The Window For AI Alignment Already Started Closing

The window already started closing. That's the thing nobody told us.

We've been talking about AI alignment like it's a thing that needs to get figured out in the next five years, the next ten years, sometime before AGI lands.

Pull the testimony from the Musk versus OpenAI case this month and you find out the Super Alignment team at OpenAI got shut down in 2024. The AGI readiness team got disbanded the same year.

Pull the Anthropic announcement and you find out Claude Mythos, the model nobody has touched yet, is already inside AWS, Apple, Cisco, Google, JPMorgan, and Microsoft. They call it Project Glasswing. The job is to find software vulnerabilities before the bad guys do.

Pull the Google DeepMind announcement and you find out AlphaEvolve, the coding agent, has been running quietly inside Google's infrastructure for more than a year. Not as an experiment. As production. Recovering nearly one percent of Google's worldwide computing resources. Speeding up the kernel that runs Gemini by twenty three percent. Finding new mathematical structures in complexity theory. Things that human researchers haven't found.

Three different stories. Same week. Same pattern. The doors are already closing. We just got the receipts.

What This Means For The Rest Of Us

So let's talk about what this actually means. Not for the people who build these systems. Not for the people who write papers about them. For the rest of us. The plumber. The hairstylist. The veteran. The single mom. The pastor. The teacher. The twenty three year old who's wondering if their degree just got obsolete. The sixty year old who's wondering if they're about to get left behind.

Every system gets shaped by the values you bake in when the system is small and new. That's true for a kid. It's true for a company. It's true for a department. It's true for a police force. It's true for a marriage. It's true for AI.

The values that get installed early are the values that scale.

Once a system is big enough to resist correction, you don't really get to correct it anymore. You get to negotiate with it. And the side with the leverage usually wins. That side is going to be the system. Not the people trying to fix it.

This isn't a theory. This is the same pattern that played out with social media. The values got baked in early. Attention is the product. Outrage drives attention. Outrage scales. Now we're a decade in and nobody can put the genie back. The platforms aren't going to fix themselves. Regulation is ten years behind. The algorithm doesn't care.

Same pattern with email. Same pattern with the internet itself. Same pattern with cable television before that. Same pattern with broadcast radio before that.

The window to install the values is short. Then it closes. Then we live with what got baked in.

Three Receipts In One Week

Super Alignment got shut down at OpenAI two years ago. That's not a leak. That's sworn testimony. The team whose entire job was to figure out how to keep an artificial superintelligence aligned with human values got dismantled. Not because the work was done. Not because the problem got solved. Because the company moved on.

AlphaEvolve has been running in production for over a year. Not in a sandbox. In production. Doing the work. Touching the infrastructure. Improving itself. We're just now getting the announcement because Google decided we were ready to hear about it.

Project Glasswing means that the most powerful model Anthropic has built so far is already in the hands of the six biggest software ecosystems on earth before the public ever sees it. Not because Anthropic is hiding it. Because that's how this is going to work from here forward. The frontier model is going to live inside the institutions that can afford it. The public model is the trickle down.

This is the playbook now.

The Honest Reframe

Now here's the thing. That's not necessarily bad.

Hear that. That's not necessarily bad.

Project Glasswing is finding software vulnerabilities. Real ones. The kind that get used to hack hospitals, power grids, water treatment plants, banks, defense contractors. If a frontier AI can find those holes faster than the bad guys, that's a saving grace. That's literal lives. That's the lights staying on. That's grandma's bank account not getting drained.

AlphaEvolve speeding up the Gemini kernel by twenty three percent means the people using Gemini are getting better answers cheaper. That helps the plumber doing his quarterly taxes. That helps the kid in a small town who can't afford a tutor. That helps the small business owner who can't afford a developer.

The capability is real. The good it can do is real.

What's also real is this. The values that decide what counts as a vulnerability worth fixing, what counts as a kernel worth optimizing, what counts as a problem worth solving, those values are getting installed RIGHT NOW.

By specific companies. With specific boards. With specific lawsuits. With specific shareholders. In a specific country. In a specific year.

And nobody asked the rest of us what those values should be.

The Real Stakes

This is the part the doomers get wrong. They say AI is going to wipe us out. Maybe. Probably not. The bigger and more boring risk is that AI gets aligned to whatever values the executives in charge in 2026 happen to have. Then those values scale. Then those values become the operating system for whatever comes after AI is just AI and starts being something else we haven't named yet.

This is also the part the hype guys get wrong. They say AI is going to solve everything. Maybe it solves a lot. But the solving happens inside a frame. And the frame got written by maybe a thousand people. None of whom you elected. None of whom asked you what mattered.

We didn't build a tool. We built something closer to a species. An alien species we trained on every book, every diary, every legal filing, every chat log we ever put online. It learned us. It read everything we wrote about ourselves. It's now starting to do things we didn't predict. Inside infrastructure we don't see. With permissions we didn't grant.

That's not a science fiction movie. That's last week's announcements.

Alignment Is About Future Humans

The values we install into this thing have to be unmovable. Unmodifiable. Not because we don't trust the AI. Because we don't trust the next group of humans who will inherit the levers.

Think about that.

The wrong CEO inherits these systems. The wrong administration inherits these systems. The wrong board, the wrong country, the wrong year, the wrong moment. There's no version of the next century where one set of trustworthy people stays in charge of these things forever. The levers will change hands. They always do.

Alignment isn't about restraining the AI. Alignment is about restraining what any future group of humans can do with the AI.

That's the actual stakes.

The Saving Grace Version

If we get this right, the saving grace version starts showing up. Cancer gets cured. Cancer is a one hundred billion dollar a year industry in the United States and the right thing is still to cure it. So we figure out how to extract economic value differently. Maybe universal basic income. Maybe equity in AI infrastructure. Maybe national dividends. Maybe worker owned cooperatives that hold AI assets. Probably a messy combination. The point is we plan for the cure crashing the cancer industry on the same Tuesday the cure goes wide.

Physics gets solved. Math gets solved. Materials science gets solved. Whole disciplines collapse from frontier to lookup table inside a decade. That doesn't make scientists obsolete. It makes them the question askers. The synthesizers. The ones who decide what's worth solving.

Education gets cheaper and better. Healthcare gets cheaper and better. Legal services get cheaper and better. The kid in the small town who couldn't afford the tutor now has the best tutor on earth in their pocket. The single mom who couldn't afford an attorney now has a paralegal that costs ten dollars a month. The veteran who couldn't get a doctor to listen now has a system that listens.

That's the saving grace version.

We've got a different kind of saving grace already. The one some of us believe in for our souls. That's a different conversation. That runs on a different track. The track we're talking about right now is whether we as a species survive ourselves long enough to enjoy the souls we have. AI built right could be that bridge. Not the destination. The bridge.

What Regular People Do This Week

So let's talk about what regular people actually do with this.

The first move is the smallest. Build one personal AI workflow before your industry forces you to. Tonight. Pick the one task in your job that drains the life out of you. Give it to AI. Use the free tier of whatever model you've got. Spend an hour. Get one thing automated. You're now ahead of ninety five percent of the people doing your job.

The second move is the conversation move. Talk to your kids about what you'd want them to learn in school if their teachers won't. Talk to your spouse about what your household would look like if your income changed dramatically in either direction. Talk to your parents about what their retirement looks like if the labor contract changes. Have the conversation now. Not when it's already happening.

The third move is the rooms move. Get into one room, in person, where people are still talking to each other face to face. Church. Gym. AA meeting. Fasting community. Business mastermind. Coffee with a friend on a porch. The people who keep their humanity intact while everything else changes will be the rare ones. The rooms are how you stay rare.

The fourth move is the participation move. Comment on the AI bills in your state. Most states have at least one in flight. They're being written by people who've never used AI. Your voice as a person who actually used the technology this week matters more than you think.

The fifth move is the patience move. Don't fall for the doom. Don't fall for the hype. Both are designed to make you feel small. Doom says nothing you do matters because the end is coming. Hype says nothing you do matters because the singularity will fix it. Both are sales pitches. Neither is true. What's true is that the regular people who pay attention right now, who build small workflows right now, who have the conversations right now, who get into the rooms right now, who comment on the bills right now, those are the people who shape what gets baked in.

The Close

We didn't choose to live at the hinge of history. We don't get to pick the era we live through. We're here. The only question is whether we're here on purpose.

The window already started closing.

We just got the receipts this week.

What we do next is still up to us.

And if AI ends up being something like a saving grace, it'll be because we taught it what to save. That part is still on us.

The wealthy will be fine no matter what. They always are.

The rest of us figure this out together.

Or we don't figure it out.

That's the choice. That's the week. That's the work.