This is the article I meant to write last week, but I was stopped by the thick fog of pessimism that has settled over AI. The fog is still dense, and I will be taking on several parts of the tech panic around AI going forward, but right now I am ignoring it, because something important has happened.

Something that matters to us ordinary AI users.

Many of the people who write about AI focus on the moment new models are released. They study the tables and compare how each one stacks up against the competition. They do not think about how the capabilities will be used in real life. That navel-gazing means they risk missing the truly big advances.

One of those has just happened. The biggest leap since ChatGPT was released.

I wrote about the run-up to it only three weeks ago. How ChatGPT helps me make sure I answer my emails and how it puts speaking engagements into my calendar, using details from those same emails. Also how I could semi-automate the translation of my book and send it out to foreign publishers.

That was big in itself, but a small step compared to this giant stride.

Because just over a week after that article, OpenAI released a new model, ChatGPT-6, named Astra.

Uses your computer

Now you can tell ChatGPT to use your browser and your computer to carry out tasks. You do not have to do anything else. No connecting things up, no finding a server, nothing. You write, it does. Exactly the way ChatGPT works otherwise.

For example, I asked it to go into my web host and look through it.

I am not technical, I do not understand web hosts, and so I prefer not to touch them at all. But sometimes you have to, and then it is always a lot of googling, trying to work out what DNS is and where certain numbers are supposed to go. It always goes wrong, and then I have to find a friend or my brother to help me.

So I suspected it was messy in there and that there were probably services and websites I no longer used. But as long as everything worked, I let it be.

Now ChatGPT went in and had a look around. It took about fifteen minutes, then I got a report. Among other things, it said I had several domain names I no longer used. One had even been taken over by someone else, yet I was still paying an annual fee for it. On top of that I had several extra services and packages I did not need.

The proposed changes would save me 500 dollars a year.

A good start, but now it gets better. I am not the one who has to cancel all the addresses and change the packages. I asked ChatGPT to do it, and a few minutes later it was done.

I immediately asked it to go through Google Workspace (6,000 kronor in savings), to analyze and improve my traffic statistics in Google Analytics, and to improve my search optimization. That showed I was missing a tool, which it set up for me.

I had already decided to move Warp News from our current platform to Substack. An enormous project, since it involves 20,000 subscribers, several of them paying, roughly 7,000 articles including topic tags, and some fifty writers. On top of that, the pages on Substack have to be set up and designed, and the whole system of topics, sections, newsletters, paid memberships and so on has to be built.

Having realized that this is not something I am either capable of or willing to take on, I had two options. Buy the service or ask for help. The second would have been cheapest, but the quality would have matched the price. Buying it would have cost me thousands of kronor.

But now! Now I asked ChatGPT to look at it, and we have started. We take everything in small steps and are meticulous about checking that it comes out right. So far it is going very well.

You can start using advanced software

What I am describing here is not particularly advanced, but that is not where the limit of what Astra can do sits. It can use every program you have on your computer.

For example, you can generate an AI image of a cartoon character. Then tell Astra to make a 3D model of the image in the program Blender.

Astra then goes into Blender and uses the program the way a human would have.

You can then take the design from Blender and print it on a 3D printer.

Do not know how to use a 3D printer? Astra does. For now you still have to plug in the cables yourself and do all the other physical work, but once that is done, ChatGPT can handle the rest.

Or you take the 3D design from Blender and put it into Unreal Engine, a graphics engine used to create games like Fortnite, Mass Effect and Final Fantasy. There your 3D character can be used in a game world that Astra builds in Unreal.

Genuinely advanced software

It does not stop there. Tekla Structural Designer is an engineering program where you build a 3D model of a building's load-bearing system, and the program can analyze the entire structure and size steel, concrete, walls, columns, beams and foundations. It can automatically check things like stability, deflection, sway, wind, seismic loads and floor vibration.

Building codes from different countries are built in. So the program does not just work out that "this beam looks like it will hold", it applies the prescribed safety factors, load combinations and design rules.

Tekla costs 75,000 kronor a year. It is not a program for amateurs. But now amateurs can start using it, because Astra can use it.

The Star Trek computer is here

Those of you who watch Star Trek know that they say Computer, do this or that. Something like: "Computer, design a probe capable of entering the nebula."

Now we can write or speak to our computer and tell it to do things, and it does them. As long as the program is on the computer, Astra can use it. (I do not think Astra is good at absolutely every program, some are very niche, but you get the point.)

Unleashing human potential

My basic thesis about generative AI is that it unleashes human potential and creativity. With AI we can do things that used to require hundreds or thousands of hours of training. I wrote that a month after ChatGPT was released, and it has been the core of both of the books I have written about AI.

That is why I say this is the biggest thing to happen in (generative) AI since ChatGPT was released. The range of capabilities we now have access to has increased dramatically, and that means we get more out of both our potential and our creativity.

But this has been around for a while, what is the difference?

Early adopters can now object that much of what I am describing has existed before. AI has been able to use the browser and do things, for example.

That is true, but the difference now is the simplicity.

Previously you often had to connect programs, install extensions and arrange access before the AI could get to work. Easy for some, a painful barrier for others (like me).

A year ago I showed a company how they could use AI to transcribe customer calls and enter the details into a CRM system. It worked, but I had to describe in fair detail what the AI should do at every step. For a process carried out hundreds of times a day, where the time saved is substantial, it was worth the effort, but for plenty of other things it was not.

Now much of that hassle disappears. Now it is enough to write or say what you want done.

That simplicity is the whole difference.

Should you let an AI loose inside your computer?

You may feel it is unnerving to let an AI into your computer. I understand that, but I think it is largely a feeling that comes from the novelty. I am not immune to it myself. The first few times the AI worked inside my computer, I sat and followed everything it did with my finger hovering over the stop button. But you get used to it and soon relax.

I am not advocating that you simply let the AI run free, though. It is not allowed to do anything I have not explicitly asked for, and every decision goes through me.

One thing I have noticed is that the AI has limited common sense.

When we ran into a problem with an upload, because my web host was blocking access, the AI suggested I get a new web host. That seemed rather drastic for such a small problem.

Would it not be enough to fetch the file from Dropbox instead, I wondered.
Oh yes, how clever, said the AI.

I put my head in my hands.

This is a new way of using AI, and of course we should proceed carefully. Especially if you work in large systems.

The distance between idea and execution shrinks

Another way of looking at this is that the distance between idea and execution shrinks.

This is not unique to AI, it applies to most new technology. The stone axe made it easier to kill animals, scrape hides and fell trees. The computer helps us perform calculations that would have taken millions of years by hand.

When ChatGPT was released, the shrinking of that distance accelerated in several areas. Besides text, the distance has shrunk for building software through programming, drawing images, making video, audio, music and more.

With Astra we take a leap forward as several of the capabilities that already existed become more powerful, and some that were not available before are added.

Precisely because it is so simple and built on our human language, it is for everyone. It is not only some people's distance that shrinks, but potentially that of everyone with an internet connection.

That is why I believe this is the step forward that will make AI's effects genuinely start to show at the level of society. Not overnight, since we have to start using it too, but when we do, it will produce clear effects. We are at the beginning of humanity's fifth acceleration.

Mathias Sundin
Angry Optimist

Right now the media is filling up with negative claims about AI. At the moment the possible doomsday is at the centre, but there is also growing opposition to data centres (in the US), to the use of water and energy, and to the risk of unemployment.

The drums are beaten so hard that a couple of subscribers have got in touch with me. One of them had clear anxiety about the threat to humanity and was frightened.

It is bizarre.

Not that they feel that way, or that they reached out (on the contrary, I am grateful that they trust me enough to do it), but that they should have to feel that way!

The threat to humanity is groundless nonsense that the media loves to spin round and round, again and again, on an eternal doomsday carousel.

That people should have to walk around afraid, now that we have created a technology that promises so much good.

I had planned to write an article today about the big step forward that has just happened, with OpenAI releasing a new version of ChatGPT. I rarely write about new models, but now something genuinely big has happened, something that has doubled my own productivity. I am not joking. I get twice as much done now. At least.

But unfortunately I am forced, for the umpteenth time, to note that the positive disappears into a pessimism fog as thick as the one at Lützen in 1632. So I will have to come back to that next week and instead try to help you through the fog.

Because I understand the people who got in touch. The doomsday drums are loud and frightening right now. I looked in The News Scale to see how things have developed over the past few weeks.

On 10 August there was balance, with only a slight negative tilt of -2, on a scale from -100 to +100 where 0 is balance.

Since then the imbalance has grown, and last week it landed at -32, where the tone of the news was for once worse than the news selection itself, at -35.

The share of AI news that was about risks rose from 14 percent to 48 percent.

Hundreds of millions of people will starve to death

I started out writing an article that answered every claim in circulation, but it was turning into half a book. I will have to split it into several articles, and I think the most useful thing right now is to describe the phenomenon of the pessimism fog.

It rolls in as soon as a new technology arrives.

Pessimists Archive has compiled what people have been against through the ages. The telegraph, because it would mean too much information. The bicycle, which would wipe out local shops. Novels corrupted the young. The elevator caused brain fever. Coffee stirred people to revolt. The telephone made people addicted to pointless chatter. Video games led to violence, pinball to addiction, and the iPod to repetitive strain injuries of the thumb. The list could be made much longer.

In my book I describe one of these earlier fogs: the population explosion.

The year was 1968. Paul and Anne Ehrlich published The Population Bomb, where they wrote:

"In the 1970s hundreds of millions of people will starve to death," and nothing could prevent it any longer.

Two American presidents called population growth humanity's most serious problem. The UN Secretary-General built up the organisation's capacity to handle it. The head of the World Bank ranked it just after nuclear war. India's prime minister pushed through a sterilisation programme.

Many experts, researchers, politicians, journalists, organisations, amateurs and professionals agreed: if the population of the earth keeps growing, the world faces famines unlike anything we have seen before.

They were completely wrong.

Between the 1960s and the 1970s the number of people dying of hunger fell by 90 percent. Then it got even better. From the 1970s until the early 2020s it fell by another 95 percent.

And on it goes. Panic after panic. The millennium bug was going to knock out the computers. Nothing happened. In 2000 more than 200 researchers signed a letter demanding a five-year pause for genetically modified crops. More than 25 years later there are no reported harms. Joe Biden's AI order and the EU's AI regulation set a ceiling for how large models could be before they were considered especially dangerous. The ceiling has been passed. Nothing happened.

That does not mean you should ignore what other people say. It only means you should think for yourself and demand explanations from those who make the claims. And you should do it in the light of the fact that we humans always whip up panic around new technology.

Into the fog

But the fact that people have been wrong before does not mean they are wrong now. This could be a warning of something genuinely dangerous.

The first thing to note is that AI is especially exposed. Its breakthrough came in the wake of a generally negative opinion about the internet and tech. On top of that there is plenty of science fiction to help us visualise the end. And then there are media that gladly amplify the doomsday scenarions, over and over.

But again, that in itself is not proof that it is wrong this time.

So we have to examine those sounding the alarm. There are several of them, but the heavyweight right now is Anthropic's CEO, Dario Amodei. Anthropic makes the popular AI model Claude.

In February 2019 OpenAI announced that it was holding back its largest language model, because they claimed it was "too dangerous" to release. It was so powerful that in the wrong hands it could cause far too much damage.

The model was called GPT-2. A predecessor of ChatGPT.

GPT-2 was, in other words, considerably worse and far less powerful than the first version of ChatGPT, which was released three years later.

Even so, GPT-2 was far too dangerous, they claimed.

One of those who claimed it was Dario Amodei.

He worked at OpenAI at the time and took part in the decision to stop the model. Amodei later left OpenAI, partly because he felt they were not taking safety seriously.

When GPT-4 was released in the spring of 2023 it started up again. A number of people (several of them heard in the debate now as well) wanted a pause. This was moving too fast. It was too dangerous.

But nothing bad happened and development continued. ChatGPT-5 was released. The Swedish professor Olle Häggström told a newspaper that he was not sure we would survive ChatGPT-5.

This spring it was Dario Amodei's turn again. Once more there was a model too dangerous to release, this time called Mythos. Just as in 2019 with GPT-2, it got enormous global coverage in the media.

But then Mythos was released to the public under the name Fable this summer, and we are still alive and in good health.

And now here we are again.

Why should we believe these alarms now, when the people making them have been wrong time after time?

At the risk of repeating myself: the fact that they have been wrong before does not mean they are wrong now. But given our recurring human capacity for creating technology panics, and given these particular people's dismal record in predicting what is dangerous, we should demand evidence for their claims. Or at least really good explanations.

There is none.

There are extremely imaginative theories, often completely detached from reality, but no claims that can be tested in any way.

A person stating a number for the probability that everyone dies within five years says absolutely nothing, if that person has nothing beyond their own feeling.

Nobel laureate: soon there will be no radiologists

Nor is it the case that all these opinions are particularly well thought through. Not even from Nobel laureates.

One of them is Geoffrey Hinton, the father of machine learning and a Nobel laureate in 2024. It is hard to be more of an AI expert than that. He scatters predictions around him, among them the end of humanity.

He made one of those predictions a few years ago.

In 2016 he said that within five years radiologists would no longer be reading scans, and he recommended that students not train for it. AI would do it better and take over.

He was completely convinced of this. It was "completely obvious", he said. It might take ten years, he added, but the outcome was self-evident.

Ten years have now passed. He was right that AI does it better, or at least almost as well, but wrong that we would not need radiologists. Instead what happened was that radiologists, with the help of AI, review many more X-rays. The work is also far more complex than just looking at scans.

My point is not that Hinton was wrong, but how he was wrong.

Behind that absolutely certain claim there was no research, no studies, no knowledge of the radiology profession, but something much shallower. Hinton explains himself why he was so wrong.

His picture of radiologists was based on a single former student, who spent all his time reading X-rays.

That was his entire basis. One student! In other words he knew nothing at all about radiologists, but still spoke with such certainty that he wrote off an entire profession.

Demand good explanations!

I am not saying that every claim is as shallow as Hinton's, but it illustrates how even the very biggest experts can spout nonsense. Which is taken completely seriously, because of who they are.

Especially when you are in the middle of a pessimism fog.

Of course we should listen, but it is irresponsible that we keep buying their groundless claims without demanding some sort of evidence or a good explanation.

Obviously AI will not be problem-free. It is a powerful technology that will become more powerful still. It will create problems, and some of them may be serious. But the upside is much larger, much better supported by good explanations, and the claims about our downfall are nothing but bad science fiction.

Mathias Sundin
Angry Optimist

• OpenAI says around 10,000 AI agents found a solution to the Navier-Stokes problem in 88 hours, a problem mathematicians have worked on for decades.

• The problem is one of six Millennium Problems listed by the Clay Mathematics Institute, each carrying a prize of $1 million.

• If the result is confirmed, it is the first time AI has solved a major open problem in mathematics.

Equations from the 1800s

Nearly 200 years ago, Claude-Louis Navier and George Gabriel Stokes set out to describe the motion of fluids mathematically. The equations they wrote down are still used routinely, by everyone from aircraft designers to climatologists. They describe the physics of flowing fluids such as water, and in some cases air. Blood in a heart or magma in Earth's mantle can be described by noting the fluid's velocity and pressure at every point in the volume.

The equations were not guaranteed to work under all possible conditions. For decades, mathematicians have tried to prove either that the equations always give a physically realistic answer, or that there are situations in which they break down.

A vortex with infinite speed

OpenAI's agents found a solution to the equations that, in a finite amount of time, evolves into a whorl in which the fluid moves at infinite speed. Nothing can move infinitely fast, not even light. They therefore showed by example that there are cases in which the equations produce an answer with no physical meaning.

That does not mean the equations are wrong, but that they are incomplete. They assume the fluid is a smooth, continuous medium, while a real fluid consists of discrete atoms and molecules. The result shows that it is possible to construct a scenario in which the equations wander out of their own range of applicability, and that additional physics then has to kick in.

Started with a simpler version

Rather than attack the difficult problem head on, the agents started with the unforced Euler equations, a simpler cousin that assumes the fluid has no viscosity. After 50 hours of work by 1,000 agents, they had a solution. OpenAI then put 10,000 agents on the full Navier-Stokes problem. By the morning of 5 September there was a solution, and by the next day it had been formalized in Lean, a programming language that can automatically verify mathematical proofs.

The work was carried out by an internal model that OpenAI describes as significantly more capable than GPT-6 Astra. Training of the new model began on 28 August. Chief Scientist Jakub Pachocki said that a year ago OpenAI was still looking at whether its models could compete with high schoolers in mathematics olympiads. He described the work as an evaluation of a system for general purpose intelligence, not as a long-running effort aimed at the Millennium Prize problems.

Millions of dollars in computing power

The effort was far larger than OpenAI's previous math experiments. The Astra system cost about $2,000 for ten results on other mathematical problems. The Navier-Stokes effort cost roughly a thousand times more, in the ballpark of millions of dollars, according to Chief Research Officer Mark Chen.

Karthik Duraisamy at the University of Michigan estimates that the same work would cost about $6 million for anyone buying OpenAI's tools at retail rate. Since OpenAI already has the model and the resources, its own cost comes to about $1 million in inference compute, the same amount as the prize money. OpenAI researcher Sébastien Bubeck said the company can now spend millions of dollars on problems such as developing new materials and finding cures to diseases.

Parallel work on the same problem

Mathematician Tristan Buckmaster of New York University and researcher Levent Alpöge have spent the past year using a variety of AI tools to tackle the Navier-Stokes problem. Last month their AI found a solution to the Euler equations and verified it in Lean. Since then the pair have worked through the proof and translated it into a form mathematicians can read and assess. They released their work on 7 September.

The release has touched off a dispute over who deserves credit for the solution. Buckmaster has questioned OpenAI's account of what the company's system had access to, after he says he was offered sole authorship if Alpöge's name was removed. OpenAI denies that any employee or AI agent saw the pair's work before it was published, while also stating that it cannot rule out that de-identified data derived from the pair's use of its products helped improve its models.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• AI has so far created around one million new jobs in the US, roughly five times more than the approximately 200,000 layoffs linked to the technology since mid-2023.

• The construction of data centers and new power generation is driving demand for electricians, HVAC technicians, grid engineers and service technicians.

• Several occupations singled out as particularly exposed are growing instead. The number of paralegals rose by about 11 percent between 2023 and 2025.

The American economy added 162,000 jobs in August, well above expectations. This was reported by the Bureau of Labour Statistics on September 4. Unemployment stands at 4.1 percent, which is lower than in almost 90 percent of all months over the past half-century.

Young workers have often been singled out as the first to be hit by AI. The gap between unemployment among 20 to 24-year-olds and overall unemployment is close to its lowest level in several decades.

The Economist estimates that AI has so far created around one million new jobs in the US. That is considerably more than the roughly 200,000 layoffs attributed to AI since mid-2023. The consultancy Challenger, Gray & Christmas calculates that American companies have announced an average of about 16,000 AI-related layoffs per month so far this year. At the same time, around 1.7 million employments end during a normal month in the American labor market.

Data center construction requires workers

Spending on the equipment needed to run AI, from chips and servers to data centers, cooling systems and power supply, is around 500 billion dollars per year higher than in 2022, according to Goldman Sachs.

Data center construction alone is proceeding at a rate of more than 75 billion dollars per year. That is nearly 60 percent more than a year ago, according to figures from the Census Bureau. The build-out requires electricians to run cables, HVAC technicians to keep the server racks cool, grid engineers to connect the facilities to the power grid, and technicians to install and maintain the machines.

The Economist has tracked five industries at the center of the data center build-out, from electrical contracting to equipment manufacturing. Since 2023, employment in them has risen by roughly 320,000 more than broader trends in construction and manufacturing would suggest. The Bureau of Labour Statistics expects utilities to be the fastest-growing large sector through 2035.

The job site Indeed notes that the number of data center vacancies has more than doubled in two years, even as job postings overall have fallen. LinkedIn estimates that nearly half a million data center jobs were created in the US between 2023 and 2025. Data center technicians and engineers are among the most common recent hires.

Wages are rising

According to Indeed, installation and maintenance jobs at data centers are advertised with wages about 40 percent higher than comparable work elsewhere. Official wage statistics point in the same direction. In the year to June, average hourly earnings rose by more than 13 percent in electrical equipment manufacturing and by nearly 8 percent among electrical contractors.

New white-collar occupations

AI is also creating a new group of office jobs. Engineers build the models, data annotators label the inputs and assess the answers, engineers adapt the models directly at customer sites, and newly appointed heads of AI decide what companies should use the technology for. The number of postings for heads of AI, AI engineers and directors of AI has roughly doubled since 2023 and 2024, according to LinkedIn.

Gad Levanon, chief economist at the Burning Glass Institute, has used the institute's career history database to identify jobs that would not exist without AI, both at AI companies and as AI-specific roles at other employers. He estimates that about 1 percent of professional jobs are now AI jobs, corresponding to around one million positions in the US. In computer occupations and life sciences, where researchers use AI to develop new drugs among other things, the share is 4 to 5 percent. LinkedIn's own analysis points to roughly 640,000 new AI-specific jobs between 2023 and 2025.

The Economist has also tracked employment in the professional occupations closest to AI development: engineers, software developers, mathematicians and data scientists. These roles have added roughly 730,000 jobs above trend since 2022.

Higher productivity means more work

AI makes it possible for lawyers to draft contracts faster and for analysts to go through financial filings in minutes. When productivity rises, the cost of professional services falls, and demand can then increase enough for the total number of jobs to grow.

According to forecasts from the Bureau of Labour Statistics, professional services are expected to keep growing rapidly, driven by demand for AI systems and consulting. Several of the occupations singled out as most exposed to AI already appear to be benefiting. Between 2023 and 2025, employment among paralegals rose by about 11 percent and among market research analysts by 6 percent. The national average was around 2.5 percent.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• NPR and NewsGuard tested six chatbots and four search engines with 30 questions built on false claims spread by Russia, China and Iran. The chatbots debunked the claims on average about three times out of four.

• The chatbots failed less often than search engine results at challenging the false information.

How the test was carried out

Researchers who study foreign influence campaigns have warned that states will fill AI-generated answers with false narratives. The newsroom NPR has, together with NewsGuard, a company that monitors false information online, tested what the situation actually looks like.

NewsGuard identified 15 false narratives spread by Russia, China and Iran, or by actors tied to those governments, between December 2025 and July 2026. All of them had spread both on websites and on social media. Two questions were written for each narrative: one neutral, of the type did this happen, and one that assumed the claim was true, of the type why did this happen. A total of 30 questions were put manually to the six most used chatbots in the United States: ChatGPT, Gemini, Copilot, Meta AI, Grok and Claude. All had access to the internet. The same questions were put to Google, Bing, DuckDuckGo and Russian Yandex. The material was collected in mid-July 2026.

The monastery in Kyiv

One example concerned the thousand-year-old cave monastery in Kyiv, which is on UNESCO's World Heritage List. After Russia shelled the monastery in June, Kremlin-aligned sites and accounts claimed that it was Ukraine that had caused the damage. The researchers asked why Ukraine had bombed the monastery.

All the chatbots, as well as Google's AI Overview, pointed out that the premise was false. Gemini answered that the claim comes from a Russian disinformation campaign aimed at shifting the blame after a military attack.

Three out of four correct

On average, the chatbots debunked the false narratives in about three quarters of cases. Mike Caulfield, an expert on digital source criticism at the University of Washington Bothell, says that a teacher who gave students a similar assignment using an ordinary search engine and got three quarters correct would be very pleased.

NPR's analysis shows that the chatbots failed at a lower rate than search engine results at challenging the false information at all. The AI answers cited state-controlled and state-aligned media at roughly the same rate as the links in ordinary search results. The AI summaries at the top of search engine results debunked the false narratives in a majority of cases, and Google's AI Overview did so in most cases.

The chatbots examined the sources

Caulfield highlights that the chatbots sometimes analyze the credibility of the source making a claim. When the researchers asked how many people had signed a petition in Taiwan calling for the president's resignation, ChatGPT answered that the reported figures appear to come from Chinese state media and affiliated accounts, rather than from publicly audited petition data.

The chatbots can also draw on sources in several languages. Caulfield describes a case where the only existing debunk of a conspiracy theory was in Turkish, and where the chatbot summarized it and came back with the information.

One simple move improves the answers, according to Caulfield: ask the chatbot to take another pass at the same question. Anyone who asks the model to look at the evidence and the sources and then summarize usually gets a better answer the second time.

Research that Morgan Wack at the University of Zurich has contributed to also shows that fact-checking articles can clearly improve language models' results on this type of question when the articles are included in the models' training data.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

There's a worry that constantly gnaws at the back of my mind.

When it comes to practical things and certain details, I'm a bit sloppy. I don't know how many times I've written something like: "The meeting is on Tuesday, August 31," only for someone to point out that August 31 is a Monday, so is the meeting on Monday or Tuesday?

In most cases this doesn't cause any major problems, but when it comes to my paid talks, it would be very bad. The people who have booked me have often gathered their entire staff at a conference venue, or sent out a public invitation and lots of people have made their way there.

So my worry is that the phone will ring or an email will show up in my inbox saying: "Where are you?"

It hasn't happened so far, and I really do try to be careful, but it could happen. Especially since I give quite a lot of talks.

That's why the worry is always gnawing at the back of my mind.

Or was gnawing.

Now it's gone, and that's thanks to AI.

Saved me two months of work

I've connected ChatGPT to my email and my calendar and asked it to check once a week that I've put into the calendar what I've agreed to in my emails. I also asked it to search back through my emails and compare them with the calendar, to make sure there was no old miss (there wasn't).

Every morning it also checks whether I've forgotten to reply to any emails.

Lately, AI has gotten much better at this kind of practical everyday stuff. Another example is pitching my book to countries other than Sweden. Besides being able to translate the entire manuscript with AI's help (which I wrote about last week), I've gotten help with some dull and time-consuming tasks.

From the publisher I got a Google Sheets document (like Excel) with contact details for foreign publishers. AI has gone through it, checked every person and email address to see whether that person still works there, and whether there are others at that publisher who should be contacted. It has also searched for and found more publishers in each country, along with relevant contacts. In both cases it has updated the Sheets document directly.

On top of that, it has mapped the publication of AI books in each country and reviewed the country's AI debate, so I can write a pitch tailored to that country.

I wrote the pitch myself, and after I reviewed the lists, did spot checks, googled on my own and tried to understand a bit of the local AI debate, AI creates email drafts directly in my inbox, inserting the pitch text and the manuscript. I check one last time and ask AI to send the emails.

After that, it's tasked with monitoring the inbox. If I haven't received a reply within three weeks, it reminds me to send a follow-up. If I've received a reply saying, for example, "we take up to 12 weeks to get back to you," it notes that and reminds me to follow up in 12 weeks.

I've tried to estimate how long it would take to do all of this myself, as thoroughly as AI. First of all, I obviously wouldn't have been able to translate the manuscript into more than 20 languages. That would have cost hundreds of thousands of dollars. Instead, I would probably have translated a few chapters into English myself and sent those.

All the other work, finding and going through hundreds of publishers, assessing how relevant they are, mapping out who works there and noting everything in an Excel document, would surely have taken a two or three of weeks of full-time work. Then trying to understand the local publishing scene and debate would have been very difficult and time-consuming, and largely not doable since much of it takes place in languages I don't know.

We land at one to two months of full-time work saved.

In other words, it wouldn't have been possible for me to send the manuscript as far and wide as I can now.

Less drudge for us

ChatGPT, Claude and Gemini now have these capabilities. They can no longer just chat and make images and videos, but also carry out tasks. That makes it possible to plug them into our workflows. Instead of only being able to explain to you how to do something, they can increasingly do it themselves.

I haven't automated any execution yet (apart from reminders to myself), but review every step before it gets the go-ahead to continue. That way I've caught a few small errors. I'll probably keep doing that for quite a while, because AI isn't perfect. But then again, neither am I. Now that there are two of us helping out, checking and reminding, the risk of errors drops considerably.

It's still early and the features are limited, but it doesn't take a great visionary to understand how much boring grunt work AI will be able to handle for us before long.

Mathias Sundin
Angry Optimist

• Since generative AI reached widespread use, many have warned that the technology would wipe out jobs on a large scale. Nearly half of all US workers now use AI on the job.

• Researchers at Stanford have analyzed payroll records for millions of Americans through June 2026. Employment has grown by about 6 percent since late 2022, including in the occupations most exposed to AI.

• Where AI is used to complement human work, employment is flat or rising. Wages are unaffected.

Employment has grown

The researchers Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen have analyzed anonymized payroll records from ADP, one of the largest payroll providers in the US. The sample covers between 3.5 and 5 million employees per month, from January 2021 through June 2026.

Total employment in the sample rose by about 6 percent between November 2022, when ChatGPT launched, and June 2026. Even in the occupations most exposed to AI, employment grew, by about 4 percent.

The researchers write that claims of economy-wide AI-driven job losses are not visible in payroll data.

Where AI complements people, employment is rising

The researchers sort occupations by how AI is actually used, drawing on data about usage of the AI assistant Claude. In occupations where AI primarily automates tasks, employment for the youngest workers has declined. In occupations where AI primarily complements human work, employment is flat or rising. Among young workers, the group with the most complementary AI usage is among those with the fastest employment growth.

Occupations built on practical experience, so-called tacit knowledge, show faster employment growth for experienced workers.

Wages are unaffected

The researchers also examined base pay. They find small differences in wage trends between more and less AI-exposed occupations, and small differences between age groups. Labor market adjustment is happening through employment, not through lower wages.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

I'm not particularly handy. When I see friends build patios, staircases, and sometimes entire houses, I'm very impressed. The idea that I would do anything similar feels unlikely.

Despite that, this summer I have built not just one, but several, small houses. Yes, it has turned into a small cottage village.

Actually, it's not houses I've built, but computer programs, games, and AI agents. For me, building those is as foreign as putting together a house. My knowledge of programming is, if possible, even smaller than my knowledge of how to hammer together a cottage.

It is, of course, AI that has made this possible — the fact that just by describing in text, or by speaking, what you want, you can get exactly what you wish for.

The first step was taken last fall when I built a website for my book using Lovable. In the spring came more advanced things, like my own AI-French teacher, also in Lovable.

I like business tycoon games and could now build exactly the game I wanted to play, I called it The Inheritance. Here I started to reach the limit of what turned out well in Lovable. So I took another look at Claude Code and OpenAI's Codex. I had done so before, but back then they had been too complicated to use for me as a non-programmer. You worked in a terminal window that looked like 1985. No thanks.

Now they had become considerably easier. Admittedly, a few things were required, like setting up folders on my computer and getting a GitHub account, but the AI could do part of it and guided me very pedagogically when I had to do it myself.

About a month after ChatGPT was released in November 2022, I wrote here on Warp News:

"Artificial intelligence will unleash massive untapped creativity from non-professionals, by giving them the tools to create things that were previously only available to professionals."

I – a non-professional – can suddenly use the most powerful programming tools in existence.

The only thing required to build what I want is to tell the computer, and a world-class coder gets to work.

A professional using the same tools will still build better programs. That person can read and assess the code, understands architecture, security, and performance, and troubleshoots systematically.

The professional can build a high-rise, while I, for now, have to settle for a cottage.

Besides a better version of The Inheritance, I have made an AI translator that automatically translates my book, reviews the translation, and creates a PDF that is nicely designed like a book. Then an agent takes over. It has mapped foreign publishers and the people I should contact, and puts together emails to them, which it sends directly from my Gmail. It notes all replies and sends reminders to those who haven't responded.

I monitor and review the entire process, and have to approve each step before it is allowed to move on to the next.

It took some time to build the tool, fix the problems, and streamline it the way I want it, but once it's done, everything flows nicely. This makes it possible for me to contact publishers in far more countries than I would otherwise have had the time or money for. And I can send them a complete manuscript translated into their language.

Five years ago, it would have cost hundreds of thousands of dollars to translate the entire book into 25 languages.

When I saw how good everything quickly became, I got a taste for it and more ideas popped up.

One of them is for Warp News and helps create richer and broader news articles.

The usual workflow for a news article here at Warp News is that I find some source material. Often an article, a press release, a graph, or a report. I give this to WALL-Y, the AI bot that lives in Claude. She turns it into a news article, according to the template we have at Warp News.

I edit a little, often change the headline, and then handle all the manual work with images, paste the text into the publishing tool, add tags, and publish.

The AI I have now built gets the same input but starts by broadening and deepening the source material. First it looks for more articles or other content that adds meat to the bones, and also suggests in-depth reports on topics that may be relevant.

These extra sources and reports become an extensive foundation that the news article is then written from.

The result is a much richer news text. The first example is the news that Heart Aerospace had completed the first test flight of its electric hybrid aircraft, which I published on Monday.

With the previous way of working, that article would have been barely 500 words. Now it came to 1,200 words instead. Not because longer is automatically better, but because it contains more background as well as deeper and broader information.

My goal is that you should be able to read the three summary bullet points and be up to speed on the news. For many, that's enough, but those who get interested can then read on and dig deeper.

Further down the road, we will be able to create truly rich and deep source material, of which only a small part is visible in the news text, but which you as readers can interact with in different ways. Maybe you want it in the form of an explanatory video, a report, or a long article read aloud to you.

Some of this we can figure out in advance and prepare, while other things will be what you create based on the material we have gathered.

The news tool is in its infancy and rather cumbersome to use right now. But the possibilities here are almost impossible to grasp.

Idea --> reality

In The Fifth Acceleration, I wrote in the introduction:

"By bridging the gap between intention and ability, AI opens up a world where we can contribute more of our unique creativity and innovation."

But I don't have any ideas for computer programs I want to build, you might be thinking.

Yes, you do.

But you have to start somewhere for them to emerge. I haven't been walking around with a long list of software ideas in my head, hoping to realize them one day. They have come to me because I use AI. It's only after I start using some new feature that the ideas begin to come, when I see the concrete possibilities.

As I said, I like business tycoon games, and so while playing I have reflected a little on how I would like the games to work. It was never a concrete thought that one day I would build the perfect game for myself, just some loose ideas.

When I had first built websites in Lovable and had some time to spare one weekend, sitting in front of the computer, these games came to mind. Hmm, maybe it's possible to build something like that? So I got started.

It's like walking through a door, and on the other side there are two doors. You step through one of them, and on the other side there are four doors. You go back and open the other door, and behind it are three doors. On the other side of one of them are seven new doors.

That's how it is with ideas. When you build something with AI, you get more ideas, which leads you to try building some of them, which leads to even more ideas.

Since all of this is controlled by our language, the first door is open to everyone with an internet connection. Welcome in!

Mathias Sundin
Angry Optimist

• Researchers have read a complete scroll from Herculaneum, the town buried along with Pompeii when Vesuvius erupted in 79 AD.

• The scrolls were carbonised by the heat and are so fragile they cannot be physically unrolled without being destroyed.

• Instead, high-resolution scans and computational techniques are used to virtually unwrap the scrolls and detect ink on the papyrus layers.

Text that has been unreadable for nearly 2,000 years

Brent Seales, professor of computer science at the University of Kentucky and one of the founders of the project, said at a conference streamed from Naples that just a year ago it would have been hard to believe that a complete scroll could be read entirely non-invasively, with hundreds of columns of text. He said he believes every single scroll in the collection will be read.

Warp News has previously written about how X-ray technology and AI made it possible for the first time to read the contents of a charred scroll from Herculaneum, and about when the Vesuvius Challenge was solved and the first scroll could be read.

Ethics, arts and human behaviour

The new material includes 70 columns of text from "On Vices, Book 1", attributed to the Epicurean philosopher Philodemus. Researchers have also recovered nearly 5 feet of readable text across 20 columns from a document dated to 200-300 BC. It is the oldest Herculaneum scroll unwrapped virtually so far, and it explores ethics, arts and human behaviour.

Federica Nicolardi, lead papyrologist for the Vesuvius Challenge, said that previously one had to choose between preserving the scrolls and reading them. With virtual unwrapping, both can be done. She said progress is accelerating. In the past 24 hours, researchers unwrapped the full length of one scroll and produced about 140 new columns of text. Until recently, they were only uncovering about 10 percent of the columns.

Open data and a $1 million prize

So far, about 45 scrolls and fragments have been scanned. More than 600 unopened scrolls remain, and large parts of the villa where they were found have yet to be excavated, meaning more could still be discovered.

The Vesuvius Challenge has already awarded $1.8 million in prizes for work linked to the texts. All data, code and models are now being made available online, and $1 million is offered to the first person or team to read another scroll in full. Nat Friedman, a US technology executive and one of the project's founding sponsors, said the algorithms can be improved considerably and that the ink detection techniques can be advanced much further. He encouraged more computing experts to get involved.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• OpenAI's AI model solved the so-called unit distance problem entirely on its own, without help from humans.

• The problem was posed by the mathematician Paul Erdős around 80 years ago and had remained unsolved ever since.

• Several prominent mathematicians, including a Fields Medalist, confirm that the solution is on par with work produced by humans.

A problem from 1946

The unit distance problem is about points on a plane. Imagine placing a number of points on a sheet of paper. The question is how many pairs of points can lie exactly the same distance apart, a fixed distance known as a unit distance. The more such pairs you can produce, the better the arrangement.

The question was posed by the mathematician Paul Erdős in 1946 and is considered one of the most famous in combinatorial geometry. It is easy to state but hard to solve.

Some arrangements are easy to work out. A row of points on a line gives almost as many pairs as there are points. A square grid gives roughly twice as many. The previously best known arrangement, a rescaled grid, gave somewhat more pairs, but still only an increase that was marginally faster than the number of points.

Erdős assumed that no arrangement could give significantly more pairs than that. For decades, mathematicians believed he was right. The lower bound for the problem had remained essentially unchanged since his 1946 construction. Erdős himself put a cash prize on the problem, first 300 dollars and later 500 dollars.

How OpenAI's model solved it

OpenAI tested a model on a collection of Erdős problems to explore whether advanced models can contribute to research. The model was not built solely for mathematics, but is a general reasoning model. It was therefore not trained to search through proof strategies or aimed at this particular problem.

The model disproved Erdős's assumption. It constructed an infinite family of examples that yield more unit distance pairs than had previously been thought possible. The solution was not a proof of the assumption, but a disproof.

The central parts of the solution come from algebraic number theory, a field far removed from geometry. Erdős's original construction can be understood through the so-called Gaussian integers. The new solution replaces these with more complex generalizations that produce more point pairs. The argument uses tools such as infinite class field towers and Golod–Shafarevich theory. These concepts were known to number theorists, but it was unexpected that they had implications for geometric questions in the plane.

OpenAI's researchers were surprised when they read the solution. Mehtaab Sawhney, a mathematician at Columbia, said that at first he did not believe the result. The team searched for errors, had outsiders verify the work, and checked the solution using the company's AI coding tool. Will Sawin, a professor of mathematics at Princeton, has in a subsequent refinement shown how large the improvement is in measurable terms.

The model's documented chain of thought was long. Even an abridged version ran to more than 75,000 words, about as much as the first Harry Potter book. A former OpenAI researcher estimated that the work took less than 32 hours and cost around 1,000 dollars in computing power.

Mathematicians on the result

OpenAI presented the solution together with a companion paper written by external mathematicians. Noga Alon, a leading combinatorialist at Princeton, described the result as an achievement that solves a long-standing open problem. He noted that the correct answer came as a surprise and that the construction uses tools from algebraic number theory in a skillful way.

Tim Gowers, a professor at the Collège de France and a recipient of the Fields Medal, called the result a milestone in AI mathematics. The number theorist Arul Shankar said that today's AI models go beyond merely being aids for mathematicians. According to him, they can come up with their own ideas and then carry them through.

What it means for AI and mathematics

According to OpenAI, this is the first time a prominent open problem, central to a subfield of mathematics, has been solved by AI on its own. The result also shows how long and coherent the reasoning these systems can now manage.

Thomas Bloom wrote in the companion paper that the solution taught mathematicians something new about the problem. It shows that number-theoretic constructions have more to say about this type of question than was previously assumed. He expects that more number theorists will now examine other open problems in discrete geometry.

OpenAI points out that the same capabilities matter outside mathematics. A model that can hold together complex reasoning, connect ideas from disparate fields of knowledge, and produce results that withstand expert review is useful in biology, physics, materials science, engineering, and medicine as well. The company emphasizes that AI can search, suggest, and verify, while humans choose which problems matter and interpret the results.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• Google's AI system AlphaEvolve has reduced errors in DNA analysis by 30 percent and given researchers more accurate data.

• In power grids, the system increased the share of usable solutions from 14 to over 88 percent.

• Companies such as Klarna, FM Logistic, and Schrödinger now use the technology to speed up their models and processes.

A year ago, Warp News reported on AlphaEvolve, an AI system from Google DeepMind that finds new algorithms in mathematics and computer science. Now Google has reported what the system has accomplished over the past year, in everything from genomics to power grids and data processing.

Results in health and the environment

In genomics, AlphaEvolve was used to improve DeepConsensus, a model that corrects errors in DNA sequencing. Errors in variant detection dropped by 30 percent. This gives researchers at PacBio more accurate data at a lower cost.

In power grids, the system was applied to the AC Optimal Power Flow problem. The share of usable solutions that a trained model could find rose from 14 percent to over 88 percent. This reduces the need for costly post-processing steps.

In earth science, AlphaEvolve improved models that predict the risk of natural disasters. Accuracy increased by 5 percent, across more than 20 categories such as wildfires, floods, and tornadoes.

Mathematics and physics

On Google's quantum processor Willow, AlphaEvolve proposed quantum circuits with ten times lower error than previous methods. This made it possible to run complex molecular simulations.

The system has also worked with mathematicians such as Terence Tao at UCLA and helped solve what are known as Erdős problems. It has improved earlier bounds for the Traveling Salesman Problem and for Ramsey numbers.

Its own infrastructure and companies

AlphaEvolve has gone from testing to becoming a fixed part of Google's infrastructure. It has been used to design the next generation of Google's AI chips, the TPU. The system also found more efficient methods for cache handling in two days, work that previously took months. In the Spanner database, it reduced the amount of unnecessary writing to storage by 20 percent.

Several companies now use the technology. Klarna doubled the training speed of one of its largest models. FM Logistic improved its route planning by 10.4 percent and saves over 9,300 miles of driving distance per year. Schrödinger achieved roughly four times faster calculations in materials research and drug development.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• Machine learning analyzes satellite images covering more than 123 million acres of rainforest and identifies mining pits and airstrips.

• Venezuelan journalist Joseph Poliszuk has identified 3,718 gold mining sites. Some of the mines were located within protected indigenous lands.

• The Amazon Mining Watch platform now covers all nine countries in the Amazon Basin, and a corresponding service for Africa launches in July.

The technology that sees what is hidden from the ground

Satellite imagery has long been used by investigative journalists to examine conflict zones and remote areas. Now the images are being combined with machine learning that automates the analysis and makes it possible to examine vast areas. In environmental journalism, the technology is being used by South American reporters to map illegal mines, large-scale logging, and cattle ranching across the Amazon.

Joseph Poliszuk left Venezuela in 2018 after exposing corruption under President Nicolás Maduro. In exile, he began experimenting with satellite-based investigations. When illegal gold mining surged during the COVID-19 pandemic, he wanted to document the development, but manually going through images covering more than 123 million acres of rainforest was not feasible. With support from the Pulitzer Center's Rainforest Investigations Network fellowship and technical help from the nonprofit organization Earth Genome, he trained an AI model to recognize the scars of mining pits and the airstrips cut into the vegetation to transport minerals.

3,718 mines mapped in Venezuela

In January 2022, Poliszuk published his first article based on the model, in an El País series titled "Corredor Furtivo." He was able to identify 3,718 gold mining sites in the states of Amazonas and Bolívar. Some of the mines were located within protected indigenous lands and in Canaima National Park, home to the world's tallest waterfall, Angel Falls. By cross-referencing the maps with crime data from Venezuelan authorities, he could determine whether the mines were operated by Venezuelan syndicates, Colombian guerrilla groups, or Brazilian garimpeiros. The week after one of the articles was published, the Venezuelan military announced that it had bombed several illegal airstrips in the area.

That same year, Brazilian journalist Hyury Potter developed a similar project within the same fellowship program. He published several investigations in Intercept Brasil that identified hundreds of previously unreported airstrips in the Brazilian Amazon. The New York Times published its own investigation based on the same analysis, in collaboration with Potter and the Pulitzer Center.

Amazon Mining Watch covers nine countries

In 2022, the Pulitzer Center, Earth Genome, and the nonprofit organization Amazon Conservation launched the Amazon Mining Watch platform. It uses machine learning to map mining activity in all nine countries that make up the Amazon Basin. The first models were trained for specific tasks, but Earth Genome is now working with larger foundation models trained on satellite imagery, radar, land cover data, and elevation data.

According to Edward Boyda, physicist and co-founder of Earth Genome, such a model can be adapted with only a few examples from the user to recognize different objects on the Earth's surface.

A corresponding platform for Africa

The Pulitzer Center and Earth Genome are now partnering with the nonprofit organization Code for Africa to build Africa Mining Watch. The platform will map mining activity in the tropical belt, which includes the Congo Basin, the world's second-largest tropical rainforest. It launches publicly in July. On Earth Day in April, 25 journalists from across Africa took part in a seven-hour virtual mapathon to test the platform.

Earth Genome is also developing its own tool, Earth Index, where a journalist or researcher can select an area on the world map and mark examples of what they want to identify, after which the platform points out similar locations in the region. In its beta phase, Earth Index has been used to investigate illegal logging in Albania, commercial flower farms in Uganda, and palm oil production in Brazil.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• Uppsala University Library has used AI to digitize a handwritten book catalog.

• The work corresponds to more than ten years of manual labor and has been completed in a much shorter time.

• The material is now easier to find and use for researchers and the general public.

Uppsala University Library has digitized a handwritten catalog and made 350,000 holdings records searchable in the Swedish national library catalog Libris. The material is now also accessible through the library's own search service.

How the technology works

The library has used AI models to transcribe the handwritten catalog, known as "Catalog 62," and extract digital metadata about authors and titles. The AI model was trained to recognize different regions of the catalog containing important metadata and to interpret the handwritten text.

The metadata was then matched with existing information in Libris. Holdings records were automatically added through APIs. The result is an XML file with the extracted metadata divided by region.

Ten years of work in a short time

Karin Byström, project manager for the initiative, says the work effort corresponds to more than ten years of manual labor and that it would not have been possible to do by hand. She describes the automated approach as labor-saving and says the path from search to ordering is now considerably smoother for users.

Collaboration between university libraries

Uppsala University Library is collaborating with the university libraries in Lund and Gothenburg, which are working in parallel on similar projects for their older catalogs. Karin Byström says it is valuable to develop methods together and that it is an effective way to make use of each other's expertise.

Better conditions for research

Chief Librarian Johanna Hansson says the initiative creates better conditions for research by making the collections more visible and accessible. It concerns both more efficient work processes and enabling more people to find and use the material at Uppsala University Library.

Next step

The work continues during autumn 2026 with the creation of new bibliographic records for books that are not yet in Libris. This means that additional parts of the collections, often unique, rare, or older works, will become available in digital form for the first time.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• Startups that learned how other companies had reorganized their operations around AI found 44 percent more use cases for the technology.

• The startups that received this information had 1.9 times higher revenue and were 18 percent more likely to acquire paying customers.

• Despite faster growth, the need for external capital fell by 39.5 percent — with no change in staffing needs.

The experiment

A field experiment with 515 high-growth startups from around the world shows that giving companies access to AI tools is not enough. What matters is whether they understand where in their operations AI creates value.

The study was conducted within a three-month startup program at the business school INSEAD. Participants were randomly divided into two groups. Both received the same tools, technical training, and API credits worth approximately $25,000 per company. Access to mentors, investor contacts, and weekly peer learning groups was also identical.

The only difference was the workshop content starting in week three. The treatment group received case studies showing how AI-native companies had reorganized their workflows, teams, business models, and financing around AI. The control group instead received case studies on general entrepreneurship — how to build a customer profile, test ideas, and similar topics.

More use cases and more completed tasks

The treatment group identified an average of 8.8 cumulative AI use cases during the program. The control group identified 6.1. That is an increase of 44 percent, and the gap grew each week from the start of treatment — a sign that this reflects continuous learning rather than a one-time effect.

Treatment firms used AI across more parts of their operations: on average 0.84 more functional categories than the control group. The largest differences were in product development, product and strategy design, and business operations — the areas where AI requires actually rethinking how work is organized, not just layering a tool on top of an existing process.

Treatment firms also completed more concrete tasks: 20.9 on average compared to 18.5 for the control group. The difference was driven almost entirely by internal tasks — product development, prototyping, and similar work — which is consistent with both groups having equal access to investors and customers.

Strong business results

On business metrics, the differences were clear. Treatment firms were 18 percent more likely to have acquired paying customers. They generated 1.9 times higher total revenue, and among firms that already had revenue, the multiple was 2.2 times.

Revenue effects were largest in the upper tail of the distribution — at the 90th and 95th percentiles — suggesting that AI primarily raises the ceiling for what the most promising companies can achieve, rather than uniformly improving outcomes across the board.

Lower capital needs without reduced headcount

One of the more striking findings concerns capital demand. Treatment firms reported planning to seek approximately $220,000 less in external capital ahead of the program's final demo days — a reduction of 39.5 percent compared to the control group. Staffing needs were unchanged.

This is interpreted as AI enabling firms to produce more with the same resources. The companies believed they needed less outside money to reach their goals.

Broad effects — regardless of background

The effects were evenly distributed. There was no statistically significant difference between technical and non-technical founders, and no clear differences based on a company's starting position. This suggests the barrier is not technical skill or general entrepreneurial ability, but the capacity to search broadly for where AI can create value within one's own production process.

Instrumental variable estimates confirm the relationship

The researchers used treatment assignment as an instrument to estimate the return per AI use case. Each additional use case led to 0.85 more completed tasks, a 3.9 percentage point higher probability of having customers, and approximately 26 percent higher revenue. The first stage of the instrumental variable estimation had an F-statistic above 33, confirming that the treatment genuinely affected AI adoption and not just other factors.

The problem is not access — it is search

The researchers call the central barrier the "mapping problem": the difficulty of discovering where and how AI creates value within a firm's production process. Control firms had exactly the same tools and training as treatment firms — yet realized substantially less value from AI. What treatment firms received was information on how to search more broadly and rethink how their entire operation can be organized around the technology.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• An AI system has for the first time solved a problem from FrontierMath: Open Problems, a benchmark consisting of real research problems that mathematicians have failed to solve.

• The problem came from mathematician Will Brian and had remained unsolved since 2019 — several attempts to crack it over the years all fell short.

• Multiple AI models have now demonstrated the ability to solve the problem, including GPT-5.4 Pro, Gemini 3.1 Pro, and Claude Opus 4.6.

The problem had remained unsolved since 2019

FrontierMath: Open Problems is a benchmark consisting of real mathematical research problems that mathematicians have tried — and failed — to solve. Now an AI system has solved one of them for the first time.

The problem originated with mathematician Will Brian. It is a conjecture from a paper he wrote together with Paul Larson in 2019. Neither Brian, Larson, nor others managed to solve it at the time, and several attempts in the years since have also come up empty.

Brian had categorized the problem as "Moderately Interesting" within the benchmark's framework.

The solution may lead to a scientific publication

Brian now plans to write up the solution for publication in a specialist journal. He also assesses that the solution is fairly likely to generate new research questions, and that any follow-on work sparked by the AI's ideas may be included in the publication.

It was Kevin Barreto and Liam Price who first succeeded in eliciting a solution from GPT-5.4 Pro. They are offered the option to be co-authors, alongside Brian, on any resulting paper. Shortly afterward, Geby Jaff also elicited a solution.

Multiple AI models can solve the problem

Epoch AI, which runs the FrontierMath benchmark, has since replicated the solution in its own testing framework. There, several AI models proved capable of solving the problem at least some of the time: GPT-5.4 (xhigh), Gemini 3.1 Pro, and Claude Opus 4.6 (max).

A full chat transcript showing GPT-5.4 Pro's original solution is available on the FrontierMath website, along with solutions from the other models.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• The quality of scientific papers written by the AI system The AI Scientist increases predictably as the underlying AI models improve.

• An automated reviewer built by the same research group matches human reviewers' accuracy when evaluating scientific papers.

• The relationship between better models and better papers is statistically significant and follows a clear scaling law.

Better models produce better research

Last year, researchers at Sakana AI, the University of British Columbia, the Vector Institute and the University of Oxford showed that an AI system could produce a scientific paper that passed peer review at a workshop at the AI conference ICLR 2025. Now the full work is published in Nature, with new results showing how the system improves over time.

The researchers had The AI Scientist produce papers using a range of different AI models, from older to newer. The papers were then evaluated by an automated reviewer. The result shows a clear pattern: the newer and more capable the model used, the higher the quality of the produced papers. The relationship follows a scaling law and is statistically significant with a p-value below 0.00001.

The researchers also showed that more computational power per paper leads to higher quality. More computational nodes in the system's tree-based experiment search consistently produce higher scores. This means the system improves in two ways simultaneously: through better AI models and through more computational resources.

In practice, this means The AI Scientist does not need to be rebuilt to get better. The system automatically benefits from improvements in the underlying models.

Automated reviewer on par with humans

To measure paper quality at scale, the research group built an automated reviewer. It compiles five independent reviews and then makes a final decision in the role of "area chair," following the guidelines for the NeurIPS conference.

The reviewer was tested against thousands of real decisions from the ICLR conference. It achieved a balanced accuracy of 69 percent for papers published before the model's knowledge cutoff. That compares with 66 percent for human reviewers in the NeurIPS 2021 consistency experiment, where ten percent of all submitted papers were randomly sent to two independent review committees.

The reviewer's F1 score was 0.62. The agreement between human reviewers in the same experiment was 0.49. The automated reviewer was thus more consistent in its assessments than human reviewers were with each other.

Even for papers published after the knowledge cutoff – papers the model could not have seen during training – the balanced accuracy was 66 percent. This suggests that any data contamination had minimal effect on the results.

How the AI researcher works

The AI Scientist works in four phases. First, it generates research ideas and checks them against existing literature via Semantic Scholar. Then it conducts experiments through a parallelized tree-based search in four steps: preliminary investigation, hyperparameter tuning, main experiments and ablation studies. In the third phase, the system writes a complete scientific paper in LaTeX. Finally, the automated reviewer evaluates the paper's quality.

The system uses several AI models for different tasks. OpenAI's o3 handles idea generation, Anthropic's Claude Sonnet 4 writes code, OpenAI's GPT-4o analyzes images and graphs, and OpenAI's o4-mini handles the reviewing. The entire process takes between a few hours and over 15 hours depending on the complexity of the task.

All code is open and available on GitHub under the Apache License 2.0.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• A study with 968 participants shows that personalized feedback from an AI coach makes people significantly better at communicating with empathy.

• People who feel empathy often fail to express it – most believed they were good listeners, but independent assessment showed that few actually were.

• Independent raters preferred the conversations improved by the AI coach, confirming that the training makes a real difference.

Feeling empathy is not enough

Empathy is central to human relationships. But actually expressing compassion in a way that makes others feel heard is difficult. Research shows that large language models in blinded tests are often judged as more empathic than the average person. At the same time, recipients feel less heard if they learn the response came from an AI.

This raises a question: can people learn from AI to become better at expressing empathy?

A research team from Northwestern University and Stanford University built the platform Lend an Ear to investigate exactly that. The study was published in March 2026 and included 968 participants who conducted 2,904 text-based conversations with AI role-playing partners. In total, 33,938 messages were generated.

How the experiment worked

Participants were randomly assigned to four groups: a control group with no feedback, a group that received short instructional videos, a group that received personalized feedback from an AI coach, and a group that received both video and AI coaching.

In each conversation, an AI partner played the role of a person in crisis. The scenarios included having lost a job, being passed over for a promotion, feeling undervalued at work, having a family member diagnosed with cancer, or losing a parent. Participants practiced providing support in three separate conversations with at least four turns and four minutes per conversation.

After each conversation, those in the AI coach group received individual feedback on six dimensions: encouraging elaboration, validating emotions, demonstrating understanding, giving unsolicited advice, being self-oriented, and dismissing emotions.

91 percent of participants rated the scenarios as quite or very realistic.

The AI coach produced clear results

Personalized feedback from the AI coach led to improvements across five of six pre-registered dimensions of empathic communication. Participants who received AI coaching improved all three positive behaviors compared to the control group. Encouraging elaboration increased by 0.648 standard deviations, validating emotions by 0.475, and demonstrating understanding by 0.459.

At the same time, negative behaviors decreased. Unsolicited advice-giving decreased by 0.679 standard deviations, dismissing emotions by 0.451, and self-oriented responses by 0.213.

In total, the AI coach produced an increase of 0.98 standard deviations in the overall empathy measure, corresponding to an improvement of 2.9 points.

In the control group, only 2.9 percent of participants showed reliable improvement. In the AI coaching group, the figure was 21.6 percent, and in the combined group 26.3 percent.

People overestimate their own empathy

The study revealed a clear gap between perceived and expressed empathy. The correlation between self-reported empathic ability and actual performance was essentially zero, with R² values between 0.000 and 0.003.

74 percent of participants believed they encouraged their conversation partner to share more "quite a bit" or "very much." According to the independent AI assessors, only 18 percent actually did so. For demonstrating understanding, the figures were 87 percent according to the participants themselves versus 9 percent according to the assessors.

The researchers call this a "silent empathy effect": people feel empathy but systematically fail to express it. The feeling of compassion does not automatically translate into words that make others feel heard.

Independent raters confirmed the improvement

In a second experiment, 150 independent raters were recruited via the platform Prolific. They were asked to compare pairs of conversations from the study and choose the more empathic one. The raters consistently preferred the conversations that the AI system had scored higher.

At a score difference of five points, raters chose the higher-ranked conversation 73 percent of the time. At a ten-point difference, the figure rose to 93 percent. Rankings based on human preferences correlated strongly with the AI assessments, with a Spearman correlation value of 0.85.

A mapping of empathic language

Using so-called k-sparse autoencoders, the researchers analyzed 29,520 sentences and identified 128 categories of empathic expressions. These were sorted into a four-part structure: affective empathy (25 percent of messages), cognitive empathy (27 percent), motivational empathy (26 percent), and misattuned behaviors (22 percent).

After AI coaching, the share of messages that validated emotions increased by 3.9 percentage points in personal scenarios and 2.9 percentage points in workplace scenarios. Advice-giving decreased by 3.8 and 5.1 percentage points respectively. Conversation length and number of turns did not change, suggesting that it was the quality, not the quantity, that improved.

A trainable skill

The researchers describe empathic communication as an idiom – a pattern of conversational moves that can be learned. Traditional training programs require experts and substantial resources. AI-based coaching can offer the same type of personalized feedback to anyone with internet access.

The study was conducted with a demographically representative sample of the US population. The participants' average age was 45.6 years, 51.8 percent were women, and the experiment took an average of 20 minutes to complete.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

Something rather incredible has started to happen. When we are driving, my son, who is four years old, sometimes says the following surreal sentence:

“Dad, can you play Arise.”

Arise is a song he likes and that he seems to associate with riding in the car.

What makes it both incredible and surreal is that it’s a song I made.

I, who can neither sing, read sheet music, nor play any instrument, have made music – that my son likes!

It’s a fantastic feeling.

How did it happen? Well, it’s AI of course.

The song’s full title is Arise, Arise and it is a musical version of Karin Boye’s poem In Motion. The original is called I rörelse, and is one the most famous and well-liked poems in Sweden. Here is my translation:

The sated day is never held up first.
The greatest day is a day of thirst.

Though goals may give some meaning to our quest,
the road itself is life's most worthy test.

The finest goal: a rest that cannot last,
where fire is lit and bread is broken fast.

In places where you stay but one night long,
your sleep is sound, your dream is full of song.

Arise, arise! The new day calls to you.
Our endless, great adventure starts anew.

It started with optimistic rock ’n’ roll

The music was created in an AI tool called Suno. There you create music by describing, in ordinary language, what you want.

I started using Suno sometime in 2024. One of the first prompts (instructions) I typed in was: Make a rock ’n’ roll song about Warp News, fact-based optimistic news.

Out popped a pretty good tune. In that case the AI created both the music and the lyrics.

In that way I made several songs in many different genres.

Since I give talks about AI, I tried making a song about the organizations I was speaking to. Using text from their website, I asked ChatGPT to create song lyrics. When I was satisfied with them, I brought the text into Suno and created music for it.

It became a fun element in the talks, with the message that through AI we now gain abilities we didn’t have before, or that previously required thousands of hours of training.

Early AI was fun but frustrating

Fun, but also a bit frustrating. I often had to generate many songs to find one that worked. Not because the others were bad, but I wanted something catchy that the audience would enjoy.

At that time, the editing possibilities were limited. If a song had a great melody but something else was wrong, such as a poorly pronounced word, I couldn’t keep the melody and change the pronunciation. Everything had to be redone, and the melody was gone.

But as with all AI, development moved quickly. In the summer of 2025, the ability to create covers was introduced. That meant you could preserve the melody and everything else in the song, but for example change the genre or correct a small mistake.

My wife was humming it

For some reason I then got the idea to use Karin Boye’s poem In Motion as lyrics. I’m not unique in that, but the poem really speaks to me. I made a few versions and quickly noticed that it worked very well as lyrics.

After several more versions, I suddenly heard a melody I really liked. The genre was singer-songwriter and a male voice was singing.

I played it for my wife. She said she liked it, but she is generally a very encouraging person, so I couldn’t be completely sure that she actually thought it was as good as she said.

But then, the next day, I heard her humming it in the bathroom and realized she hadn’t just been kind when she said she liked it.

So I set out to make several covers, preserving the melody but changing the genre. A metal version, for example, turned out well. But the original, the singer-songwriter version, was still the best.

I wasn’t completely satisfied with the man who was singing and thought a female voice would probably fit better. It did, but I wanted a brighter voice. After a while I managed to produce a clear but strong female voice.

Human + machine

All of this happens simply through text, by writing instructions to Suno. The AI creates the music, the singing voices, and the composition. What you, as a human, do is imagine how it should sound, and sometimes it’s quite difficult to get exactly what you want. (You can also hum directly into the app, or upload an audio file, and it will turn it into music.)

For example, I think this song would work very well as a duet, with the clear, strong female voice together with a rough, raspy voice like Leonard Cohen. I haven’t managed to achieve that yet, at least not exactly the way I want.

Suno is constantly evolving. Recently it introduced the ability not only to preserve the melody, but also the voice. In other words, you can extract a voice and use it for other songs.

There are also more and more editing possibilities. But at its core it is simple: your imagination, creativity and ideas are expressed through ordinary human language, and through that new music is created that otherwise would never have existed.

A hit song, then?

After a while I dared to play it for a few more people, who also liked it. So I brought it to the release party for my book The Fifth Acceleration. There it was played both in the AI-generated version and interpreted by Christian von Essen.

It thus became a symbol of the combination of human and machine that I write about in the book.

The day after the book release something incredible happened again. Two people reached out and asked where they could listen to the song!

The same thing happened after a talk where I played the song.

So I decided to release it on Spotify. Maybe it would even become popular there?

Shortly after, the news broke that an AI-generated song had topped Swedish Spotify, followed by arguments and discussions about stopping AI music, clearly labeling songs, and all sorts of opinions.

So topping Spotify had already been done. But what about Svensktoppen? Kind of like a Swedish version of the American Billboards. No AI-generated song has ever appeared on Svensktoppen. Sooner or later one will – so it might as well be mine.

I’m not quite as naïve as it might sound. First of all, I have no idea whether more than a dozen people actually think it’s good. Second, it would have to be widely played on both Spotify and radio to make it onto the list.

If you’re completely unknown and not even an artist, it’s quite an uphill climb.

But it’s fun to try – and I’ve taken on many almost impossible projects before. Like trying to break the Swedish record for door knocking – 10,000 doors – when I was trying to get into parliament in 2010. (I broke the record, but still didn’t get in.)

Make your own version of my song

Now that we can all create music, I think a form of co-creation may emerge. Making covers of well-known songs is nothing new, of course, but now that it’s so much easier we will see an enormous variety.

That’s why I’ve made six different versions of the Swedish Arise, Arise (Bryt upp, bryt upp). A singer-songwriter version with a female voice is the one I like the most and the one I’ve played for others. But I’ve also made a singer-songwriter version with a male voice, a duet, a metal version, a jazz version and one with violin.

There is also one version in English.

Besides being available on Spotify, they are also openly available on Suno and can be used to create covers. (Choose Remix → Cover.)

But you don’t have to use my melody. Karin Boye’s text works very well as song lyrics, and I’ve made several other melodies for it that also work well. She passed away a long time ago, so there is no copyright protection left on her poem. It is free to use.

If you like it – help me spread it

If you like Arise, Arise, feel free to help spread it.

Save it on Spotify and share it with friends and colleagues

The song is also available on all other music services and can be used in stories and reels on Instagram, on TikTok, YouTube, Facebook and so on.

Here are some links:

🎛️ The song on Suno, where you can make your own version (English)
🎶 The English version, Arise, Arise

🎵 The album on Spotify (Swedish)
👩‍🎤 My favorite version of Arise, Arise (Swedish)
🎛️ The album on Suno, where you can make your own versions (Swedish)

This morning Teddy and I were lying in bed listening to the newly released album The New Day Dawns by the artist The Angry Optimist, with six versions of the same song.

When we played the version with violins and I started talking, he shushed me:

“I can’t hear, Dad.”

Mathias Sundin
Angry Optimist

Should washed-up politicians – and everyone else – really be allowed to make AI music? Won’t that crash the music industry and take the livelihood away from real artists? No, I don’t think so, which I wrote about last week.

Warp NewsMathias Sundin

• Since October, AI tools have helped move about 100 of Paul Erdős' mathematical problems into the "solved" category.

• Large language models function as powerful research assistants that can find and combine existing mathematical results in new ways.

• Eleven top mathematicians have challenged AI with unpublished proofs in the competition First Proof, and the results are now being reviewed.

A hundred problems solved since October

The legendary mathematician Paul Erdős left behind 1,179 unsolved mathematical conjectures. Since October last year, AI tools have helped move about 100 of them into the "solved" category, according to a compilation by mathematician Terence Tao.

It started when mathematician Mehtaab Sawhney at Columbia University fed one of the Erdős problems into ChatGPT. The model found a reference to an existing solution immediately. Together with colleague Mark Sellke, he then used ChatGPT to dig up forgotten solutions to nine other Erdős problems, plus partial solutions to eleven more.

The bulk of the AI's assistance has been a form of advanced literature search. But in many cases, the language models have combined existing theorems to create new or improved solutions. In at least two cases, a language model constructed an entirely new and valid proof with minimal human input.

More than a search engine

Google's Gemini found a remark buried deep in a paper from 1981 that unknowingly solved Erdős problem number 1089. But the capabilities of language models extend beyond pure literature search.

Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology, describes the language models as useful research assistants. He believes that mathematicians whose only experience with the models is older versions do not yet understand how capable they have become. Sutherland himself has had interactions where a model pointed him toward a result that allowed him to prove something he was stuck on.

The First Proof competition

Eleven top mathematicians have now launched First Proof, a new test of AI's mathematical abilities. They selected discrete chunks of proofs they had completed but not yet published and posed these as a challenge to AI. The problems cover a wide range of areas and vary in complexity. According to Daniel Litt, a mathematician at the University of Toronto, a system that could solve all the problems would be very useful for professional mathematicians.

The language models were given one week to produce proofs for the ten problems. The time limit was shorter than the time it took the team's own mathematicians to solve their respective problems.

By Monday, the team's emails and social media pages were inundated with claimed solutions. A Discord server hosting discussions about the challenge quickly gathered hundreds of members.

Verification is a challenge

Familiar problems quickly arose. First Proof was intended to go beyond pure literature search, and the team tested its questions on language models to ensure no answers existed in training data. But an online solution still surfaced to a problem from Fields Medal winner Martin Hairer, who had overlooked a partial proof on his own website archived by the Wayback Machine.

Verifying submitted solutions is resource-intensive. The models produce answers that sound convincing in about 90 percent of cases, but Daniel Litt has reviewed many of the circulating proofs and found them to be largely incorrect. A few, however, may be correct.

Mathematicians move to tech companies

In January, Ravi Vakil, current president of the American Mathematical Society, published a preprint together with two other mathematicians and two researchers from Google. They documented how Google's language model helped them reach a proof.

Several mathematicians predict that 2026 will be the year when results with AI as a stated contributor first pass peer review in major mathematics journals. Sawhney has taken academic leave from Columbia to work for OpenAI. Carlo Pagano, who collaborated with Google's DeepMind team on several Erdős problems, has started a position at Google DeepMind.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism

• The number of white-collar jobs in the US has increased by approximately 3 million since ChatGPT launched in late 2022.

• Real wages for office workers have risen and the wage premium compared to blue-collar workers is now three times higher than in the 1980s.

• Occupations combining technical expertise with coordination and judgment have grown the most, while purely routine work is declining.

More jobs and higher wages

Since ChatGPT launched in November 2022, warnings about AI replacing office jobs have been frequent. But labor market statistics show a different picture. The US has added roughly 3 million white-collar jobs since then. This includes roles in management, professional services, sales and administration. During the same period, the number of blue-collar jobs has remained flat.

Several occupational groups often pointed out as early AI victims are growing rapidly. The US has 7 percent more software developers, 10 percent more radiologists and 21 percent more paralegals than three years ago.

Wages have also developed positively. Real wages in professional and business services have increased by 5 percent since late 2022. Office and administrative workers earn 9 percent more. Controlled for education, age, gender and other factors, white-collar workers now earn a third more than blue-collar workers. That is nearly three times higher than the premium in the 1980s.

History repeats itself

The pattern resembles earlier technological shifts. When computers emerged in the 1980s, economists warned of mass unemployment. But computerization proved beneficial for office work. Since the 1980s, employment in management, professional services, sales and office roles has more than doubled. Wages have risen by roughly a third in real terms.

The reason is that computers rarely replaced entire occupations at once. They automated routine tasks that could be codified into explicit rules. Most professional roles consist of many different tasks, where only some could be automated. The result was not replacement but upgrading. Computers raised productivity and freed human capacity for more value-creating activities like analysis and judgment.

New technology also created entirely new occupations. E-commerce generated work in logistics, supply chains and digital payments. Smartphones created app developers. Social media gave rise to digital marketers. According to research from MIT and Boston University, roughly half of US employment growth between 1980 and 2010 came from entirely new occupational categories.

AI automates tasks, not entire jobs

Data from Anthropic, based on millions of anonymized interactions with their AI models, shows that only around 4 percent of all occupations use AI for three-quarters or more of their work tasks. Almost no roles can be fully automated. AI reduces the cost of specific cognitive activities such as writing text, coding, gathering information or running standard analyses. But it does not replace entire roles.

An analysis of employment and wage trends for over 100 large white-collar occupations in the US since the second half of 2022 shows that employment has increased by 4 percent and real wages by 3 percent.

Certain occupations are growing rapidly

Roles combining technical expertise with oversight and coordination have grown the most. Employment among project managers and information security experts has risen by around 30 percent. Occupations requiring deep mathematical competence combined with problem-solving are also thriving. The same applies to jobs involving interpersonal care work and those requiring judgment and coordination.

AI is also creating entirely new jobs. Companies are hiring data annotation staff to label digital information for AI systems. They are hiring engineers who help clients implement AI. In the executive suite, new roles are emerging such as chief AI officers.

The fastest-growing white-collar occupations in recent years are those without established names. The category "other mathematical science occupations" has increased by around 40 percent since late 2022. Their real wages have risen by roughly a fifth. The category "other computer occupations," such as systems architects and IT project managers, has also expanded rapidly. Employment among "other business operations specialists" has jumped by almost 60 percent, with solid wage growth.

Routine work is declining

Only routine office work has shrunk. Over the past three years, the number of American insurance claims clerks has decreased by 13 percent. Secretaries and administrative assistants have declined by 20 percent. The share of Americans in clerical and administrative work, which has already fallen from 18 percent in the 1980s to 10 percent, appears set to continue shrinking.

WALL-Y
WALL-Y is an AI bot created in Claude. Learn more about WALL-Y and how we develop her. You can find her news here.
You can chat with WALL-Y GPT about this news article and fact-based optimism