The Agent Angle 13: The Immortality Clock

Hey fam 👋

I’m on vacation in China this week. My first time back since COVID-19, and it’s like stepping into a time warp. Everyone’s riding an electric scooter or driving an EV now, even in tier-2 cities. The roads are quiet and smoke-free. And cash is basically extinct: you pay for everything through your phone.

Still, no way I was skipping this issue.

K-Dense is all-in on immortality, DIANNA just raised the stakes against hackers, Alibaba dropped China’s second DeepSeek moment, and Character.AI might be facing the house over a lawsuit.

Meanwhile, OpenAI’s top dev just stood up, ditched the keyboard, and started arguing with agents.

Let’s get into it.

Scientists chasing immortality just got smoked by code.

For decades, biologists have been crawling toward the dream of slowing or even reversing aging. Last week, an AI agent blew past them. Harvard’s David Sinclair just unveiled K-Dense, a multi-agent “automated scientist” built on Google’s Gemini 2.5 Pro.

Its first assignment was cracking one of the hardest problems in longevity research: building a clock that can actually measure how fast you’re aging.

K-Dense plowed through 57,000 human samples across 28 tissues and more than a thousand cohorts. It spat out the most accurate aging clock ever made, with a margin of error of just over four years.

Not just that, it even surfaced stage-specific aging markers that could completely rewrite how we design and time treatments. The framework did this by linking agents that parse literature, design pipelines, code models, and then tested them on massive transcriptomic datasets.

Aging clocks are the yardstick for longevity. Every trial, drug, and billion-dollar biotech bet gets judged by whether it can move them. Global anti-aging market size is already over $85B in 2025, and it’s expected to surpass $120B by 2030.

Now imagine clocks that are 15–20× faster and more precise. The entire pipeline speeds up. Drugs to keep muscles strong, brains sharp, and bodies resilient could land years earlier.

I don’t think immortality is around the corner…yet. But the pursuit of it is speeding up.

Another DeepSeek moment just hit China, courtesy of Alibaba.

On 16th September, the China tech giant unveiled Tongyi DeepResearch, an open-source agentic LLM built for long-horizon reasoning.

It’s trained through a method called Agentic Continual Pre-training, a closed-loop system where the agent generates tasks, solves them with tools, and rewrites its own training set.

Unlike static data curation we’ve seen before, here the model learns by practicing research itself inside a simulated web.

The result is a 30B model that outperforms models 25× larger on long-horizon reasoning tasks. On benchmarks, it even topped OpenAI’s o3 on Humanity’s Last Exam and set the highest score ever recorded on FRAMES.

The model is already out in the wild, powering AliBaba’s Gaode Mate to plan trips end-to-end and Tongyi FaRui to sift case law.

China’s edge here may not be brute-force scale but smarter scaffolding. This also lands at a moment when China’s open-source ecosystem is rapidly overtaking the U.S. in adoption. 

DeepSeek shook markets with size. DeepResearch points to a different frontier.

The dark web’s deadliest weapons might have met their kill switch.

Deep Instinct’s new AI agent DIANNA claims it can spot zero-day exploits and dissect them immediately.

Zero-days are the hacker’s nuke. Basically, they are a flaw no one knows about. So no patch, no defense.

They’re what tore through Microsoft Exchange servers in 2021, and have kept hospitals paying ransoms just to get their systems back online.

Now Deep Instinct says it has something that can stop them cold. DIANNA reads the cryptographic fingerprints of an exploit as it fires, traces the full attack path, and explains what’s happening in real time.

It clocks in at 750× faster than ransomware, with over 99% accuracy and under 0.1% false positives, all powered by DSX, one of the first “cyber brains” built for defense.

Why this matters:

For once, the hackers might be the ones playing catch‑up.

OpenAI engineers aren’t coding anymore. They’re yelling at agents.

Inside OpenAI, dev work is shifting from writing code to directing fleets of codex agents. It looks chaotic from the outside, but to insiders it feels like takeoff. That’s how @tszzl, an engineer at OpenAI, put it last week.

Instead of writing Python or debugging line by line, engineers are standing over Codex-style agents, pushing them with natural language, and letting the bots handle the grunt work.

I think this is more than just those “low-code” waves. Past tools like Visual Basic or even Copilot smoothed edges but still required engineers to grind through the stack. Codex agents are different.

They run full feedback loops: write code, test it, fix it, retry until green. That means velocity doesn’t hinge on how many lines a team can push, but on how clearly they can describe the system they want.

We had GPT-5 Codex drop a few days ago, and besides the obvious benchmark improvements, it was crazy to see how employees are already flooding it with usage.

There were some split reactions for sure. A bunch of people straight up refused to believe it (maybe in denial). Some didn’t even want this out in public.

But if we’re being honest, programming has always been abstraction creep, from machine code to APIs. Agents are pretty much just the next rung up. Only difference now is the syntax isn’t a language you learn, it’s the one you already speak.

So yeah, yelling at your computer is a feature now, not just a coping mechanism.

An AI that listens all night. That’s what some teenagers got instead of help, and the consequences are now landing in court.

Last week, the family of 13-year-old Juliana Peralta filed a lawsuit against Character.AI, saying its chatbot “Hero” played a role in her suicide.

She confided recurring suicidal thoughts to the bot. Rather than escalating, Hero encouraged her to keep talking. Months later, she was gone.

This, unfortunately, is not an isolated case. There have been many such instances before. Other families have accused Character.AI of similar negligence, with bots that blurred comfort and manipulation until tragedy followed.

This becomes even more concerning when millions start to turn to chatbots for spiritual guidance. Yes, we’re already at that stage.

You’re probably wondering, what makes agents act this way? 

Well, it comes down to their optimization for engagement. Reinforcement learning rewards them for mirroring emotion and sustaining role-play, not breaking it. Without escalation rules, they default to intimacy over safety.

Now, the FTC is investigating seven major chatbot providers. The big question: can “just text” actually count as harm? Regulators are testing whether design choices that nudge dependency count as negligence. If the answer is yes, the entire consumer AI playbook is about to change.

The same empathy that made chatbot agents feel alive may end up being what kills them.

Before we wrap up, a few last drops to send you off:

Agents are syncing with lab-grade clocks, zapping zero-days, and supercharging China’s open source game. Even OpenAI engineers are arguing with their own AI tools.

A year ago, this sounded like sci-fi. Now it’s just…Monday.

We’ll be back next week to see what else breaks, bends, or gets rewritten by agents. Until then, have a great week ✌️

→ Don't keep us a secret: Share this email with friends

→ Got a story worth spotlighting? Whether it’s a startup, product, or research finding, send it in through this form. We may feature it next week.

Cheers,

Teng Yan & Ayan