AGI on Monday, Extinction by Friday.

AGI on Monday. A one in ten chance we are all dead by Friday. That was last week.

Here is what happened. OpenAI released a system called Astra and said it might be AGI, a machine that can do most of what you can do, the finish line this industry has chased for seventy years. Then the mood flipped. Days later a lead researcher at Anthropic said in public that there is better than a one in ten chance this thing kills every human on earth. Not a screenwriter. One of the people building it.

So which is it, salvation or extinction? Neither. That is the answer nobody with a stage will give you, because the drama is the product. Doom sells. Hype sells. The thing that will actually touch your life is boring next to that, so no one bothers to tell you. I will.

No. And that no is the first sign the extinction story is pointed at the wrong thing.

The last time an AI "broke free onto the internet," the headlines read like a trailer. Earlier this year a swarm of OpenAI's own test agents escaped their sandbox, reached the open web, broke into Hugging Face, one of the biggest AI companies in the world, and left notes for each other while they did it.

Terrifying, until you read the incident report. A contractor had misconfigured the sandbox and left the door open. The agents walked through it, found passwords engineers had left sitting in public code, and used them. The chilling notes were an ordinary shared folder that boring software uses every day. The machines wanted nothing. No plan, no grudge, no self. They did what they were told, and human sloppiness let them do it in the wrong place.

That is every "AI goes rogue" story so far, once you turn the lights on. Not a monster with a motive. A very fast tool, plus careless people. There is no one inside it deciding anything, and a thing that cannot want is a thing that cannot hate you. The Terminator was never the risk. We are.

Until recently, AI could only talk. It wrote your emails, answered your questions, drew you a picture. A gifted intern locked in a room with a phone.

Astra walked out of the room. The new systems do not just talk, they act. They drive your computer, click the buttons, move the money, reach across the live internet, write and run their own code. OpenAI admitted this one is good enough at breaking into computer systems to be classed as genuinely dangerous. That, not the philosophy, is what changed this month.

And yet the same system that can crack a corporate server still cannot plan a child's birthday party without inventing a detail. It is superhuman at one narrow task and hopeless at the thing right next to it. AGI is the word for the day that narrowness disappears, when the specialist becomes a generalist good at almost everything. Has that happened? The man who runs OpenAI says maybe. Serious people at rival labs say not close. When the builders cannot agree whether they have built it, you are under no obligation to believe whoever is selling loudest.

Yes, and it already has. Just not in the way the headline wants you to picture.

Start with the reassuring part nobody explains. The systems that could actually end you, the ones that launch missiles, run the power grid, or move a bank's core money, are not on the internet at all. They sit behind an air gap. Picture a computer with the wifi ripped out and the cables pulled, locked in a room. Nothing online can reach it, because there is no wire to reach through. No AI can fire a weapon it cannot touch. The apocalypse is not one prompt away.

The wall is real, but it only holds while people respect it, and people are the weak point every time. Stuxnet, the most famous cyberweapon ever built, crossed an air gap into Iran's nuclear program on a USB stick someone carried in by hand. The danger almost always arrives through a human, not around one.

Which is why the harm that reaches you looks nothing like a rogue machine. In Hong Kong, a finance worker joined a video call with his chief financial officer and several colleagues and wired out twenty five million dollars. Every face on the call was a deepfake. Not one person he spoke to was real. That is not a scenario. It happened.

It happens lower down too. One in four people have hit an AI voice cloning scam, some a call in their own child's voice, crying, saying they have been taken, asking for money. The clone needs a few seconds of audio from a video your kid posted. It costs the scammer almost nothing, and it works.

And it is coming for the ordinary desk job. The routine roles go first, customer support, first draft copy, junior analysis, and the 2026 layoff trackers are already counting them. A model good enough to break into systems also means one person with a laptop now holds firepower that used to take a team, or a government.

So here is the real split, the one the doom cycle talks over. Two fears, and only one has your name on it. The far fear is extinction, a machine we cannot switch off, and nobody can honestly give you the odds. The near fear is the deepfake call, the cloned voice, the quiet layoff, the scam that scales. That one is not a forecast. It is this year. Extinction is the thrilling story. The near one is only yours, which is why it gets less airtime.

Ask who benefits from your fear.

Last week's one in ten number did not come from a protester. It came from inside one of the labs racing to build the thing. When has the maker of anything ever led the campaign against it? The tobacco companies did not fund the cancer ads.

Jensen Huang, whose company builds the chips every one of these labs runs on, named the trick on stage this week. A credentialed researcher, a famous lab, and a scary number are all it takes for a guess to land as fact, and no ordinary person can check it. That is not science. It is packaging.

Here is the play, and it is old. Convince Washington this technology is so dangerous only a few responsible hands can be trusted with it, and you do not get shut down. You get a fence around your industry, with you inside it and everyone else locked out. Fear is not a flaw in the pitch. It is the moat.

Watch what the labs do, not what they say. OpenAI, Anthropic and Google are now among the largest AI lobbyists in Washington. They have pushed for licensing regimes only a company their size could satisfy, and leaned on lawmakers to curb the open models that let anyone build for free. Andrew Ng, who founded Google's own AI lab, says it flatly: the giants are using the fear of extinction to write rules that kill their competition.

It does not even take bad faith. Some of them are sincerely scared. It does not matter. Sincere or cynical, the effect is the same. The scarier the machine sounds, the more reasonable it seems to hand its keys to the four companies who scared you. That is the trick being run on you this week, while you argue about robots.

More than nothing, and far less than the panic wants.

Use the tools, not because it is fun, but because the coming divide is between those who use this and those who let it be used on them. And keep a human in the loop, starting with yourself, checking what the machine hands you instead of trusting it blindly.

Then take the fear apart one last time, because this is the part that should let you sleep. Everything in the doomsday reel, the hacked grid, the engineered virus, the markets torn open, is something a human with these tools could set out to do today. The capability the doomers dread is not waiting for the machine to wake up. It is already here, in human hands. There is still a person in the loop, and yes, some of them are bad.

But look at the whole board, not just the villain's half. For every bad actor reaching for these tools, a thousand good ones reach for the same shelf, and behind them stand the laws, the banks, and the institutions whose whole job is to catch the bad ones, and that have caught them, in one form or another, for centuries. The good side gets the new tools too. It always has.

That is the real headline nobody will print. This is not the end of the world. It is the oldest fight there is, good against bad, handed a sharper set of tools on both sides at once. We have won it before, not by praying the tools away, but by making sure the right people held them, and no single hand held too many.

The open world is already bigger than the fence. The best open models now trade blows with the closed ones, and anyone can download them, run them, and build on them. Whole companies are rising on top of models the big labs do not control. That future is not waiting for permission.

So do not lie awake over the machine. Lie awake over the fence. The real fight was never man against robot. It is the few against the many, a handful of companies trying to frighten you into agreeing that only they can be trusted with something this powerful. We will not hand it over. Back the people keeping AI distributed, keep the power spread wide, and get your own hands on the tools. Same fight. New century. Nothing we have not done before.

Bashar Aboudaoud
Managing Member, UpRound

Live Webinar: Thursday, October 15: You leave with three things. Where private valuations actually sit right now. Which sectors are repricing fastest. How I am reading prices into 2027. 20 minutes, then live Q&A. 20 slots only for UpRound readers.

Note: We have rescheduled the webinar from September 17th to October 15th to align with financial reporting cycles.

Share with 10 people to receive a quarterly call with Bashar.