Toggle light / dark theme

Anthropic research warns AI could build itself by 2028

In this exclusive interview, Axios co-founder Mike Allen sits down with Anthropic co-founder Jack Clark to discuss his warning that by 2028, AI systems may be able to improve and build better versions of themselves.

Clark explains why Anthropic is preparing for the possibility of an “intelligence explosion,” how advanced AI could accelerate breakthroughs in science and medicine, and why governments, companies and researchers need new plans for cyber threats, bio risks, economic disruption and the future of work.

Timestamps:
00:00 — Introduction: the future of AI
00:41 — The 2028 prediction: AI building itself.
01:49 — The risks of rapid acceleration.
03:11 — The 3D printer metaphor.
05:21 — Intelligence explosion and fire drill scenarios.
06:55 — Building a \.

A Token of Our Imagination: The Invisible Economy Powering GenAI

Ever wonder what actually happens inside the AI after you hit “Enter”?

You type a prompt into your favorite generative AI, and within seconds, your screen fills with exactly what you asked for—whether it’s a quarterly report or a cinematic image of a cyberpunk golden retriever. It feels like absolute magic.

But behind that seamless curtain lies a bustling, microscopic economy running entirely on a digital currency you’ve probably heard of but might not fully understand: the token.

Most of us only ever see the input and the output. We don’t see the internal cash register ringing, the mathematical gymnastics, or the sprawling “assembly line” churning through billions of calculations.

What actually happens between the moment you hit send and the moment your final masterpiece appears? In my newest blog post, I peel back the curtain to trace the fascinating journey of an AI token.

I break down this invisible economy—from the “toll booth” of the input phase to the heavy lifting of the output phase—and show you exactly how the machine balances the books.


Mathematical framework solves asteroid route planning exactly for first time

A new publication from Bielefeld University sets a benchmark in optimization research. Together with an international team, Professor Michael Römer from the Faculty of Business Administration and Economics has developed a mathematical framework that solves a complex problem from space logistics exactly for the first time: the optimal planning of a route to visit several asteroids under conditions that are as close to reality as possible. The study is published in the INFORMS Journal on Computing.

At the center of the research is the so-called Asteroid Routing Problem. It addresses the question: In what order should a spacecraft visit multiple asteroids if both travel time and fuel consumption are to be minimized? The challenge is that, unlike in classical routing problems, the travel time between destinations is constantly changing because all celestial bodies are in continuous motion.

The idea for the study originated in Bielefeld, sparked by a success in a competition organized by the European Space Agency (ESA). During a research stay in Bielefeld, lead author Isaac Rudich revisited the topic and, together with the team, developed a new solution approach.

Time-varying magnetic fields can engineer exotic quantum matter

Quantum technology has promising potential to revolutionize how large and complex amounts of information are processed. While already in use primarily in laboratory and research settings globally, quantum technologies are in a transition phase for broader industry applications across many economic sectors.

In researching fundamental aspects of quantum physics, or the behavior of nature at the smallest scales—involving atoms, electrons and photons—a study led by Cal Poly Physics Department Lecturer Ian Powell analyzed how a changing magnetic field can make matter behave in unusual ways.

Powell and student researcher Louis Buchalter, who graduated with a Cal Poly bachelor’s degree in physics in 2025, published the article “Flux-Switching Floquet Engineering” in the journal Physical Review B, highlighting how changing magnetic fields over time can create quantum states that do not exist in any stationary material (remaining in the same state as time elapses).

OS Orchestration: Stepping Into a Frictionless Future of AI Sparks and Endless Abundance

There’s a very specific reason the tech giants are suddenly racing to get AI running locally on your phone, watch, and smart glasses.

The traditional Operating System (OS) is quietly being retired. Soon, the OS as you know it will be replaced entirely by an omnipresent AI hub.

But if the OS becomes an AI, what happens to that grid of static apps we rely on every day? And when the friction of swiping and searching disappears, how does the underlying economy of the Internet shift?

In my latest piece, I explore what happens next: the death of the app, the rise of dynamic AI “Sparks,” and a hidden token economy where your device doesn’t just cost you money—it generates it.

Want a glimpse at what your digital life looks like when you stop swiping and start orchestrating?


I have been on a breathtaking journey, for decades I have been watching how we connect with the world and each other. If you’ve been around tech long enough, you remember the humble hum of single twisted-pair copper wires, and the sheer, brick-like weight of early cell phones. Fast forward to today, and we are streaming the entirety of human knowledge over millimeter-wave antennas onto super-thin slabs of glass in our pockets.

Firehorse superstition helps uncover why women’s education may not drive Japan’s fertility decline

The rapidly declining marriage and fertility rates across developed East Asian societies strain pension and health care systems, threaten economic growth, and reshape entire societies. To tackle this issue, governments in Japan and across East Asia have invested heavily in pronatalist measures, but often with limited success. For instance, Japan’s government has repeatedly expanded childcare subsidies and parental leave provisions, yet the total fertility rate hit a record low of 1.20 in 2024.

A common narrative in media commentary, policy circles, and even within families is that women are “too educated” or “too career-focused” to marry and have children. However, the exact causal relationship between women’s education level and family formation is not well understood.

To fill this knowledge gap, a team of researchers from Japan and Singapore, led by Associate Professor Rong Fu from the Faculty of Commerce, Waseda University, Japan, used a novel quasi-experimental approach to understand the relationship between education, fertility, and marriage in Japan.

The Why Is a Discipline: Goodhart’s Law and AI

A reader asked me a question this week that I have been thinking about ever since.

She did not ask whether AI could malfunction. She did not ask whether bad actors could misuse it. She asked something sharper:

Can a system produce bad outcomes systematically, even when intent is good, and nothing is broken?

The answer is yes. And it is the most dangerous category of bad outcome, because nobody is at fault and nothing is broken.

We have all the evidence we need. Amazon ran into it. YouTube ran into it. Hospitals are running into it now. AI labs are about to run into it at a planetary scale. And almost nobody is talking about why.

A 1975 economic principle explains it cleanly. A reader’s question forced me to refine an argument I have been making for years.

New essay: [ https://www.singularityweblog.com/goodharts-law-ai/](https://www.singularityweblog.com/goodharts-law-ai/)

Neutrinos caught on camera: Testing the first prototype of a new elementary particle detector

Some innovations in physics come from entirely new technologies, others from fresh theoretical insights. Others still take shape by bringing together existing tools in new ways, working out how to combine them to outperform other solutions. The branch of particle physics that studies weakly interacting particles—such as neutrinos and some types of dark-matter candidates—could use innovative detection approaches: technological challenges in this research area quickly become practical as well as economic, as increases in detector volume and spatial resolution improve the sensitivity to the processes producing the particles of interest. Similarly, demanding targets on instrument capability apply to the calorimeters used in collider experiments.

Three-dimensional (3D) tracking of elementary particles in large-volume, dense materials is required in most particle physics experiments. In a scintillator, this is commonly achieved through fine segmentation of the material into many smaller active units, with each unit emitting light in the visible frequency range when a charged particle passes through it. Typically, the photons produced in every active unit are collected by optical fibers and carried outside of the scintillator to the photomultiplier tubes or silicon photomultipliers used for photon counting.

In the T2K neutrino-oscillation experiment in Japan, for example, one detector boasts about two tons of sensitive volume assembled from approximately two million cubes and 60,000 fibers. Over at CERN and the Paul Scherrer Institute, the LHCb and Mu3e experiments achieve sub-millimeter spatial resolution thanks to millions of thin scintillating optical fibers. With these figures, it’s clear that the scalability of this kind of scintillator material segmentation may turn into a bottleneck when larger volumes become necessary.

A humanoid robot sprints past the human half-marathon world record in Beijing race

The winner from Honor, a Chinese smartphone maker, completed the 21-kilometer (13-mile) race in 50 minutes and 26 seconds, according to a WeChat post by the Beijing Economic-Technological Development Area, also known as Beijing E-Town, where the race kicked off.

That was faster than the human world record holder, Uganda’s Jacob Kiplimo, who finished the same distance in about 57 minutes in March at the Lisbon road race.

The performance by the robot marked a significant step forward from last year’s inaugural race, during which the winning robot finished in 2 hours, 40 minutes and 42 seconds.

/* */