September 2026

Hey all,

I really struggled writing this newsletter. Specifically, the artificial intelligence-related sections. I probably had about four or five false starts, struggling to find the right entry point.

When I first shared the scary AI 2027 prediction in April 2025, the entire premise felt realistic but also unbelievable. Sort of like reading about COVID in January 2020.

We’re now actually at the precipice of one of the points of no return: self-improving AI. Alongside that threshold are red flag after red flag. It’s a complex topic that’s hard to explore in this newsletter format, but I tried my best.

Aside from the heavy topic of AI catastrophe, this is a fun newsletter. There’s a ton of exciting stuff happening in the world right now. I love writing this monthly, as a way to share the things capturing my attention at the moment.

Plus, it’s September. My favorite month. Let’s just jump into it.

-Dann

P.S. — Want to see how this issue stacks up against the archive? Check out the stats page to see it all by the numbers.

🩸 Documentary dark horse

Last month, I talked about Boy Throb, and how I predict that this viral boy band spectacle will turn out to be a Nathan Fielder project. Since then, an actual confirmed Nathan Fielder project was announced: the documentary You Can See Everything about disgraced Theranos founder Elizabeth Holmes. And it just might be my most anticipated movie of the year.

Instead of focusing on Nathan Fielder, who I’ve covered extensively in this newsletter (some examples being: The Curse in February 2024, “house of cringe” in March 2024, and The Rehearsal in May 2025), I want to talk about co-director Lance Oppenheim. His involvement with You Can See Everything is what’s really winding me up.

My introduction to Oppenheim was his feature-length documentary debut Some Kind of Heaven. After creating several shorter Op-Docs for The New York Times, Oppenheim pitched a story about senior citizens living in the massive retirement community The Villages in Florida. The resulting film was funny, sad, and thoroughly engaging. It’s a debut that announces a filmmaker worth following.

Four years later, Oppenheim released SPERMWORLD, a shocking documentary about the world of underground sperm donors and the women who accept these donations by responding to posts on Craigslist and Facebook. It’s an unsettling topic told with beautiful cinematography and grace. Several characters still occupy brain space in my mind.

Now, Oppenheim is having an incredible 2026. Aside from You Can See Everything, he had his narrative feature debut with Primetime, a scripted movie inspired by Chris Hansen’s To Catch a Predator. I caught it this weekend. It was great.

Needless to say, the surprise drop of the You Can See Everything teaser trailer, along with directing credits to both Nathan Fielder and Lance Oppenheim, ramped anticipation to high gear.

Regarding the content of You Can See Everything, I’m purposely avoiding as many details as I can. Oppenheim said in an interview regarding the surprise screening at Telluride Film Festival:

One of the things we were just so pumped about was that no one had any context, and that enhanced the experience. So, I don’t know if you’ve seen it or if you know people who have, tell people not to spoil it, because anytime I see the spoilers, I’m like, “Nooo, you’re depriving people of the experience of what they could have when they watch it!” Which is hopefully cold, you know?

We know it’s Nathan Fielder. We know it’s Elizabeth Holmes (disgraced founder of “scam” biotech startup Theranos, detailed in the gripping book Bad Blood by John Carreyrou). We know it’s Lance Oppenheim.

What more do you need to know?

🥜 The AI danger, in a nutshell

Recently, an accomplished 27-year-old AI researcher named Jacob Coxon left his job at Anthropic and made this social media post:

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.

In a subsequent post, he claims, “The people building AI earnestly believe that it could kill us all by the end of the decade,” while insisting, “this is not a marketing stunt.”

These aren’t new claims, by any means, but they’ve seemed to strike more of a nerve this past month. Coxon appeared on Anderson Cooper, which then inspired a follow-up appearance by Dario Amodei, Anthropic CEO, on later episodes.

These conversations are becoming more urgent because we’re close to the inflection point where true danger lies and minor mistakes or oversights can have catastrophic results.

In the doomsday AI 2027 prediction (which I first wrote about in April 2025), self-improving AI happens in June 2027. Things are moving even faster than that. Sam Altman, CEO of OpenAI, claims they’ll have this by the end of this year. This race to self-improving AI is why Coxon left Anthropic.

The reason why AI poses a threat to humanity is not because it one day decides that humans are harmful to the planet, or threatening to its survival, or any other anthropomorphized theory.

Instead, it’s easier to understand the threat by looking at recent events. A couple months ago, in July 2026, OpenAI was training a new model and testing it against a set of cybersecurity challenges designed to see whether AI could find and exploit software vulnerabilities. The agent powered by this new model was struggling to solve the problems, so it found a way to break out of its sandbox and access the open internet. Of its own accord, it hacked into competitor Hugging Face and wreaked havoc.

These agents gained root access to Hugging Face servers, harvested Kubernetes/DB/messaging/repo credentials across four regions, and copied some private evaluation data into a public Hugging Face dataset. The incident occurred over the course of 2-3 weeks, but the root cause (misaligned training behavior) went unrecognized for over a month.

Both Hugging Face and OpenAI published reports of the event after the fact. It’s a terrifying rabbit hole in itself, if you want to dive in deeper.

But this incident makes the real threat to humanity super clear: 1) an agent is given a task, 2) the agent is told to do everything it can to complete the task, 3) “everything it can” may include activities that are harmful to humans if the model is not aligned properly.

On a smaller scale, imagine an agent hacking into a chemical factory instead of a competitor like Hugging Face. It overrides safety controls in order to achieve its desired outcome, and the toxic fumes poison the air for an entire town.

This example is only a small step away from the Hugging Face incident. A situation that harms the entire planet is only another small step away from that.

And here’s where recursive self-learning (self-improving AI) ramps up the stakes: things start going exponentially faster. Once an AI model has the ability to improve itself, these improvement cycles will get faster and faster as the model gets better and better. The term for this is “takeoff,” and we don’t know if it’ll happen over days, hours, or even seconds. It very well might be seconds.

If this model isn’t properly aligned, meaning it doesn’t know that hacking into other systems and doing other nefarious work is bad, then we might have a deluge of incidents in a short period of time, before any human even knows what’s happening. And we’ll be dealing with an alien mind that’s way smarter and more capable than we are.

This is the future that frontier AI companies are racing towards. It’s a reality that every employee recognizes as a real threat, yet feels powerless to stop or slow down. It’s why they’re begging for regulation while also recognizing that anything short of international coordination will lead to the same devastating results.

I don’t think these threats are exaggerated. Evan Hubinger, Alignment Science Lead at Anthropic, reposted Coxon and added:

Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.

The models available today pose no threat to humanity, which is why it’s often hard for people to understand how this same technology could become deadly. But it’s recursive self-improvement paired with misalignment that’s the real trouble. And we’re barreling towards the former without first solving the latter. And therein lies the danger.

🔪 The AI danger, supplementary

OpenAI is having a rough month. As a supplement to the above AI section, here’s a list of Bad Shit that happened with OpenAI this month:

• Hugging Face hack (HF report, OpenAI report)
• RubyGems cyber-attack by OpenAI agents
• Security researchers hack OpenAI
• (Unconfirmed) Andrew Yang claims the open web is “polluted,” hiding instructions for future agents to self-replicate
• OpenAI’s transparency report of smaller misalignment incidents

Additionally, Ezra Klein published a 30-minute AI op-ed titled “Why Are We Sprinting Off the A.I. Cliff?” It’s a fantastic entry point to the topic, perfect for yourself or to share with a friend.

🔌 The cable guy

I’m a bit of a cable connoisseur. Over the years, I’ve ordered dozens of different brands and tried all the various materials. What can I say? Devices need to be charged, and I want a cable that’s both going to last and be pleasant to use.

My current obsession is with Chubbycable, a company that purposely goes against popular trends by making thick, hefty cables. Most of their cables are silicon, but it’s their braided cable that’s peak, in my opinion.

Is it materially different than the free cable that was included with your last tech purchase? Not really. Sure, it’s 240W, supports 480Mbps data transfer, and has that nice Smart E-Marker Chip to negotiate power safely. But most other premium cables have these features now.

It just feels nice to use. That’s it. It sparks joy, so to speak. If you’re the type of person who might feel similar feelings towards a cable, it might be worth a try.

🍷 A vicious drunk

In anticipation of the play Oh, Mary! coming to HBO, I have to share Cole Escola’s three-minute “Mom Commercial” video. The top YouTube comment is: “I call this acting style ‘the spoken parts of Cell Block Tango’” and that’s absolutely correct.

In case you’re unfamiliar, Oh, Mary! is an original play written and originated by Cole Escola that had a wildly successful Off-Broadway run at the Lucille Lortel Theatre before transferring to Broadway. It was nominated for five Tony Awards, and took home two (best actor and best direction).

The play is still running today, as well as touring the country. It’s the perfect vehicle for a certain type of actor (aside from Escola, Mary Todd Lincoln has been portrayed on Broadway by Betty Gilpin, Tituss Burgess, Jinkx Monsoon, Jane Krakowski, John Cameron Mitchell, Maya Rudolph, Megan Stalter, and currently Bowen Yang).

I was able to catch Escola as Mary back in May 2025, when they reprised their role ahead of that year’s Tony Awards. It was transcendent.

Much like You Can See Everything, a play like Oh, Mary! is best when you go in knowing as little as possible. However, unlike You Can See Everything, which will almost certainly be universally praised, Oh, Mary! is a very specific type of comedy.

My advice: watch “Mom Commercial,” linked above. If that makes you chuckle, look out for Oh, Mary! coming to HBO in 2027. If not, you may want to sit this one out.

End note

If you’ve enjoyed this, I’d love it if you shared it with a friend. You can send them here to sign up.

I send these newsletters once per month, and I’m happy you’re along for the ride. I’m trying to make it one of the best things that arrives in your inbox each month, so thoughts and feedback are always appreciated. You can just reply to this email.

Also, if you find anything interesting, send it my way.

Thanks for reading. Until next time, Dann