The New Silicon Valley

On a valley made of receipts, and what's left when the fog lifts.

I've lived in Silicon Valley for twenty-nine years. Junior high, high school, college, and everything after. I've watched every arc of it, and every arc has a cost.

So I want to be careful about what I mean by the title. I'm not writing against the place, and I'm not accusing a company or a person. I'm saying that Silicon Valley was never really a location. It was a set of defaults, and defaults can be rewritten by anyone willing to build with different ones.

This is a post about the different ones. It's a summary of what a year of research adds up to, and then a forecast: what I think the world looks like if it all ships. The research is the evidence. The forecast is my bet, and I'll mark which is which.

(Drafted with A.I. from my own notes and repositories, then edited by me. The decisions, and the bets, are mine.)

The old defaults

Here's the short version, described as mechanism and not as motive. Nobody had to intend any of it. The incentives did the work.

• Collect first, explain later. Data flows up to the center because the center is where it's valuable. Consent arrives afterward, if at all.
• Optimize for attention. The metric that's easy to measure becomes the thing the product is for.
• Keep the influence invisible. Persuasion that's quiet works better than persuasion that's labeled, so the steering lives where nobody can inspect it.
• Ambiguity as a safe harbor. Vague wording, vague policy, vague feedback. Nothing is committed to, so nothing can be held against anyone.
• Move fast, ask forgiveness. Boundaries get found by crossing them.

I wrote about the product side of this in The Case for Personal A.I.: a feed can't tell you which of your thoughts were yours, and a chat window can't tell you whether it's steering you away from harm or toward a purchase.

I don't think those are two problems. A culture that runs on fog builds products that run on fog. An organization that can't be asked directly will ship software that can't be asked directly.

What the research adds up to

When I lay the past year's projects side by side, they look like separate efforts. Seer, Totem, Thread, Mary, Veil, Scorpion, CLR. But they all answer the same question, and the question isn't about AI. It's: can I ask? Can I ask where this came from, where it went, what it did, who it's pulling from, who's paid for it?

Each project is one version of making the answer a yes.

1. Every output knows where it came from. Seer began as a retrieval system and became an attribution layer: every response carries a Contributions sheet, the share owed to each source, and a ledger that settles it. The royalty work extends the same idea to images, an attribution matrix built step by step through diffusion. The claim underneath it all is one I make in Attribution Decay: when a thousand sources each contribute a little, "nothing is responsible" isn't an answer. It's overdetermination, and overdetermination isn't absolution.

2. Memory belongs to the person. Totem and Thread moved storage out of a central cloud and onto nodes people own. Users decide whether their data joins the network at all. STTP pushes the thought to its end: devices prepare semantic tokens as things are created, share them peer to peer, and a shared aggregation layer, not one monolith, predicts the completion. A slogan from one of my first jobs keeps coming back: "Big brains, small devices."

3. Agents act with receipts. Mary, and Ambient as its macOS vessel, answers three questions most assistants skip: where does the screen go, where does the memory live, does it act or advise. It reads the accessibility tree in-process, so pixels stay home, and every action comes back as a validated step with a receipt. Ambient Abilities goes further and shows the router's work before it runs, instead of asking you to trust it. Software that can drive a real screen is safest when it's open, because then the code you can read is the code that runs.

4. Safety happens at the source. CLR tries to read truth out of the pixels themselves. Scorpion and Veil are meant to run before a model ships: one a memorization check at the upload door, the other closing routes to a likeness. Detection alone is an autopsy. The point is to move the checkpoint earlier.

5. The A.I. scales with the person, not past them. The personal-A.I. thesis is that a tool should ride just above your current craft and widen with it, so what it makes carries your earned work and not a prompt "vibed to completion." That's also a statement about hiring and identity: a way to communicate capability without sacrificing the person.

Five projects, one pattern: legibility. The thing can be asked, and it answers with a receipt.

What isn't proven

An honest accounting matters more here than anywhere, because a post about legibility that hides its own gaps would be a joke.

Most of this is early. The payout ledger I've shown settled a single response, with a single owner. Detection has limits I've published myself: newer generators have to be covered with data, and precision is a slow climb, not a switch. STTP is a question I've been circling, not a result. A registry of protected people is only as good as who gets to be on it. The Linux convergence I keep mentioning isn't public yet.

I'm saying so out loud because the old defaults rely on you assuming the best. The new ones have to survive you checking.

If it all ships

Here's the forecast. Take it as a bet, not a result.

The receipt becomes the default unit of an AI product. Right now, "why did you say that?" and "who contributed to this?" are things you ask a person. Once a few products can answer them, the ones that can't start to look like the fog they are. I think within a few years it will feel as strange to use an assistant that can't show its work as it feels today to use a website with no HTTPS.

Creators get a line item for being used, not only for being found. The old internet bargain was that your work got traffic in exchange for being indexed. Generative systems broke that balance, which is the whole argument of my Fair Use post. Attribution at the layer before serving turns the dispute from litigation into infrastructure: not a lawsuit after the fact, but a payout at the moment of use. It won't be perfect. It only has to be better than nothing, and nothing is the current price.

Intelligence gets smaller and more numerous. My strongest bet is the least fashionable one: that millions of small models, each tuned to a person or a field and aggregated by consensus (a Product of Experts, not a monolith), end up better than a few giant ones, and that the frontier model becomes one expert in the room instead of the room. "Decentralized intelligence centralizes growth for the whole" is still my best summary.

The machine you own becomes the primary site. Memory, retrieval and action move home, and the cloud turns into a peer you call when you need more, not the center you report to. That changes who a product is accountable to, because software that runs on your machine can be inspected by you.

Verified craft becomes a language for hiring. This is the one I care about most as an engineer. Hiring right now is a fog problem: employers can't measure value, newcomers can't express it, and so everything gets read from proxies. If your work carries provenance (what you made, how, what it drew on, in public, with a record), then the unasked question has nothing left to feed on. The work answers for itself, and the person never has to defend a self that was never on trial.

Safety moves upstream and gets argued in the open. Checks at the hosting layer, registries that people can opt into, adapters that close routes to a likeness, watermarks that let on-device users prove provenance. None of it is airtight, and all of it will be attacked. But "we'll fix it after it spreads" stops being the only option.

Companies start to look like their products. This is the cultural part. A team that ships receipts can't comfortably run on whispers. Boundaries still exist, since secrecy for non-competitive disclosures is a real and respectable need, but they're stated: here's the edge, here's the reason, here's who can answer. The new Silicon Valley isn't a place without secrets. It's a place where the secrets have names.

How it goes wrong

Every new valley is one decision away from the old one, so here's the list of ways I'd expect this to fail.

• Receipts become surveillance. A system that records everything it did can record everything you did. Legibility has to point at the machine, and stay out of the person.
• Attribution gets gamed. The moment a payout exists, someone will manufacture the contributions to claim it.
• Local-first excludes people. Not everyone owns a good device. A world of owned compute needs a plan for those who don't.
• Decentralization fragments. A thousand small nodes that can't agree on anything are just a thousand silos.
• A registry becomes a gatekeeper. Whoever decides who's protected holds real power, and that power needs the same audit as the models.
• The incentives quietly return. The oldest failure of all: a better design with the same business model underneath it drifts back to the same behavior.

That last one is why it comes down to guarantees: say what's being evaluated in advance, say what gets recorded, let the person read it. The test of a new valley is the same as the test of a new workplace. Not whether it says it's different, but whether the guarantees are written down.

The point

I used to think the next valley would be defined by a technology: a model, a chip, a protocol. I don't anymore. The research has taught me the defining feature is much plainer. It's whether, when you ask a direct question, there's an answer with a name on it.

The old valley optimized what it could measure and hid what it did. The new one is whatever makes the doing legible. Where it came from. Where it went. What it did. Who it drew on. Who got paid. Who can ask, and who has to answer.

I'm building toward it on the side of everything else, in public, and I'm happy to be wrong where it can be seen. That, more than any architecture, is the culture worth building.