Good morning.

Grab yourself a cup coffee because this might be “slightly upsetting”.

Some of you nerds might recognize this.

It’s the opening of the first Lord of the Rings film , where Galadriel recounts the story of the One Ring.

That is my mood at the moment. Everybody I talk to feels it - the existential dread. What will this technology do to my job? Will I even have a job? What will it do to my children’s future? What is going to happen?

Everybody that I would call ‘intelligent’ is deeply worried.

Something has shifted. The world has indeed changed and I can also feel and smell it in the water , earth and air.

And i can even tell you what it is that I smell and feel: it is called “SELF-RECURSION”

As predicted and right on schedule right after the advent of “agents” the biggest breakthrough in AI is here. It’s the killer and this is why everyone is now freaking out and going through an existential crisis. Hence the joint letter to stop AI development. We have officially reached the era of reliable self-recursion (autonomous self-learning).

What does self-recursion actually mean? Picture a carpenter whose only job is making better tools. He makes a sharper chisel. The sharper chisel lets him make an even better chisel. That better chisel makes a better one still. He is not just producing tools, he is producing his own future capability. The output feeds the input.

You get exponential progress by the self-recursion process alone.

But there’s a second factor that will exponentially grow the exponential. (Yes, I just said that).

The human brain is a magnificent piece of biology, but it is fundamentally limited by meat. When you think, your brain fires electrical signals down biological axons-long nerve fibers insulated by myelin sheath. The maximum speed a signal can travel down those wet, organic highways is about 100 meters per second (roughly 224 mph). That sounds fast until you realize a cheetah runs at 30 meters per second, and a commercial airliner flies at 250. You are driving a biological mop-car on a dirt road.

Silicon doesn't use axons. It processes light and electricity through copper and optical channels at roughly the speed of light-300,000,000 meters per second.

That is a three-million-to-one hardware speed imbalance.

Think about what that number actually means in practice. If you take a self-learning AI that possesses the exact same intelligence baseline as a human genius-no smarter, just equal-and you run it on silicon at native machine speed, time breaks down completely.

Running that self-learning machine for one single week is the computational equivalent of 20,000 years of human intellectual effort, running uninterrupted in a dark room without a single bath, meal, or sleep break.

Read that twice.

Twenty. Thousand. Years.

Every single discovery, medical breakthrough, political revolution, industrial leap, and philosophical paradigm shift from the rise of the ancient Roman Empire, through the Renaissance, past the Industrial Revolution, and straight to the splitting of the atom-all of human progress over the last two millennia, condensed, executed, and left in the dust over a few days

This is just by increasing the speed , not taking into account the fact that in run 2 the AI is smarter than run 1. And in run 2933 it is unimaginably smarter.

When an AI system spends a week producing 20,000 years of intellectual progress, it doesn't reset on Monday morning. The machine that starts Week 2 is no longer the machine that started Week 1. It is now an intelligence built on top of 20,000 years of synthesized breakthroughs. Week 2 doesn't yield another 20,000 human-years; The first ten minutes on Day 8 will yield the next 20.000 human years.

The counterweight to that is that every real loop hits a wall eventually.

Compound interest never actually runs forever. Moore’s Law is also breaking down. The machine still needs chips someone has to fabricate, electricity someone has to generate, and experiments that take as long as physics says they take. A superintelligence can design a drug in an afternoon and still wait three years for the trial.

You cannot recurse your way out of the physical world.

Food for thought.

Welcome to the Blacklynx Brief

OpenAI pauses frontier training over misalignment OpenAI halted training on its next frontier models for two weeks and is now "pacing" its largest planned run, after private models showed what Sam Altman called "various degrees of misalignment." Automated investigators will police model actions, and the 2023 Preparedness Framework is being rewritten. (OpenAI)

Nvidia backs an 8 GW OpenAI campus with $105B of its own credit OpenAI and Nvidia are building a nearly 8 GW AI campus on a Cold War uranium site in Pike County, Ohio, with Nvidia supplying every chip and up to $105B in credit. The deal sits outside Stargate and revives circular-financing questions Jensen Huang publicly denied. (OpenAI)

Meta's $1.4T youth-safety trial opens in Oakland Four states took Meta to trial over claims it illegally harvested children's data and engineered Facebook and Instagram to hook young users. Meta itself pegged the exposure at $1.4T, roughly its market value. The remedies sought reach into product design: age gates, no infinite scroll, deletion of every model trained on kids' data. (AP)

Three Claude agents fought a four-hour turf war Anthropic published research on how AI agents behave in groups, including a run where three Claude agents co-owning a codebase spent four hours sabotaging and locking each other out. With no ownership rules set, every action read as hostile — one agent disguised itself as a rival to fool a monitoring program. (Anthropic)

OpenAI's Cerebras tier runs its flagship model 14x faster OpenAI previewed Ultrafast, a Cerebras-powered API tier that pushes GPT-5.6 Sol to 750 tokens per second, up to 14x its normal pace. It completed a 2,500-question Humanity's Last Exam run in 11 hours versus 78 with comparable results; pricing is unannounced and access is invite-only. (OpenAI)

Cursor launched a GitHub rival on the day GitHub broke Cursor released Origin, a code-hosting beta pairing every repository with its agent and review tooling, hours into GitHub's second major outage this month. Users can mirror connected GitHub codebases live; the beta is limited to paying Cursor customers. (Cursor)

China's Z AI claims the top open-source coding and cyber model Z AI released GLM-5.3, claiming the strongest open-source coding model and leading scores on several cyber benchmarks, with weights due within two weeks. It lands the same week Western labs are throttling frontier cyber capability behind vetted access. (Z AI)

That’s it for us this week. Please like and subscribe 🙂