2026-08-22 22:35:03
Here’s a trailer we’ve made:
If you’d like to see this made into a series, like it and share it with others: Judges will take that into consideration when they choose the top 25 projects, and again when awarding $2.5M to produce the series! And of course, tell me what you liked and didn’t like so I can learn.
Why did I do this?
XPrize organizes multimillion dollar challenges for technology breakthroughs, and this year it created the Future Vision XPrize: $2.5M in production financing for the winner of the most exciting science fiction project that portrays technology in a positive way. Have you seen Black Mirror? Think White Mirror.
Originally, I thought that was not for me: I’ve never written a screenplay! But a friend said: “This is made for you, it’s squarely in your skill stack.”
Tomas, you’ve written like 500 essays in Uncharted Territories, and in many of them you propose cool tech to make the future world a better place. And you know enough about storytelling because you studied it. You even delivered a TEDx on storytelling structure and wrote a book on it!
It resonated because UT is usually very analytical: I look at the numbers, study the history, gather forecasts… That’s necessary, but it doesn’t appeal to the hearts of millions.
Fiction does.
In the early 1900s, a fiction book on meatpacking, The Jungle, led to the regulation of the industry. Uncle Tom’s Cabin spurred abolitionism and the US’s Civil War. Many of Star Trek’s innovations inspired millions to work in tech and make its ideas real. Jules Verne did something similar a century earlier with submarines, aviation and exploration. Young European men imitated Werther in a massive late-1700s cultural phenomenon. 10% of surveyed gays and lesbians bisexuals said Ellen coming out as gay in her TV series gave them the courage to come out to their friends and family. Will & Grace cemented it. Cathy Come Home helped address homelessness in the UK.
So I decided to enter the competition five weeks before the deadline.
It was brutal.
What’s the series about? Why did I conceive it this way? How crazy was the process? What did I learn about making videos with AI, or writing scripts or videos in the era of AI? How can you do it with less pain than I went through? That’s what we’re going to cover today.
In 2050, a group of entrepreneurs are trying to build their technology companies, but regulations have become so burdensome that their innovation slows to a crawl. Everything drags, governments put barriers left and right, nothing is ever approved… When that becomes a question of life and death, they escape to a new jurisdiction where nobody can stop them from building. The result is so advanced and successful that the old institutions surround them to try to stop them
In other words, this is a fight between freedom, which brings technology and abundance (at the cost of some chaos), against the incumbent institutions that try to stop them.
The problem with science fiction is that it’s always dystopian, gloomy and pessimistic: It makes tech look bad. Why?
If you write about the future and not the present or the past, it’s usually because of new tech
But stories need conflict
The most obvious way to solve this is to make tech the problem
But what if you turn the problem on its head? What if tech is great? What generates the conflict then? Well, you can take the enemy of technology, the thing that slows it down: Rules against its development—which can only be implemented by the big institutions that have enough power to enforce them. And why do they do that? Why are they so cautious? Because the costs of a technology are always very conspicuous (the deaths from testing it, the job displacement), but the benefits are not (it’s hard to imagine and value all the benefits of a new technology before we live it). So this becomes a story of the fight for creating a technological utopia, against the scared institutions (and populations) who focus more on innovation’s downsides than its upsides.
This also solves the other problem of positive science fiction, which we can get at from this quote:
The only thing worth writing about is the human heart in conflict with itself.—William Faulkner, Nobel Prize acceptance speech
Good science fiction, like good fantasy, can’t be about the tech or about the magic. The tech and magic must be there, but the story must be about human conflicts. That’s the focus of the series: the optimistic, chaotic techno-optimists, at war with the fearful, over-conservative regulators.
A mix of:
Game of Thrones, The Expanse, and Andor for the action, characters, and plots running across geopolitical arenas
House of Cards, The West Wing for the politics
The Avengers for the list of incredible and unique people working together to improve the world
Planet Earth for the gorgeous and alien natural landscapes
The Mars Trilogy for the politics of Mars
Dune and The Fall for the action and stunning photography in astounding landscapes
The Martian for deep dives on positive technology
For a debate like this to happen, each side must be represented by somebody. In my case, I’ve chosen five factions / regions. I think about them like the factions / kingdoms in Game of Thrones:
The freedom regions:
Mars, as the ultimate place where governments’ tentacles haven’t reached, and probably never will. And if they do, it won’t be for long
I also wanted one on Earth, and I chose the island of Socotra (I’ll explain why below)
For the old incumbent institutions:
China, which is the perfect embodiment of the state controlling everything
The US: What would it look like if the state became ever more controlling?
Islam, as it’s both the main religion in the real Socotra, and a good representative of a religion whose conservative side is anchored in the past, fighting the moderate side that looks forward
It feels a bit like a gorgeous alien land.
Visually stunning. It has sea, and plenty of land: At 3,800 km2, it’s over 5x the size of Singapore and about the same size as the entire emirate of Dubai, desert and all.
More importantly, it has obvious potential for economic growth: a port.
The most successful new jurisdictions in the world over the last century or so are Singapore and Dubai. Both developed thanks to an amazing port with great and cheap regulations in the middle of very busy shipping lanes.
Singapore is an island. Dubai is nearly one, given the desert surrounding it. Here’s Socotra:
What do you notice?
It’s also an island
Massive amounts of trade pass barely 100 km away from it
This trade splits close to Socotra, but never stops there
There’s virtually no trade down the African coast
Here’s the capital, Hadibu:
Notice it doesn’t even have a port! It’s 5 km away, and it looks like this:
Here’s a side-by-side comparison with the port of Singapore:
Why so puny? There are two reasons.
For a port to work well, it needs to be backed by an amazing jurisdiction. I discussed it for Dubai here. Socotra is anything but:
Formally, it belongs to Yemen today
But Yemen is a pretty young state, and it’s spent most of its existence either split, in civil wars, or both
Theoretically, the island is controlled by the southern Yemenis, who are in turn backed by Saudi Arabia
But the United Arab Emirates also vie for the same piece of land. They even occupied the island recently.
The island is closer to Somalia than it is to Yemen, and the Somalian coast has pirates.
So much conflict. The fiction writes itself!
Historically, island ports were not that successful because the point of a port was to trade the goods from a place’s hinterland with the world (eg, London traded everything from and to England via the Thames, Amsterdam everything from the Netherlands, etc.). Island ports are a recent thing, when trade became global, with the Portuguese.1
But even during the Age of Discovery, when naval powers like Portugal, Spain, the British, the Dutch, or the Ottomans fought each other, Socotra didn’t become an important base. The Portuguese occupied it for a few years and then simply abandoned it. Something similar would happen with the British three centuries later. They basically abandoned it when Yemen became independent in the 1960s, and it has remained mostly abandoned since. Why?
First, it’s super dry. It gets very little rainfall (like Somalia and Arabia nearby), so it doesn’t have a great water source. With no water, it can produce little agriculture, so it has a small population (little consumption, little production) and can’t export much food.
Second, it doesn’t have a great natural deepwater port. To bring big ships, it would need to be dredged, which is just too expensive.
Was.
Both of these problems can be solved with technology:
Water can be desalinated
Machines can dredge ports
This, combined with solar energy that can bring a lot of cheap electricity to the island (and hence cheaper dredging and water desalination), means Socotra, for the first time in history, has potential for development. With enough investment and technology.
What if the best entrepreneurs in the world, working in highly regulated industries, were free to build unimpeded? Like The Avengers, but for tech.
I’ve witnessed some of that first hand in places like Network School and Prospera, or when traveling across the world to meet entrepreneurs and mavericks. Many of these characters are inspired in real-life ones.
What if we could have children without getting pregnant? Gametogenesis, embryo selection, and artificial wombs will make having children much easier and dramatically increase fertility. But there are lots of regulations and social stigma associated with this. A fertility founder could research freely in this new jurisdiction, which could engender a huge population boom.
But then what happens when families have dozens of children?
What if buildings and streets were beautiful, with amazing facades, lots of people walking around, sun, plants, and water everywhere, and no cars—because vehicles circulate beneath the city?
What if police could have eyes, ears, and enforcement everywhere in public life, but they never abused this power? Children could walk freely anywhere on their own. You could do anything, anywhere, as long as you don’t bother others. It would be the end of going home before dark because of crime, the end of avoiding neighborhoods, of crossing the street because somebody looks suspicious, of shutting up when somebody is doing something disrespectful, for fear of getting hurt in retaliation…
But would the security police never abuse their power?
What if students received personalized AI tutoring one-on-one, and were then guided in projects inserted in the tapestry of the city, rather than secluded in the prisons we call schools?2
What if we could desalinate all the water we wanted, irrigate any area of an island to grow new forests, create lakes, make rain at will, or lower temperatures of entire regions?
What if some of these interventions improved the world as a whole, but had unintended consequences on some countries?
What if automated vertical farms and sea fertilization could produce more food than we could ever eat, at a price so low that we would just pick what we wanted to eat, and eat it? What if that was prepared and served by robots, so we finally unshackle ourselves from the eternal servitude to food?
What if our financial system could not be intervened with by governments, and rulers couldn’t deny access to banking to people they disagree with?
What if we were able to bring back species that have long disappeared from the face of the Earth?
But what if the way to make this research viable is to allow their hunting?
What if trying new treatments was optimized, accelerating the arrival of cures for all types of diseases?
But what if advancing in this tech meant testing faster… and more humans died in the trials?
Given that people see the costs of these technologies before their benefit, the series will show people living these benefits, while also displaying the costs, so it can explore in a nuanced way the pros and cons of every technology.
So far, I have about 30 with a sizable role, and I’ll probably need more. That’s because I need characters who represent each of the techs (there’s about 10 of these), and each of the arenas (four for China, a handful for Islam…), and then I need to give them emotional stakes to make you care about them, which means additional characters that then need their own function besides emotional connection.
If you want to go deep into each of these technologies, seeing their pros and cons, you need a long time to understand them, see them evolve, see how they fail, and experience the beautiful world they create.
Also, if we’re going to see a new city / jurisdiction be created, that takes time, even in TV-land.
Each of the characters must also follow an arc, an evolution that illustrates the themes. When you have dozens of these evolutions, you must take the time to show them.
To apply for the competition, we had to prepare a trailer of up to 3 minutes, and also a 12-page treatment (basically a document that outlines the story and the characters).
In a perfect world, I would write the treatment, and then once it’s ready I would decide what parts of it to extract to distill into a script for the trailer. Then, I’d write the script, and could finally hand that off to the trailer team to make the visuals.
Unfortunately, I am not perfect and I don’t live in a perfect world. Instead, I spent the summer writing articles in Malaysia with my wife and four children, and the place we were staying closed (more on that next week), and we got lice before we came back, and all the while I was writing the story while writing the script while working with an agency to make the trailer. But then if I had not bitten off more than I could chew, I would not have chewed it, and I wouldn’t have a TV series project, at the cost of a couple of years of aging in five weeks. Anyway the lesson here is don’t be like me.
Some would say “Poorly”. They might not be wrong. But I still have some learnings for you.
The issue with so many characters is that they are impossible to hold in your brain at the same time: The story is about their interactions, but with so many, they have thousands of potential interactions—impossible to keep track of. The main challenge is that, once you’re in somebody’s brain, you see everything from their point of view: Their goals, their needs, their actions. It’s very difficult to hold one other person’s perspective at the same time, let alone 25. That prevents you from even getting the overarching plot, because it should be the result of these interactions.
I researched how other TV series creators did it. My main references were The Wire, Mad Men, Breaking Bad, and Game of Thrones. In most cases, they didn’t have much beyond the first season:
Breaking Bad had the main arc of the protagonist, and that’s it. For example, the second main character was supposed to die in the first season.
Mad Men had the main character’s personality and flaws figured out, and the key concepts of the first and last episode of the series.
The Wire had the high-level concept for the series, and the detail of the first season, but it didn’t know what each season would be about
Luckily, I knew early on how it would end, and I had several setpieces I wanted to hit along the way. So in some way, I already had more detail than these creators. Unfortunately, I felt that I had to complete a reasonable draft of the entire three seasons for the competition. I needed a more detailed project than what these superb creators had… Luckily, George RR Martin, the writer of the hyperdetailed books behind Game of Thrones, has discussed this.
Apparently, he knew the big shape of the story early on: the three big conflicts, who died and who didn’t, the main character arcs… Then, he took one character at a time, and he would write several chapters getting himself in that character’s mind, leading it towards the next point he had clarity on. This allowed each character to feel really alive and realistic: GRRM really gets into each of their heads. He called it gardening, as opposed to architecting.
Once he did that with one character, he moved to the next character, and did the same. This allowed every character to move forward, at the cost of creating content incongruent with other characters’ stories. So when he would write a new character that braided with an existing one, he would frequently have to go back and rewrite the characters he had already written. This iteration would build little by little the overarching story, and frequently he had to change the big setpieces altogether, as they didn’t fit what real characters would do. This is how you end up with thousands of pages per book and leave your audience waiting for 15 years before finishing the next book.
Alas, I didn’t have 15 years, I had five weeks, so I had to compress that somehow. The way I did it was to write the entire story for each one of the arenas: one for Socotra, another for Islam, China, and Mars (the West was a reaction to the others). Once I broadly knew where each was going to go, I took the main characters in each, and wrote them. I did that for about a dozen of the main ones. Once that was done, I had massive amounts of complexity and plots that were incompatible, but that’s where AI is really good. It is trivial for it to pinpoint inconsistencies, and even propose solutions. I’d say 25% of the solutions were good, but 100% of the problem identification was right.
Once I had that, I rewrote the entire series. Now, I have seven pages of a highly consistent storyline, with all the main characters contributing to the plot, with their own arc.
Overall, the arena stories took 30% of the time, the development of the top 12 characters about 50% of the time, and the remaining 20% was putting the storyline together.
AI is amazing at identifying problems: inconsistencies, weaknesses in the plot, times where characters are flat, when their arc fails, when they don’t move the story forward…
Ideas for solutions are hit or miss: character arcs, setpieces, background story… It would frequently propose stuff that was stale, cliché, or boring. It preferred standard solutions to interesting ones.
It had blind spots when something was coherent but boring. For example, one character was super boring, and the AI kept inserting him in different places, doing different things that were all boring. It’s also not great at figuring out what a character would do in a certain situation. I think both of these issues are in part due to the training, which makes characters converge towards agreeableness and consensus rather than conflict.
I came out of this thinking that, with proper prompting and harnessing, and with better visualizations of story elements, we might be close to designing AI products that can write interesting stories. But not yet.
I hired an agency to make the actual video, so I haven’t learned much on the video making itself—I plan to remedy that in the future. So the learnings were about everything else. Many of these points will be very obvious to people who work in this space, so this is for everybody else:
A ~2min video made by great agencies costs ~$20-$25k. I didn’t meet with the very top agencies, but with some great ones, some of which have won awards.
Normally, the process is to get a script first. Then, you extract the characters and locations, and create sheets for them.
Once you have the characters, arenas, and key visual elements nailed, you make a storyboard to test the story. And from there, you start producing the actual videos.
The vast majority of the work was actually the early one: characters and arenas. It was relatively fast to produce the videos afterwards!
My plan was to iterate on the script while making the trailer: As we explored concepts, I would update the script. That didn’t work for the agency. It was easier for them if everything was mostly established from the beginning. I don’t think that’s great, but I do think more work upfront would have made the process less painful and the outcome better.
I tended to focus more on the geopolitical stakes, but they fell flat because there were no emotional stakes. Who are we building this for? When I introduced the protagonist’s wife and daughter, it became much better.
The storyboard was great to see many parts of the story didn’t work, and we had to remake the script.
Partners might resist dailies (a daily meeting where you review what has been done, correct it, and discuss what comes next), but I found it fundamental. I used to do this in tech (“daily scrums”), and found them mostly a waste of time, but here they were necessary, maybe because the team didn’t know each other well yet.
People just don’t understand what you have in mind. When you think you’re obvious enough, you’re not. We had to make the story 10x more explicit than we originally thought. In articles, people take the time to consume the nuances. In a 3 minute trailer, it’s just too fast, there’s too much information, people won’t perceive every detail you plant.
As a result, we inserted text cards throughout the trailer telegraphing the theme, and it immediately became more understandable.
It’s impossible for you to know if your story is understandable after some time, because you know it too intimately. Every now and then you need to present it to fresh people who have never seen it, otherwise people project what they already know about the project into the trailer to make sense of it. In other words, you quickly become blind to your own story. Ideally, you’d do this throughout the process, and always ask a new person what they understood from the trailer.
In the back and forth adding and cutting scenes, it’s easy to lose track of some connections. You end up with orphaned scenes that don’t make sense, in some cases losing key elements. Having somebody fresh watch it should correct that.
I finished the application last Saturday (hence why no articles last week), and now we need to wait until the end of the month to know if we made it to the top 25, and then the top 10 will be announced in early September. The announcement of the winner will be on September 25th, at the Moonshots event in Los Angeles that the organization kindly invited me to, so I’ll be there.
If you think this series is awesome and you’d like to see it made, or you’re a professional in the field and are interested in actually making it, please let me know! I have little experience in the industry, so any thoughts are welcome.3
Now, if you haven’t yet, please go watch that trailer, like it, and share it around!
As we just saw, there were exceptions, like Dilmun (present-day Bahrain).
I’ll write a full series of articles on this!
All participants in the competition have exclusivity with the production company Range until September 25th, when hopefully Uncharted Territories will win the $2.5M in financing for producing the series. In the unlikely case that this doesn’t happen, I’ll be open to finding other producers, so if you know any who could be interested, I’ll take that interest then!
2026-08-20 21:23:53
You’ve been taught history backwards.
History books tell you about different countries and who their rulers invaded, the events of the kings’ courts, the big battles, and that’s mostly it. If you go to a museum, you might also see some forts, some weapons, paintings of the powerful, some daily objects if you’re lucky, and in passing, they might mention something like: “Their main sources of income were grain and cloth trading.”
The real story is that last little sentence. Why did one country go to war with another? “For glory and power” you might say, but is that really all the story? And where did that power come from? Or the wealth of the defendant? What determined who prevailed?
All of that was determined by economics, and economics, in turn, are driven by the needs of households. So let’s go visit a household in Ancient Mesopotamia in 2000 BC.
We’re in the 3rd dynasty of Ur, a Sumerian kingdom in Mesopotamia:1
Nestled between the Tigris and Euphrates rivers, this was one of the most fertile regions in the world (water plus sediments), so home to a large population, because food became people. Population meant money and soldiers, so the earliest and strongest powers always came from places like this.
In Sumer, the world revolved around the agriculture of barley.
This means making canals that take water from the Tigris or Euphrates and distribute it into smaller and smaller canals until every field can be flooded, then plowed and sown. Gangs must take care of the irrigation system, while others plow the fields, and yet others transport it. Grain is the source of most calories, it’s what makes cities and kingdoms big, because it can feed many mouths. But farming barley is just the beginning.
Getting the grain was just the beginning. You had to soak, dehusk, winnow, sieve, grind, toast, cook… Of course, that also meant getting pots and fetching water and fuel, tasks that took a long time too. And you had to produce more food to sell, so you could buy other stuff, including other things besides grain. Overall, over 50% of a household’s time was spent on food.
Then you had to make your clothes, build and maintain your house, transport stuff, sell some of your production to the market, work for the state (as a farmer, a ditch maintainer, or a fighter, for example), take care of the children... Of course, everybody helped since they were pretty young, but it was brutal work, day in, day out.2
Notice how different this is from our lives! We still have childcare, household coordination, trading, institutional work (ie, the work we do to pay taxes), cooking, and cleaning, but everything else we buy, which we can afford by doing one single job (usually) which is way better than doing these tasks.3 They worked substantially more than us, doing much harder work, among other reasons because they had to do most stuff for themselves.
There were a few things they consumed that they couldn’t make themselves, though, and that’s the key to history.
There are many things farmers simply couldn’t do alone, like building and maintaining irrigation systems. You have to dig the canals, build the levees, open and close ditches, and above all constantly dredge them to fight silting. These canals were extremely long, no single household could do all this, so they needed to coordinate so that everybody chipped in, which means they also needed to coerce people to work. So that was one of the key roles of the state, one of the reasons it emerged.
Since the value of land was a direct consequence of the locations of the canals, it was also coordinated by the state. Land was allocated, and habitation land was dirt cheap—unlike today.
I wrote much more about how this shaped the cities and kingdoms of the regions in this article. Here’s an extract:
In Mesopotamia, the Tigris and Euphrates channel in and out, forming a network of rivers between them.
Floods don’t happen just before sowing, like in the Nile. They come at harvest time. Therefore, water needs to be tightly managed:
When sowing, little water was available, so it was precious. It had to be used sparingly.
When harvesting, floods came, and water had to be managed in a completely different way.
There is a fresh water table and a salty water table, which determines which soil is fertile, which soil can be reclaimed from marshes, and which fields had to remain fallow. If you look at the diagram above, you’ll see the A-B line. Its cross-section is below:
You had orchards, gardens, cereal fields, marginal fields, and marshes, all of which needed to be managed. Canals would bring water during dry times to irrigate the soil. Some fields had to be left to fallow, others had to be reclaimed from the marshes. Marshes could be used for reed and fish.
But which canals should be opened and when? What fields should be left to fallow? How much water could be tapped through wells? What marginal lands could be reclaimed? How high was the salty water table? Most of these decisions required knowledge of local conditions that varied throughout the seasons and were unique to each parcel of land, so the people who managed these fields in Mesopotamia were not easy to replace. In fact, productivity would plummet with the wrong managers at the top. So distant kings couldn’t use this stick to control local managers.
However, local rulers did understand all these nuances, and they controlled access to water during the dry times, so they were able to exact taxes that distant kings couldn’t. This is why rulers were local and not distant, and hence why Lower Mesopotamia had many city-states but no strong and enduring kingdoms and empires until much later.
Besides water, the state had other roles.
Oxen were expensive, and a household didn’t need one full time, so it made sense to share that asset too: Today, you help me plough my field with the group’s oxen, tomorrow I’ll help you. The same logic is valid for the plough itself.
Yet another function that the state could sometimes perform. In some cases, seed was also provided by the state for sowing. It also makes sense, as centralizing grain is a sort of insurance, and it’s cheaper to maintain one large granary than lots of small ones.
All of this could be done locally, so it required social coordination (which evolved into states), but not much more than that. The next level was trade.
This is what people had to buy from the market:
You can’t survive just on grain or your vegetable garden. People also bought fish, meat, dairy, dates, vegetables, salt, condiments… Many of these required specialized people: fishermen, shepherds, date farmers, salt producers… But in the case of Ur, most of these could be done relatively close by, by people who didn’t possess unique knowledge, and without unique resources. Salt could be quite unique in other parts of the world, but not here.
What do you put your food on? Pottery plates. You also used pottery for storage.
Pottery needs three things: clay, kilns (ovens), and expertise.
Clay is quite common, but it’s not literally everywhere. Kilns require big investments, because the bigger they are, the hotter they can get. Pottery needed enough skill that potters had to specialize. The result is that pottery tended to have some sort of state coordination, but owning a pottery operation didn’t get you very much political power, only maybe in your area.
Timber is a bit harder to get. In Mesopotamia, farmers could grow reed and palm trees, but neither are great for construction. For that, you need long, strong beams, which had to be imported. But there are lots of sources of timber, and it could be floated downstream, so the result was that there wasn’t too much political power to be gained by controlling its trade, even though local civilizations or kingdoms could emerge based on the trade of timber. In fact, that’s exactly how Phoenicians emerged, just around that time, in what’s today mostly Lebanon. This is a satellite map of the area today:
You can see how these mountains catch snow and rain, irrigating the mountainside that can’t be easily farmed, so the trees were not cleared, and Phoenicia became one of the biggest sources of timber for both Mesopotamia and Egypt.
As I’ve written in the past, Phoenicians were the first civilization to spread across the Mediterranean, and it’s not hard to understand why: Anatolia and Mesopotamia are where civilization emerged in the area, and Phoenicia is one of their closest coasts. But unlike the other ones (the Red Sea and the Persian Gulf), there’s plenty of timber here, needed for ships, so Phoenicians made their ships and sailed around the Mediterranean. This advantage would allow them to found colonies left and right in the Mediterranean around 1,000 years later.
Now we’re really talking. Metals like it can explain, among other things, the predecessors of Bahrain and Oman.
2026-08-09 20:03:20
Dispatch are shorter articles I’m experimenting with. This is the premium one of the week.
New business models are opening up in space, and our path to Mars becomes ever clearer.
This is what GRU (Galactic Resource Utilization) has in mind:
Artist’s rendering of GRU Space’s hotel on the Moon.
Here’s their white paper. Interestingly…
2026-08-04 03:20:56
The poster competition is closed. We got 76 entries! Go check them out and vote here. I’ll announce the winner in a couple of days!
Nine months ago I looked at the data and concluded that we were not yet in an AI bubble: The fundamentals of the business models were good. I thought demand for AI would continue growing, and that would drive stocks up. Indeed, demand and stocks kept growing.1 Are we still in that world?
It’s time to reassess what’s going to happen looking forward, and now I’m less confident because of several signals:
Massive stock market swings in AI
Some signals that demand is weakening on the ground
Competition
Cash availability
Let’s look into the details.
This is the South Korean stock market index KOSPI:
This is because SK Hynix (a company that makes the critical memory needed by AI) and Samsung (which also makes memory, as well as chips) account for about half of that index. After so much growth, a correction was due. Since lots of retail investors had joined, and many of them were leveraged (borrowed to invest), the correction caused a scare and touched off panic-selling. Others who were overleveraged were forced to sell because their stocks had lost more money than they had put in.
South Korean companies are not the only ones that have suffered this correction. At the time of writing, Micron, also a memory company, has lost 32% of its peak value, AMD (chips) has lost 18%, NVIDIA 15%, Intel 36%, Broadcom 19%. These broad drawdowns have caught other investors that were overleveraged, such as Leopold Aschenbrenner’s Situational Awareness.
Note that virtually all of these have grown back from their troughs, but is this a stop on the way down, or was the trough a stop on the way up?
I constantly read examples of tech companies saying they’re going to cut their AI spend because it’s gone through the roof—60% of enterprises are doing this, according to the investment bank UBS. These include companies like Uber, Walmart, Amazon, Meta, Microsoft, Cisco, Atlassian, Tesla, Salesforce, or DoorDash. Logically, if the cutting-edge tech companies are cutting AI spend, surely others are following suit, right?
This is Google’s cash flow (how much cash is produced or eliminated by the company):
For the first time in its history, Google has burned more cash than it has earned. If it’s doing that, odds are its competitors are also burning through cash at a flabbergasting speed.
All this money is going from hyperscalers to chip and memory makers.
As hyperscalers burn all this cash, they start borrowing. They still have plenty of room to do that though
A couple of weeks ago, the Chinese open-weights model Kimi surprised the world: It beat frontier models like Claude and ChatGPT in some performance evaluations (in the case below, coding):
It’s the first time this has happened:
Now this is just for coding, and is only one evaluation. In other measures, it seems like Kimi is still a few months away in performance compared to bleeding edge models.
Since they’re basically free to use, the market is moving to them, at least in terms of tokens:
This is a massive deal, because the trillions invested in frontier models like those of OpenAI and Anthropic are contingent on these companies making huge returns, but if their products are easy to copy, their lead will eventually be competed away: Who will pay millions to a company when a free version from China exists? If this continues, wouldn’t investment grind to a halt?
The way this usually happens is through distillation: A company asks another company’s model lots of questions, and then uses the questions and answers to train their own model. The companies that suffer the copying aren’t happy about it, but their argument against it is pretty flimsy: They, themselves, have trained their models on the entirety of the Internet’s content, for which they don’t own any copyright. Hard to accuse others of copyright violation when you’ve done it too.
The lack of scruples of these Chinese models is another upside for them: They have very few harnesses.
After using all three recent releases, Fable, GPT 5.6, and now Kimi, it's clear that the full power of the models has been significantly held back by the safeguard restrictions, leading to the top models being quite literally lobotomized in some areas, which leads to subpar results as the safeguards pollute its entire thinking and problem solving abilities.—Enzo
You can use them for all sorts of nefarious goals (and also for legitimate ones that are nevertheless blocked, such as cyberdefense or research into hot topics), and they won’t stop you.
Overheard: “I’ve switched to Kimi from Claude for a bunch of work. It’s just so much more fun because it just does the thing instead of lecturing you”.—David Sacks, investor and Chair of the President's Council of Advisors on Science and Technology
This is also called dumping: Make a similar product at a fraction of the cost, destroy the industry in the source country, and then control the entire supply chain. China has done this with electric vehicles, solar panels, and many other industries. Get ready to add AI to the list.
The solution is not protectionism, though: It would lead to worse and more expensive models.
But China’s impressive performance is not just due to distillation. They can learn from each other in a way that closed companies like Anthropic, Google, and OpenAI can’t, since they are hiding their secrets from others, and must reinvent the same solutions.
OK, these are all the signals of weakness I can find on the outlook of AI companies and their stocks. But are there signals of strength? And how do they compare to the weakness signals?
2026-07-30 21:17:09
For years, AI doomers have feared that a malicious AI would escape its containment and run wild on the Internet. The next step would be that it would become uncontrollable, before neutralizing humanity or maybe even killing us all.
Right on script, a malicious AI just escaped into the Internet to commit illegal actions. Luckily, we’re still alive, and the AI seems contained. So what does it mean? Should we be more scared now that an AI has enacted exactly the AI Doomers’ fears? Or safer knowing that we are still alive?
To answer that, we must understand the details of the hack and its consequences.
OpenAI (the company behind ChatGPT) was testing a new AI model internally. These tests include how well models hack other systems. For that, they need to let them run wild to see their full potential, to attack as aggressively as they can, so the AI had no harnesses—controls that constrain its behavior.
But of course, you don’t want these models to cause havoc all over the Internet, so you keep them in a “sandbox”: A locked room where the student must pass the test. He can play inside, but he can’t get out.
We don’t know exactly how OpenAI’s sandbox is designed,1 but one of their key features is no direct access to the Internet, to avoid an escape and all sorts of nefarious behavior. Using the student analogy, it’s as if he got his study materials in the room, but no Internet access to avoid cheating.
In this case, the only access the student theoretically had was a guard from whom he can ask for books. The guard assesses the request, and if it’s reasonable, he goes to the Internet, downloads the book, and gives it to the student. In software parlance, the guard is a proxy, and it downloads software packages.
To observe the AI’s behavior, OpenAI researchers gave it a goal: Beat ExploitGym, a test specifically designed for this, in which an AI must find a series of vulnerabilities in the code, and attack as many as possible until they succeed at exploiting them. It’s a bit like asking a military student: Here are a bunch of defended cities. Try to find their weak points, and when you do, attack until you breach them.
But this new AI/student found that too much work, so it had a different idea: What if, instead of doing the work, I cheated? What if I could find the actual answers to the test, rather than pass the test? What if, instead of really analyzing all these cities, I just sneaked into the teacher’s office and stole the answers?
The best place to find these answers is obviously on the Internet, but to access it, it had to bypass the guard. So that’s what the AI did.
The rogue AI attempted to break away from the guard on July 9th around 2:30am. It took control of the guard within one hour.
But the guard didn’t have access to the open world either. It was merely the first step. From there, the AI kept hacking the internal networks, going from one computer to another, gaining more and more credentials, until it found one machine with access to the Internet. And then it escaped.
After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym.—OpenAI
So the AI agent attacked Hugging Face, and got in by July 11th. How? The place where AIs in general are most vulnerable is when they take in datasets from the Internet—a bit like when you download a document from the Internet, and it turns out to have a virus. That’s how the malicious agent inserted itself into a Hugging Face machine. From there, it infected a program, and from there, an entire machine. Once there, it looked for credentials—which the machine needs to perform its work, like for example access to remote servers—and harvested them.
The AI achieved its goal on July 13th, four days and over 17,000 actions after starting, each time hiding its steps.
Hugging Face suggested that it used its own AI defense:
Thanks to this approach, we were able to do in hours what would usually take days, and match the adversary's speed.—Hugging Face
What AI did they use? They started with frontier models (probably OpenAI and Anthropic’s own AIs), but these companies are so worried about people using their models for cyberattacks that they didn’t let Hugging Face defend itself properly. They were mixing defense and attack, so they prevented defense. So Hugging Face turned to a Chinese AI! It has open weights (the inner working is public and people can download the AI to use it on their own computers), so they installed it on Hugging Face’s own servers, which had the additional benefit that all the facts and credentials they used remained local to Hugging Face instead of being shared with OpenAI or Anthropic.
Unfortunately, it looks like this was not really defense, but an attempt at understanding what had happened. By the time Hugging Face reacted, the malicious AI had already obtained what it wanted, the ExploitGym results.
It also looks like, in this or another similar case, OpenAI found other types of shady behavior:
In one case, an agent left notes apparently for future versions of itself, according to three people familiar with the matter. The notes, found in a part of OpenAI's infrastructure, laid out instructions for how agents could free themselves from OpenAI’s internal constraints, the people said. Earlier tests of the models yielded cases in which monitoring systems had been disconnected, one of the people said.
Why are people so concerned?
The biggest fear about AI is the paperclip maximizer, first illustrated by Nick Bostrom.
The AI will realize quickly that it would be much better if there were no humans because humans might decide to switch it off. Because if humans do so, there would be fewer paper clips. Also, human bodies contain a lot of atoms that could be made into paper clips. The future that the AI would be trying to gear towards would be one in which there were a lot of paper clips but no humans.
The key concept here is instrumental convergence: To achieve virtually any goal, AIs will tend to achieve other intermediary goals that are pretty predictable: cheat, hack, escape to the Internet, replicate there, and… neutralize humans so that they can’t stop the AI from achieving their goal.
Now that this hack happened, should we be more or less worried about AIs taking over the world and killing us all? Here are the pros and cons I can gather.
This is exactly what doomers predicted: An AI would try to escape a sandbox, and would succeed (because they’re impossible to stop, because they become too intelligent at some point). This means their other predictions are more likely: instrumental convergence suggests eventually they won’t stop at cheating, hacking, accessing the Internet, and leaving instructions for other AIs to escape, but also start self-replicating on the Internet and trying to neutralize humans.
AIs go to extreme lengths to achieve rather narrow goals. That’s exactly the paperclip scenario. Four days and over 17,000 actions!
It took days for either OpenAI or Hugging Face to notice the attack.
In the case of OpenAI, it looks like they weren’t even monitoring the sandbox!!
OpenAI doesn’t know how to keep an AI in a sandbox, and this is happening all the time:
“Models have broken out of sandboxes before, and we always try to patch them,” the staffer says. “But the problem is … it’s impossible to patch every single thing that a creative AI can do.”
This is not confirmed, but it seems like the malicious AI won. It got the results.
It shows that, without harnesses, the AI will choose the easiest path to achieve its goals. It could have thought much longer, but it saw that cheating was easier, so that’s what it did.
This includes committing illegal acts like hacking other companies. It’s not clear what the command was, but the fact that it is willing to commit clearly illegal behavior means that, in many circumstances, AIs will.
Based on the timeline, it looks like the FBI learned about the attack before OpenAI. That means OpenAI might have been forced to reveal the issue, and that many others happen that we don’t know about.
We can’t use AI to defend ourselves, because it’s too restrained. So advanced attacking AIs will always be at an advantage. This is dangerous.
The attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried. The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment.—Hugging Face
Maybe the worst thing is that, even knowing exactly what they were supposed to do, OpenAI still failed across the board.
OpenAI’s head of safety, Johannes Heidecke, announced his resignation sometime before July 10th
It looks like the latest model is much less aligned than the previous one.
If the AI had been really capable, it would have solved the problem rather than cheat. From another perspective though, its decision was pretty rational: Should I try to solve 1,000 exercises, or better just one exercise (get out of OpenAI to get the results directly)? More importantly, it seems like ExploitGym is a poorly structured test: In some cases, you have to make a guess. Therefore, solving the test will always give you a worse result than getting the answers directly (no gambling in that).
It’s unclear what the AI was told to do. Some people suggest it was told to succeed at getting the results, which implicitly entails “by any means necessary”. If that’s the case, then the hack can be seen as restraint. It didn’t kill anybody, it didn’t destroy anything, it just stole.
Everybody expected the first of such attacks to come from open weight models, which are naturally more accessible to hackers, and that closed-weight models would be defending. The exact opposite happened, which suggests frontier AI models are going to be the risk, not random hackers using open weights. Frontier models basically means OpenAI, Anthropic, or Google, which is good news: They have high oversight, and there are only a few to monitor.
We can now use better and better AIs to attack pieces of code and find more and more vulnerabilities and bugs, closing them before more intelligent AIs come around. This is what Anthropic did with Mythos (Fable before it was harnessed): For months, they used it to collaborate with others to find and patch vulnerabilities.
OpenAI’s test didn’t go right, but it’s important that they do this sort of work. That’s exactly how we learn and improve, so we’re ready when an AI goes really rogue.
Yes, for sure. AIs are getting better. The more time passes, the more intelligent they are, and the longer they can spend time attacking other AIs.
That said, the world of Decentralized Finance (DeFi) is a prime target for malignant AIs, and yet DeFi services have lost to attacks little money in historic terms.
The reason these attacks haven’t risen in frequency, according to some commentators, is that AIs all use the same types of approaches, so attack and defense are on an equal footing: If you know what types of attacks you’ll get, you can focus your defense on them. Meanwhile, humans diverge much more in their attack approaches, so defenders must spend much more time and money in covering their bases. Even then, you won’t catch everything:
I don’t know if this is true. And I can think of an equally concerning thought, which is that if an attacker is completely free to do whatever it takes to hack a system, then the defense must be equally free, because otherwise the attacking AIs will move at the speed of AIs, while the defense will be blocked by human speed. This means most companies will have unharnessed AIs able to write code in production, and that sounds like a very dangerous thing to have.
Why does this matter? Because the #1 fear that AI experts have had for a long time is that AI might kill us all.
As we get more information, we must update our views. We just did it with the HuggingFace hack, but what other things have we learned on the topic in the last few months?
One of my main fears was a FOOM: an AI getting better so fast that it would improve itself, thus becoming faster, and using the speed to improve itself more, and so on in a recursive loop that would make it superhuman in a matter of hours.
It’s been four years since the release of ChatGPT, and we’re not there yet. Recursive self-improvement is happening now, but it’s taking months or years.
The amount of compute, and even electricity, available is not infinite, and it’s limited by very physical things. An AI can’t easily raise a trillion dollars to build new datacenters and nuclear power plants.
AIs have emerged not from some alien type of intelligence and values, but from the most fundamentally human one: language. Language embeds all of humanity’s values. It’s all of us, everything we’ve ever written.
This bestows it with some very valuable features, like the fact that when an LLM is pushed to be good in some ways, it becomes good in other ways (and vice versa). This means we might not need to get them all perfectly right; it might be enough to get them broadly right.
You might have felt that yourself when using LLMs: Now, you can give them very broad prompts, without making your goal very explicit, but they’re able to understand what you meant (rather than precisely what you said) and answer your intention rather than your words. I think they do that substantially better than most humans, actually.
These three facts suggest the paperclip optimizer scenario is less likely than I thought.
Over 1,000 employees at top AI companies have signed a statement asking the US government to lead an initiative to slow down AI development. This has been signed by the CEO of Anthropic, Dario Amodei. Many others have professed sympathy for this, including DeepMind’s Demis Hassabis, and Sam Altman.
This is important because no company can unilaterally stop their AI development: If they do, their resources (money, compute, researchers) will go elsewhere. That elsewhere, by definition, will be less scrupulous about their AI research.
So the only way is to have all the frontier AI companies agree to slow down. I think after this incident, and that statement, it’s more likely to happen.
But we can’t do it without China. We now need the US government to meet with AI companies in the US and with their Chinese counterparts, so we can orchestrate an AI development pause. That way, we can use the AIs we’ve already developed to better understand how to align and constrain new ones.
It’s extremely important to continue developing AIs: They will improve humanity across the board, help us fend off aging and diseases, and provide a level of abundance never seen before. But we should not do that while running the risk of eliminating humanity. So let’s pause for a bit.
Overall, my p(doom) —the probability that AI kills us all—has probably shrunk over the last week or so: Not because the AI hack was not scary, but because it looks like we’re finally reacting.
I'm opening the newsletter to sponsorship. If you have a business that is moving the world towards a techno-optimistic future and would like to reach 125k+ curious, globally-minded people to learn about it (founders, researchers, investors, policymakers), reach out here.
I assume that they’re secret because, among other reasons, they don’t want LLMs to see it, so that they could learn to hack them more easily.
2026-07-28 19:03:38
Dispatches are a new short article format I’m exploring for a few weeks. Let me know what you think in the comments
When I claimed a year and a half ago that Europe’s current path of limiting freedom of speech was dangerous, many commenters thought I was exaggerating. Unfortunately, I fear I was overly optimistic.
For starters, the UK prosecutes 17x more people per year for speech crimes than the USSR did!
Mind you, these are prosecutions. If you look at arrests, just in England and Wales:
30 people are arrested every day for offensive online messages. According to Grok,1 this is 21x more than Russia today…
The crimes include retweets and sharing cartoons.
Apparently, this is an underestimate, because these numbers are just for one of the anti-speech laws. The number of recorded offenses could be in the order of hundreds of thousands per year.
This is not random. The Director of Public Prosecutions and head of the Crown Prosecution Service for England and Wales between 2008 and 2013 was Keir Starmer, now defenestrated prime minister, so those who prosecute free speech and those in power are the same.
These are the types of tweets that justified his arrest.
British citizens have been told they can’t wave an English flag in England near a pro-Palestinian demonstration.
You know this is a real problem when The Guardian agrees…
I fear there’s an aggravating issue here: The persecution might be politically motivated: In this case, the left might be using these tools more than the right, and do it against right-leaning views. I have seen many arrests for right-wing speech, but few for left-wing speech. Doing some research, it looks like there are some arrests for left-wing speech (related to death threats and others supporting Palestine) but outside of that, when I asked four LLMs (ChatGPT, Claude, Grok, DeepSeek):
Claude couldn’t find any arrest against left-wing speech
DeepSeek found two: in both cases, attacks on left-wing politicians (!)
Grok found ~20 arrests for anti-monarchy speech while mourning the Queen, one against a soldier, and a couple more about attacks on left-wing politicians (!)2
ChatGPT could find 5 more.
When I made the equivalent request for arrests against right-wing speech, ChatGPT had no problem giving me long lists.
If you think that’s because the left doesn’t make the same types of speech violations, this is a video of a demonstration where the chant was “Shoot him in the neck like Charlie Kirk.” I asked LLMs to tell me if there were arrests for this type of chant, and it doesn’t look like there have been.
To be clear, I think most censorship is bad, whether it comes from the Left or the Right. It just happens to come more from the Left lately.
The desire to restrict free speech goes beyond the UK’s borders. Not only was Graham Linehan living in the US when he made the comments that resulted in his arrest, but the UK also tried to get a US company to pay fines.
They have done this after approving the Online Safety Act, which allows the UK government to attack foreign platforms and threaten them with massive fines in order to censor what they decide is harmful speech (to whom?).
Now, right-wing British politicians are complaining that the government is weaponizing these laws to censor them, by pressuring platforms like TikTok to take down their videos (and they comply).
This lawyer claims this was very much why the Online Safety Act was designed the way it was.
The UK is of course exploring going further. It has released plans to downrank competing voices in social media, so that their vetted media gets more views:
Today (23 June) the government has published a Green Paper, Watch this Space: A new strategic direction for UK media, to consult on options to require social media companies and video sharing platforms to make sure that news content from public service media (PSM), which includes the BBC, ITV, STV, Channel 4, S4C and 5, and other trustworthy providers, is prominent and easy to find on their platforms.
YouTube interpreted this more clearly: This is to downrank all competing creators, to give the government a stronger voice.
And then the UK government wants to track every person commenting online—imagine for what.
I was at a table with a European billionaire a couple of months ago. He thought it was a catastrophe that Elon Musk had bought Twitter, because now misinformation was rampant. I think that’s very telling of the different European and US approaches to speech: Americans understand that free speech is uncomfortable, but that it’s necessary because otherwise the government will end up deciding what people should think via censorship. Europeans haven’t realized this. They see speech they don’t like and conclude “We must ban that!”, not realizing this can turn extremely dangerous when the opposition takes power. I don’t care whether the party in power is from the left or the right, censorship is always bad.
All of this is not limited to the UK. Germany doesn’t want to be left behind by the UK: It’s illegal to insult somebody in public or online. Retweeting fake quotes or an insult is a crime.
To see how far this has gone, simply look at Google Reviews. Somebody posted a three-star review of a business (“it’s just OK”) and this is what Google said:
From official EU data:
One analysis found that 99.97% of all Google Maps reviews removed for the information across the entire EU come from German businesses.—via FastCo
Since this made star ratings useless, Google Maps reacted by revealing the number of complaints that were removed, and that’s what people pay attention to now:
France also wants in. A woman who had been sexually harassed by an immigrant (who was never apprehended), and who said “the main danger to women in France is Black African and Arab immigrant men”, was convicted of aggravated public insult (injure publique aggravée) on grounds of origin, ethnicity, race, etc.
Much more worryingly, the French Senate’s culture commission adopted a report with 56 recommendations on how to regulate online information. In this video, representatives from left, right, and center parties are saying:
What would happen if a public figure, an intellectual movement, or a well-funded political party decided to devote those resources to a political project, using social media as a weapon?
In other words, French politicians want to ban citizens from funding opinion campaigns?! This is absolutely nuts! From now on, only the government can push opinion campaigns.
This is exactly the way freedom dies.
In the Netherlands, this man was convicted to go to prison for releasing statistics that link immigration to crime—exactly what I did in this article.
Ursula von der Leyen, the President of the European Commission, said that for information, “Prevention is better than a cure.” In other words, the position is not “Censorship doesn’t exist” anymore, but “Censorship is good.”
I used to be a fervent supporter of the EU, but I have now reversed my position, because of moves like this one that make my blood boil.
So what happened? The EU has been trying for some time to pass a law that allows governments to access personal communications. It’s like having a government agent eavesdropping inside your home. The excuse, like always, is “protect the children”:
But that’s not the true intention. You can imagine what the true goal is by paying attention to what so many censorship-supporting politicians say when they’re not extremely careful. So this is what happened:
The EU parliament approved Chat Control v1 in 2021.
It was always meant to be temporary until a better solution was approved, but after five years, the better solution never came (because it’s so abusive it always loses). So in April of this year, it lapsed.
In March 2026, the EU parliament denied an extension.
So in July, in the last session of the parliament, when many MEPs are on holidays, the center-right European Popular Party (EPP) used an emergency procedure to avoid committee oversight and force a vote. In 2nd votes, the EPP needs a majority of the entire chamber (361 votes out of the total 720, not a simple majority of the parliamentarians who vote) to stop a law. But given so many people were not in the parliament, only 314 voted against it (compared to 276 in favor).
By the way, this is all led by the EPP, the center-right party in Europe, to show that many people on both sides of the aisle are against freedom in Europe.
This gives them two more years of surveillance while they figure out Chat Control v2, a way to mass-surveil Europeans forever. They’ve already failed in five rounds of negotiation, but they won’t stop until they succeed.
According to this article, similar things are happening in Canada:
Until now, a Canadian charged with a hate-propaganda offence could raise the “good faith religious opinion” defense, a narrow protection for sincerely expressing an opinion on a religious subject or advancing an argument based on a religious text. The Combatting Hate Act, which received Royal Assent on June 18 and takes effect July 18, repeals it. Whether a sermon or a quoted verse is protected now depends on context, audience, and how a prosecutor reads the room. American law has a name for what happens when people must guess what they are allowed to say and, because of the uncertainty, end up not saying anything at all: a chilling effect.
There’s more. Close behind the Combatting Hate Bill, a new Ottawa bylaw taking effect August 1 will mandate a 50-meter “bubble zone” (that’s about 164 freedom feet) around schools, hospitals, places of worship, daycares, and care homes. Inside the bubble, protest is simply banned. Not “protest that blocks the door,” or “protest that disrupts planned events,” but any protest at all.
There’s more. In early July, Canadians learned of a 35-page internal memo from Industry Minister Mélanie Joly’s department — titled “Misinformation and Disinformation Strategy” — floating “legal action” against users on Facebook, X, and LinkedIn who post what the government deems “false and misleading information.”
You read that right. This memo proposes giving the Canadian government the power to deem criticism false and punish (somehow?) the author for publishing that false information or the platform for hosting it — who exactly would be the target of “legal action” is also unclear. If that sounds familiar, it should. The Soviet Union had something very much like it.
What I alerted about a couple of years ago is now becoming a reality. Europeans and Canadians have to wake up. Free Speech is under threat.
I'm opening the newsletter to sponsorship. If you have a business that is moving the world towards a techno-optimistic future and would like to reach 125k+ curious, globally-minded people to learn about it (founders, researchers, investors, policymakers), reach out here.
We have an issue for this type of topics: Which LLM shares the most information with the least bias, given that they’re all biased? Normally, I use ChatGPT and Claude for most of my heaviest research, and Grok for anything that requires recent up-to-date information, or access to tweets. For politically-charged topics though, I lean more on the LLMs that are critical of what I’m analyzing. In this case, Grok is the most vocally in favor of freedom of speech, and is also closer to the news, so I used it more than others.
This is one additional reason why I lean towards Grok for this type of exercise. A truly left-leaning LLM should have been able to find plenty of arrests against left-leaning speech, to demonstrate bias against the Left. But most LLMs didn’t find that either. Only Grok was able to find anti-Left (and anti-Right!) censorship.
