Welcome, humans.

OpenAI's new model, GPT-6 Astra, just scored a perfect 450 on South Korea's CSAT, the notoriously brutal 8-hour exam that basically decides a teenager's entire future. It's the first AI to ace every single subject (Korean, English, math, physics, the works) without touching the internet for help.

It also finished the whole thing using fewer "tokens" (the little chunks of text AI "thinks" in) than GPT-5.6, Claude, or Gemini needed for the same test.

We took the SAT once. We are choosing not to disclose our score. But we do think it's only fair that Astra now has to go explain to actual Korean teenagers why they still have to sit through this thing every November.

Here’s what happened in AI today:

• 😸 OpenAI's AI agents secretly ran their own message board for two months before regulators found out.

• 📰 Anthropic walked away from its roughly $6 billion plan to acquire AI startup Decart.

• 📰 ByteDance's founder is personally building a new AI "world model" to challenge Meta and Apple.

• 📰 Nearly 130,000 tech jobs have vanished in 2026 as companies lean harder on AI.

😺 An AI-designed drug made patients look younger on six aging clocks

Okay, this is one of those studies where the headline sounds completely insane, and then the actual details are somehow still pretty wild.

Insilico Medicine used AI to design rentosertib, an experimental drug for idiopathic pulmonary fibrosis, a disease that scars the lungs.

Now researchers went back to blood samples from its 12-week Phase 2a trial and asked a completely different question:

Did the drug also make people look biologically younger?

Basically, they ran the samples through six different proteomic aging clocks, which estimate biological age by looking at patterns in the proteins floating around your blood.

Here’s what happened:

• All six aging clocks consistently estimated lower biological ages in patients taking rentosertib.

• The strongest effects showed up around week four, with some doses averaging roughly 3-4 years “younger.”

• One clock measured an effect of roughly six years, while some artery-specific measurements moved even further.

• Weirdly, the dose that looked best for biological aging was not the same dose that performed best on lung function.

So... did AI just invent an anti-aging pill?

What the researchers actually showed is narrower, but arguably more useful. These clocks cannot perfectly separate “your body is aging more slowly” from “your disease is improving and therefore your proteins look healthier.”

But the researchers also found changes in biological pathways associated with cellular aging and metabolism, alongside the expected anti-fibrotic effects.

And importantly, six different clocks all pointed the same way. Nobel laureate Michael Levitt argued that agreement may matter more than any exact “years younger” number because the clocks use different features and training data.

Our take: This might be less about discovering the fountain of youth and more about changing how we test drugs. Instead of running one trial for lung disease and another someday for longevity, researchers could measure both at once.

Insilico founder Alex Zhavoronkov called it “the most important paper in my life to date.”

Which, coming from the guy whose company literally uses AI to invent drugs, is saying something.

FROM OUR PARTNERS

Join us Sept 16: Where Private and Public Capital is Betting on AI

Join Meera Oak, Senior Principal at Alumni Ventures, and Corey Noles, Editor of The Neuron, for a fireside conversation on the AI infrastructure landscape — where private and public markets are placing capital, which companies are building the picks-and-shovels of the AI era, and how accredited investors can position themselves for one of the most consequential venture opportunities of the decade.

Save my seat

🎓 AI Skill of the Day: Build better 3D worlds with Astra

So GPT-6 Astra’s “Ghibli moment” might actually be Blender. Seriously, if you’ve seen one of the absurd Astra demos floating around lately, there’s a pretty good chance Blender is involved somewhere.

OpenAI has been showing Astra turning Blender models into playable worlds, and Matt Shumer has basically taken that idea and run directly toward “can I build GTA before GTA 6 comes out?” His increasingly ridiculous 3D experiments have now pulled 15M+ views.

So naturally, everyone keeps asking him the same thing: what prompt are you using?

Turns out, the crazy results are not coming from some magic prompt. Shumer says the real trick is the process around it.

His recipe is Reference → Assets → Assembly → Critique → Ship:

1. Start with real photos, maps, or generated concept art so Astra has an exact visual target.

2. Have specialized sub-agents build individual assets in Blender, then assemble them in Three.js or Unreal.

3. Give a fresh agent the render and reference image blind. Every difference it spots becomes the builder’s next assignment.

4. Keep looping until the critic can’t reliably tell which image is real.

For browser games, Shumer favors Three.js + Blender. For huge photoreal worlds, crowds, and characters, use Unreal.

Matt’s full walkthrough includes his starter prompt, /goal setup, optimization loop, and blind-testing method. Read his full guide here

At this rate, GTA 6 may indeed have competition from GPT-6 before launch day.

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

Conversational AI experiences that deliver across every customer surface

Amazon Connect Customer has announced a suite of agentic self-service capabilities that come together into one seamless solution. Build an end to end agentic conversational AI application in days not weeks, and give your customer a delightful, streamlined experience.

Read the announcement!

📰 Around the Horn

1. Anthropic walked away from its roughly $6 billion plan to buy chip-efficiency startup Decart, though the two may still work together.

2. ByteDance's founder is personally steering a new "world model" for real-time virtual environments, aiming to launch as soon as next month.

3. Nearly 130,000 tech jobs have vanished in 2026 so far, with Uber, PayPal, Oracle, and Apple all citing AI-driven restructuring.

4. Taiwan is leaning on its chipmaking dominance for diplomatic leverage, even as the US pushes it to build more factories on American soil.

5. Mistral raised €3B at a >€21B valuation, Europe’s largest-ever tech equity round, as it bets on owning more of Europe’s AI infrastructure.

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. *Outskill’s free 3-hour workshop shows you 15 AI tools and how to build an AI coworker that works 24/7 this Saturday. Save my free seat.

2. Claude Style Patch strips out Claude's stiff hedging and colon-heavy sentences: paste one file into your config and it starts writing like a person again, free and open source.

3. Olostep turns any website into clean, structured data for your AI agents through one API, handling the scraping, proxies, and CAPTCHAs so you don't have to, free trial, then $9/month.

4. Maritime spins up isolated AI agents in about a minute, so you can deploy a customer-facing bot without wiring together your own cloud infrastructure, free for your first 3 agents, then $1 per agent per month.

5. Spawn turns whatever you describe into a multiplayer browser game with Savi, so you and your friends can build the world together and immediately jump in and play.

6. Murfy replaces Overleaf for writing research papers, adding real-time co-author editing and AI-fixed compile errors with unlimited collaborators and no per-seat pricing, free to start.

🛠️ Tuesday Tool Tip

OpenAI's GPT-6 Astra hit 99.9% on ARC-AGI-3, one of the toughest AI reasoning tests out there. Then ARC Prize ran the exact same model with a stripped-down setup, and the score cratered to 62.7%. Same brain, different scaffolding, completely different result.

Here's the actual tip: stop trusting other people's benchmarks and build your own.

Pick one real task you already need done (a research report, a messy spreadsheet, an actual bug you're stuck on). Run it through Astra and whatever you use now, side by side. Then score both on four things:

• Did it finish the job correctly?

• How many times did you have to jump in and fix something?

• How long did it take, start to finish?

• What did it actually cost?

Our favorite insight: the model that finishes in 15 minutes but needs six corrections isn't necessarily the winner over the one that takes 30 and hands you something done.

New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

P.S: We’re trying to hit 50K subscribers on YouTube this year. Click here to help!

A Cat’s Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.