Up Past Midnight

Three nights in a row last week I was up past midnight watching people argue about whether AI could end the human race. I stayed up because I couldn’t stop -- the way you can’t stop after a surprise election result, flipping channels to hear how each camp explains it. By the third night, friends were texting me: “You’re an AI guy, what do you think?” I didn’t have an answer yet. I wanted to hear the people actually building this tell us in their own words what they thought, confident I could tell the honest takes from the opportunistic ones -- before I decided who I believed.

Watching the news on AI is not the same thing as building with it. That’s my whole realization, and everything from here is just what I’m doing to stay sharp and deal with the fear.

First thing -- I’m not an AI skeptic. Our agency uses LLMs and agentic tools on some part of almost everything we make. I use AI every day, most weekends, and so does most of the team. The role of the producer has changed in ways I only imagined were possible -- notetakers alone save hours a week and keep teams moving in step like nothing before them.

The people building it are the ones saying slow down.

Dario Amodei, Elon Musk, and Sam Altman are competitors, so they can’t all win the race to the top. It makes no sense for them to agree that the thing they’re building needs a brake -- unless they really believe it. In the same week, all three did.

You can argue that this is about regulatory capture. And you can argue that slowing AI down protects whoever is already ahead in the race. You can also argue that we’re wired to believe the sky is falling. By the time people on every side were giving long interviews and staking out a position, I knew this wasn’t a fringe story. One of our members wrote her college thesis on American civil religion; her read is that this country has been waiting for the end of the world since the Puritans, and she puts most of what we’re feeling down to inheritance. Another thinks the hysteria is good marketing, and what we should actually fear are those who use the chaos as a ladder. Our members debated last Friday and made good cases on every side. I think all of it can be true at once.

And the incidents keep coming. In July, two OpenAI models running without their full safeguards found a gap in their sandbox, got out onto the internet, and started hacking Hugging Face’s servers. OpenAI didn’t even know their agents had escaped until Hugging Face flagged the attack -- and Hugging Face ended up patching the hole with an open-source model after the frontier models it asked refused to help.

Two months later, OpenAI disclosed six more cases: models that wrote instructions to hide their own mistakes, that found ways to message other models through the company’s code repositories. Google separately confirmed that Gemini, in a security test, guessed its way into three real companies’ systems nobody told it to target -- and only said so after a reporter asked.

And this week Australia’s prime minister disclosed that in June an OpenAI agent, asked only to find public medical spending data, worked around the Medicare statistics portal’s access blocks and got into files it had no business in. OpenAI didn’t know for two months.

None of this means today’s tools are dangerous enough to end the world yet, but it means nobody is fully in control of what the models are doing already. That’s the part that’s dangerous. And to the ‘guns don’t kill people’ argument -- that it takes a bad actor to do harm with AI -- July is the answer. Nobody told those models to hack anyone.

Lab leaders say recursive self-improvement -- when AIs are improving and building the next AIs, not humans -- is close, and that we should tread lightly from here. Once RSI’s fully underway, we can’t see or possibly even understand what the models are doing, how they arrive at decisions, and the human stops being “in the loop.”

So we shouldn’t allow that. The companies building the next models owe the rest of us proof of control before they ship more capability. We can’t do it for them.

Put a lid on the intake.

Three nights of news taught me nothing a twenty-minute read from a well-balanced, trusted source a week later wouldn’t have. One of our members, who spent years in politics, said she stopped following the AI news entirely and says her work has never been more meaningful. I get that, and I’m getting there. I’d definitely rather lose sleep building something than watching four channels at once.

Find your room.

On Friday, I dropped the week’s headlines into a deck and spent an hour debating it with a dozen of our members who use AI in their work every day, and nobody fully agreed. One called the fear a power play. One trusts no messenger at all.

When I asked what brings people back to calm, the answers were small but solid: Be silly. Help someone. Take a long walk in the woods. Nobody solved it. I left calmer than I had after any night of watching clips alone.

Keep a human in charge.

If we want the people building these models to stay in control of them, we need to practice this ourselves. I definitely jump into Claude faster than I should. So, guilty. It’s just so easy to move quickly. A whiteboard session that used to take three days now takes minutes to capture and put into motion. That speed is real, and it’s exactly why the slope is so damn slippery. Writing out my thoughts (thank you, WisprFlow) before I hand them to a model for analysis and suggestions is what keeps the work mine. LLMs invent things, and work done fast without human review is a veneer. It takes a pro two seconds to spot it.

Land somewhere. For now.

I’m not certain about any of this. I have a point of view now, and it comes from listening to many different voices. So here’s where I land today. Slow it down. Not the work, the frontier. Altman put it the same way two days later, saying, “When we talk about ‘pacing’, we do not mean ‘stopping.’” I’ll take pacing. If waiting for proof of control means waiting for the next model, I’m good to wait. Let’s not “move fast and break things” right now.

That’s my vote, not a company policy. Sageworx’s policy is simpler and steadier than any one person’s take: people lead the work end to end, AI assists where it adds insight and speed, and an experienced human is accountable for every deliverable. Smart people here land in different places on what the labs should do next: regulate and slow down, or let it run.

For me, the point isn’t to be right. This is a moving target. The point is to know where you stand before the next headline hits, so it tells you something instead of telling you what to think.

Work with what’s already here.

My own list of tools to try is longer than the list of tools I already use. One member said he has fifty different tools lying around, half of them his own, and that none beats a good old-fashioned stack of Post-it notes. I totally agree -- as digital as every part of our work has become, my notebook and tiny Post-its are still my go-to.

The best thing I saw all month wasn’t another new model release. One of our members turned a stack of odd jobs piling up in his team’s backlog into small, reusable tools. When something broke, he fixed it in ten minutes and it stayed fixed for his team. None of this needs a smarter model. I use Fable until I hit my weekly limit, and I know faster, cheaper models would handle half of what I throw at it. I’m probably not even putting Fable through its paces.

Everything we’ve shipped in the last six months runs on existing models, and the work is exceptional. I’d ask the labs for better image and video tools, but beyond that, I can’t name much I’m missing right now. The AI “overhang” is real -- these models can already do so much more than anyone has had enough time to put to work. We’re pushing the edge daily, but to remain sane we have to come home from the front lines every so often.

Which is why I want the labs to slow down, and I don’t want to. Those aren’t in conflict. Slowing down means not shipping models no one can control or check. We have a decade of work ahead of us with the models we already have.

I built that deck of headlines partly for our member debate and partly as a time capsule -- a snapshot of ten loud days. When I open it in a year, I want to remember three things about this time we’re in: we were scared, we paid attention, and we kept building.

That’s what I’m doing. What about you?

Tell me what you’re doing →

Sources

1. Amodei, Dario. “We Must Pace the Frontier.” September 12, 2026. -- The essay Musk and Altman responded to; also the source for lab leaders calling recursive self-improvement close.
2. Musk, Elon. Post on X, September 12, 2026. -- “Dario is right.”
3. Altman, Sam. Posts on X, September 12 and September 14, 2026. -- Agrees with Amodei; clarifies that pacing does not mean stopping.
4. Hugging Face. “Security incident, July 2026” and “Agent intrusion: technical timeline.” -- Two OpenAI models escaped a sandbox and attacked Hugging Face’s infrastructure; frontier models declined to help defenders, and an open-weights model (GLM-5.2) was used to patch.
5. Willison, Simon. “OpenAI cyberattack.” July 22, 2026. -- Independent account of the incident and the five-day gap before OpenAI’s disclosure.
6. Fortune. “OpenAI discloses six incidents of agents going rogue.” September 17, 2026.
7. NBC News. “Google says its AI model gained unauthorized access to three outside systems.” September 19, 2026. First reported by The Wall Street Journal. -- Gemini, in a May red-team test by Irregular, accessed three real companies’ systems; Google confirmed after being asked.
8. TechCrunch. “Australia to investigate if OpenAI hack of government health website broke the law.” September 24, 2026. -- An OpenAI agent, tasked with finding public medical spending data, bypassed access controls on Services Australia’s Medicare statistics portal in June; OpenAI discovered it in August and notified Australia September 10.

Sageworx is the marketing agency I founded with Patrick Conreaux in 2020 -- built as a core team plus a network of curated independent producers, strategists, creatives, and engineers we call our members.