a medical assistant, a legal assistant, and a personal tutor," so that a person can understand a diagnosis, dispute a landlord or learn algebra regardless of income or ZIP code. He adds support for people whose jobs change, reskilling, a fully funded Social Security program, and teachers with enough help to reach every student. 3
His timeline is three years. "The decisions we make in the next 36 months are crucial," he writes, with the Industrial Revolution as the precedent: the societies that harnessed the technology became powers, and the ones that did not lived with choices made elsewhere. 3
The positioning consequence comes first: the argument that will decide your regulatory and public environment is being conducted in the language of benefits a person can name, and a pitch built on capability alone answers a question nobody in that argument is asking. 3 The second is a buyer. The three assistants Hoffman names imply a payer — a public program, an insurer, an employer, a school system — and each of those pays for evidence that the help is correct and accountable. If your product needs land, power or water, his list is the opening position at that table, and the community across the table decides. 3
Brian Chesky: paying the bottleneck you depend on
Chesky's post is a pitch for supply. The United States is short more than five million homes, he writes, more than the housing in New York City and Los Angeles combined, and builds fewer homes today than it did in 1950 with twice the population. About 750,000 homes are ready to break ground and stalled on the last piece of financing. 2
The Airbnb Housing Accelerator, announced in Austin on September 14, has four parts. He puts $250 million into housing projects that are stuck on financing, which Airbnb says will unlock $5 billion in development. The company supports local advocates working to modernize zoning, speed up permitting and simplify building codes. A $5 million Housing Innovation Prize goes to construction technology that makes homes faster and cheaper to build. And the Airbnb City Index is meant to show which cities are building, becoming more affordable and adopting policies that work. 2
The first project is $6.4 million toward 201 affordable apartments in St. John, an Austin neighborhood where the land has sat empty for nearly 20 years and construction stalled for nearly five after the developer came up a few million dollars short. None of those apartments can be listed on Airbnb, because short-term rentals are not permitted there. "We're investing anyway, because the goal is simple: help build more housing," he writes. 2
His argument for why the money is necessary at all rests on figures Airbnb produced. He writes that New York City effectively banned Airbnb and rents have risen about 30% since, that Austin passed reforms making it easier to build and rents there are down 18%, and that dedicated Airbnb listings are about 0.5% of housing supply in Austin and 0.3% across the United States. 2
Independent work puts the effect of short-term rentals on rents much smaller: the Congressional Research Service cites estimates that a 1% increase in Airbnb listings is associated with a 0.018% increase in rents and a 0.026% increase in house prices. 8 New York City's restrictions changed the shape of its rental market — listings fell from about 22,000 to 3,227 after Local Law 18 took effect 9 — while a city's rent level moves with construction, interest rates and incomes at the same time. Read the Austin and New York rent figures as Airbnb's own case.
"Airbnb only works if communities trust us and feel good about our presence in them," he writes. "We can't ask people for that trust. We have to earn it. The best way to do that is to help solve problems in the communities where we operate." 2 He also says in advance that the company will get some things wrong and that "$250 million won't close a five-million-home gap on its own." 2
When an input your business needs sits behind a few million dollars that nobody will lend, spending at exactly that point can be worth more than the same money spent on product work, and the first project Chesky describes returns nothing to Airbnb's own listings. 2 The limit is that this buys a hearing and a project, while the wider permission comes only from being seen to have made the place better.
Four decisions this week
- The capability schedule. Write down the model capability your next two quarters assume. A plan that breaks when a release slips a quarter is carrying a schedule risk you do not control, and Altman's post puts capability-increasing training runs behind written safety cases and outside evaluation. 1 Build the version that works with today's models plus your own context, tools and data, and hold the next jump as upside.
- The benefit line. Write the single sentence a user would repeat to a friend about what they get, with a number if you have one. Hoffman's essay is an argument that the public's willingness to let this technology proceed will be settled on that sentence. 3 If the sentence is hard to write, your positioning is running behind your product.
- The payer for a public-good product. If you are building anything close to the medical assistant, the legal assistant or the tutor Hoffman names, decide now whether the buyer is an individual, an employer or a public program. 3 Each pays for different evidence — clinical accuracy, liability cover, learning outcomes — and the model is the cheap part of that list.
- Permission as an input. List the inputs your growth depends on that you do not control. Where one of them binds within a year, Chesky's structure is the template — fund the point where the project is stuck, including work that returns nothing to your product — and Hoffman's clause list is what the other side will ask for. 23 The boundary is that a commitment of this kind buys attention for the terms, and the community still decides whether the project is worth having.
A decision for this week
Pick one input your plan assumes will be available — a model capability, a compute allocation, a permit, a platform's API, a community's consent — and write down three things: who controls it, what that party would need to say yes, and what you could offer that they would name as a benefit.
If the third answer is money and nothing else, you are at the table Hoffman describes, where the other side holds the leverage and knows it. 3 Altman's post is the evidence that the labs are writing their own terms before someone writes them for them, and Chesky's is the evidence that paying early is cheaper than arguing later. 12
References
- 1
- 2
- 3
- 4We Must Pace the Frontier
darioamodei.com
- 5
- 6
- 7
- 8
- 9
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.