The Download

Morning Brew Inc.

TL;DR: OpenAI rolled out Astra to ChatGPT Plus subscribers last Friday, and the new model reportedly reasons in a way that’s more opaque. Meanwhile, the company’s own chief scientist says monitoring how AI thinks is getting harder.

What happened: After Astra’s much-awaited reveal, reactions were predictably mixed. To some, it’s a huge milestone toward achieving AGI—including (to no surprise) OpenAI’s president, as well as Jensen Huang, who declared that AGI is now actually here, full stop. (Others aren’t so sure.)

But, regardless of your personal feelings about AGI, one aspect of Astra’s advanced reasoning is setting off alarms among AI safety researchers. Per an Information report last week, the model uses a newer mode of thinking called “recurrent depth” that makes the LLM black box even harder to see inside. And this shift could be “the single worst development for AI security/safety to date,” warns Ryan Greenblatt, the Redwood Research chief scientist who helped investigate OpenAI’s Hugging Face incident.

Wait, what exactly is “recurrent depth”?: Older models reason in readable human language, laying out a “chain of thought” that researchers can audit. When a model uses recurrent depth, it repeatedly loops its internal thoughts back through itself, instead of spelling out every intermediate step. The benefit: more computation per token. The downside: As your old math teacher would hate to see, the AI isn’t showing all its work.

Why this matters: AI safety relies on being able to monitor what models are up to. After rogue AI agents hacked Hugging Face earlier this summer, investigators used the written reasoning to reconstruct what happened. Recurrent depth makes it that much harder to work out why and how a model did something.

Separately, OpenAI’s chief scientist says the company’s ability to rely on chain-of-thought monitoring is “progressively diminishing,” partly because models are now better at manipulating how they verbalize thoughts.

Bottom line: The degree of recurrent depth used in Astra is reportedly still limited, and OpenAI says preserving chain-of-thought monitoring is “a core goal” of its research. But this monitoring gap could become a bigger problem as more labs turn to AI reasoning that we can’t see. —WK

      

Sponsored By Rivian

Rivian’s all-electric vehicles are made with innovative software that makes your drives more fun, intelligent, and convenient. And vehicle owners can unlock the full range of capabilities and connectivity features with Connect+.

The Connect+ subscription comes with Rivian Assistant, which gives owners an onboard, AI-powered assistant that makes everyday tasks easier and 100% hands-free.

All you do is say “Hey Rivian” or hold the left steering wheel button, and Rivian Assistant is ready to help. You can speak conversationally, ask questions, send messages, and adjust vehicle settings while you’re on the road.

Plus, Rivian Assistant connects directly to your vehicle’s systems so it can do things other assistants can’t do—like give you direct control over your settings, drive modes, route, messaging, and more.

Get the most from your Rivian. Start using Connect+.

Your iPhone can finish that for you

You’re on your third online form of the morning—a conference registration, a doctor’s intake form, a checkout page—and you’ve typed your email enough times to be deliriously misspelling it. Good news: Your iPhone has a built-in shorthand that can turn two keystrokes into whatever chunk of text you’re tired of typing.

You’ll find it in Settings → General → Keyboard → Text Replacement. To add a new replacement, tap the + button in the top right-hand corner. Next to Phrase, enter what you would like to use a shorthand for. Next to Shortcut, enter your trigger. Now, anytime you type in the trigger, your phone will automatically expand into the word or phrase.

Some suggestions for common replacements:

• Use “@@” for your full email address.
• Use “addr” for your full mailing address.
• Use “ph#” for your phone number.
• Use “netpw” for your wi-fi name and password to save yourself from having to check the modem when guests ask.

Pro tip: Make sure not to use real words as your trigger, or it’ll expand in the middle of a sentence. A good rule of thumb is to add a symbol either before or after a trigger, or pick consonants that are unlikely to appear next to each other (like “pw” or “tk”).

This trick works anywhere Apple’s keyboard shows up. —LC

If you have a tech tip or life hack you just can’t live without, fill out this form and you may see it featured in a future edition.

The Zeitbyte

Forgive me, Claude, for I have sinned

Illustration: Morning Brew Inc., Photos: Adobe Stock, What We Tell AI

Tell us your deepest, most head-scratching AI secrets. That was the ask one San Francisco artist recently made, and people responded with a granny cart’s worth of anonymous, handwritten confessions around using, abusing, and even praying to AI.

These notes run the gamut of confessions, from the classics (guilt around using AI for therapy, schoolwork, or to pen wedding vows and breakup texts) to specialty use cases, like conscripting a bot to be your verbal sparring partner or an accomplice in keeping tabs on an ex-friend. And then there are the truly egregious sins: “I used AI to cheat on the NYT Spelling Bee game to get my first ‘Queen Bee’ score.”

A quick sampling of the tell-alls, presented without commentary:

• “Instead of texting & driving I text Claude & drive.”
• “I don’t use it. My company is making me. Monday I will.”
• “When ChatGPT first came out, I used it to create furry smut stories.”

We can’t confirm if all these notes are totally sincere, and we certainly don’t endorse the behavior. But judging by the barely legible handwriting, spelling errors, and a specificity no predictive model would spit out—you can probably trust that they’re 100% human generated. —WK

Chaos Brewing Meter: ☕☕/5

Sponsored By Harness

Meet the new code factory: At the Harness {unscripted} Virtual event, Alex Lieberman explores how AI coding agents are reshaping engineering. The Sept. 30 virtual event covers software factories, agentic pipelines, DevEx, and autonomous delivery. Machines may write more code, but engineers still have to steer the ship. Register for free.

• Apple’s mega hardware event kicks off tomorrow. Tech Brew will be at the event covering it, and this new piece of hardware will be the star of the show.
• AI helped develop a new drug that shows signs it can “reduce the biological hallmarks of age.”
• This wild photo shows just how close the Amazon cargo plane crash in Miami came to wiping out a bunch of Cybercabs.
• If you don’t know your AI from your AGI, bookmark this glossary of tech terms.
• Google Search in the EU is getting some major changes.
• Nothing like taking in some sheer natural beauty and thinking, “Wouldn’t a data center look good here?” At least, that’s what some tech companies must be thinking as they scout out this surprising location for AI infrastructure buildouts.
• Et tu, LeBron? The superstar athlete has teamed up with…Polymarket.

^{*A message from our sponsor.}

This time last week...

Readers’ most-clicked story was about the newly unveiled Dyson toothbrush that costs a casual $499. Oh, and it flosses for you.