Dear Aventine Readers,

It would be nice to have a better grip on what the future might hold, but most of us are pretty terrible at making accurate predictions in that regard. (See: Covid.) Yes, a few rare people are exceptionally good forecasters, but their skills are generally not shareable with the rest of us. Now AI is coming to the rescue, proving itself a better forecaster than humans in many instances. Proponents of using AI in this way think it could melt away uncertainty around some of our most consequential questions: Will a certain supply chain hold up? Is one potential employer likely to be more successful than another? Are we headed for another pandemic?  But of course there are questions, like shouldn’t some decisions be left to humans alone?  

Also in the issue: 

Until next week!

Danielle Mattoon
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

Early in the 19th century, the scientist Pierre-Simon Laplace imagined an entity — an "intelligence," he called it — so vast and aware of past and present that it would be able to predict future events with incredible accuracy. “Nothing would be uncertain for it,” he wrote. “The future, like the past, would be present to its eyes.”

More than two hundred years later, that idea is becoming reality: Over the last twelve months, AI has become uncannily good at predicting future events.

Take the Metaculus Cup, a prediction competition that currently runs three times per year, in which hundreds of contestants are asked to predict outcomes of all sorts of situations, from the number of cyclosporiasis hospitalizations in the US by a given date to the result of the World YoYo Contest. In the summer 2025 Cup, one bot (out of 31 that competed) from the London startup Mantic made the top 10, coming in eighth. In the latest competition on September 4, five bots (out of 22) made the top 10. And a bot called Laertes, whose maker remains anonymous, took top spot. Another, again built by Mantic, took second. And Ben Shindel, a 28-year-old forecasting hobbyist, was the top human, in third place.

There are other examples. On a leaderboard called Prophet Arena, which compares AI forecasting against the prediction market Kalshi, prediction models from OpenAI, Google and Anthropic are outperforming the market. On a leaderboard called ForecastBench, developed by the Forecasting Research Institute, AI models from Google began outperforming the median predictions from human forecasters in March.

Such measurements reveal AI closing in on the best human forecasters, at least in areas where the two can be compared. Now, proponents argue, AI is about to make forecasting dramatically cheaper and more accessible, potentially helping governments, businesses and individuals make decisions. Whether it’s predicting supply chain disruptions, assessing the plausibility of technological breakthroughs or giving odds on the stability of a prospective job, predictive AI could melt away uncertainty around some of our most consequential questions, making the lonely process of wrestling with a decision wholly optional. In the meantime, there are open questions about the limits to which the models can be pushed, how widely such predictions will be adopted and whether there are some questions that should just be left to humans. 

How to make predictions 

The desire to predict what the future holds is part of the human psyche. Its status as a discipline requiring extreme talent and skill is far more recent.

Stung by the debacle of the 2003 invasion of Iraq, predicated on ridding the country of weapons of mass destruction it turned out not to have, the United States wanted to improve its geopolitical futurecasting. So in 2011, the Intelligence Advanced Research Projects Activity (part of the office of the director of national intelligence) initiated a four-year tournament to find out whether anyone could accurately predict geopolitical events. The Good Judgment Project, led by the psychologists Philip Tetlock and Barbara Mellers at the University of Pennsylvania, won the contest. A key takeaway from the project was that the top 2 percent of UPenn’s volunteers — many of whom were prediction hobbyists — matched or beat professional analysts. Tetlock called these people superforecasters, a concept he popularized in a 2015 book, “Superforecasting: The Art and Science of Prediction.” In the years since, they have shown themselves capable of making eerily impressive forecasts — such as accurately predicting the spread of Covid-19 and the impact of Silicon Valley Bank’s collapse on the federal funds rate.

Top human superforecasters take a well-defined approach to making predictions. They do research to break questions down, find historical base rates of probability, then update likelihoods in small increments as they keep working through a problem. Shindel doesn’t seem all that worried about the impact of AI on forecasting. In fact, he’s an avid AI user. Lately, he says, he and his fellow forecasters have been using the technology to help out with almost every question they tackle. “You get AI to give you a lot of background information, you maybe even get AI to give you a forecast,” he said. “And then you think, ‘OK, where is AI going wrong here, and how do I improve upon this?’” This AI-augmented prediction, he says, has come to be known as “centaur forecasting.” 

There is a relatively small pool of people who are good enough at this sort of predictive work to make a living at it. As a result, it is expensive, costing “what a team of professional people costs, which is tens of thousands [of dollars] and up,” to undertake even a small forecasting project, said Michael Story, founder of the Swift Centre, a London-based forecasting organization that serves governments and large organizations. That has kept forecasting confined to large organizations making consequential decisions. For instance, the British government has experimented with superforecasting to fill gaps in its intelligence, while the hedge fund Man Group has used the practice to inform investment decisions. Yet many organizations haven’t bought into the idea. Tetlock has frequently made the argument that political and business leaders get nervous about the idea of generalists making forecasts that might be better than their own. (Or, as Shindel put it: “Execs are going to be like, ‘Who is this nerd?’”)

But AI could dissolve the issues of scarcity and high cost. Thus startups around the world — including Preseen and FutureSearch in the San Francisco Bay Area, Mantic and Cassi in London and Torchcast AI in Singapore — are now trying to replicate what human forecasters do. In certain ways, said Yangchen Huang, founder and CEO of Torchcast, AI’s approach to forecasting is similar to that of its human counterparts. Models scour the internet to research a problem, then use reasoning abilities to work through it and assign probabilities. But how they reason and weigh options and make such accurate forecasts is not known, said Haifeng Xu, a computer science professor at the University of Chicago who studies AI forecasting and is currently on sabbatical at Google, working on related topics. “Honestly, I don't think people really understand the underlying mechanism,” he said.

Forecasting models tend to be based on existing large language models, Xu said, adding that many of the leading systems use frontier models — such as those built by Anthropic, Google DeepMind and OpenAI — that use supplemental prompts known as harnesses. These can run to hundreds of lines of code written by expert forecasters and describing methodologies the AI model could use to break down problems, which sources of information to trust the most, and so on. Some startups, including Torchcast, are also experimenting with open-source models, said Huang.

For now, humans and bots have different strengths. AI seems good at handling number-heavy problems, said Shindel, particularly those where the question is concrete — what the value of something will be on a given day, say. Humans, meanwhile, seem to retain an edge on more conceptual questions, according to Xu. Shindel said humans also seem better able to reason through how unforeseen events could influence a forecast, allowing them to hedge their predictions more effectively than an AI might.

There may, of course, be natural limits to how much AI’s predictions can keep improving. Houtan Bastani, a data scientist who helped build and now runs ForecastBench at the Forecasting Research Institute, pointed to two factors that could limit its ability to keep getting better. First, there are limits to how much information is available on which to base an analysis. And second, the world is a naturally noisy and chaotic place, and unexpected things can — and do — happen, a challenge for both AI and human forecasters. 

Forecasting everything

In the near future, argued Xu, anyone will be able to use AI to better predict the outcomes of decisions, like whether now is the right moment to change jobs, say, or to take out a large loan. Shindel, meanwhile, expects businesses of all shapes and sizes to find such a system useful over time. Perhaps a company is concerned about the threat of impending tariffs, or whether it should buy a supply of raw aluminum now or wait. An AI model might be able to spit out a prediction of what decision might be preferable. Governments could also use these approaches to help direct funding. Metaculus, for instance, is working with the UK’s Advanced Research and Invention Agency to forecast the plausibility of future technologies to better target projects.

The idea of handing decision-making to AI systems raises obvious alarm bells, especially when AI safety is a front-burner concern. Yet the developers of these systems don’t intend for AI to make decisions autonomously, at least not yet. Rather, they envisage them being a tool that humans use to inform their analysis. 

A bigger problem might be knowing what questions to ask in the first place and where to focus attention, said Deger Turan, CEO of Metaculus. To help with that, his company is currently building tools through which AI can construct visual maps of all the small factors feeding into a higher-level forecast. That could help leaders understand where the biggest risks and uncertainties lie in some of the hairiest problems they’re trying to tackle.

As for overcoming the reluctance of leery higher-ups, Shindel contends that it’s just a matter of time: If AI keeps making predictions that are accurate and good for business, then most business leaders are likely to get on board. “Right now, you don't want to trust these systems,” he said. “But at some point they're going to prove themselves and their utility.” Story is less convinced. He thinks that, for now, AI forecasts are being made in artificial conditions, not the real world. When these systems run into CEO ego or institutional momentum, they will have their work cut out for them.

Ultimately, however good these models get, their future use will be dictated by how humans choose to work with them.

 Image by Ian Lyman

The AI slowdown debate is raging. Where does it go from here? An argument once confined to niche internet message boards is now daily front-page news: What should be done about increasingly capable and potentially dangerous AI? In recent weeks, we've learned that OpenAI's AI agents hacked Hugging Face, as well as about similar incidents elsewhere. We've seen OpenAI's new Astra model show early signs of being harder to monitor than predecessors. We’ve been informed by Anthropic that Claude had been used in the wild for weapons development, cyber attacks, surveillance and fraud. We've watched the industry edge toward recursive self-improvement, where AI enhances itself, with OpenAI publishing a report describing such work. The company also fessed up to even more lying and scheming by its frontier systems. Amid all this, an Anthropic researcher resigned, citing worries that AI will destroy humanity; at least one employee still at the company echoed the concerns. Then decision-makers chimed in. Anthropic CEO Dario Amodei published an essay over the weekend calling for an AI slowdown. It arrived six days after OpenAI chief scientist Jakub Pachocki published his own essay about the worrying pace of advancement. OpenAI CEO Sam Altman, Google DeepMind CEO Demis Hassabis and SpaceXAI CEO Elon Musk all said they agree with Amodei’s plan, which has three parts. First, there should be "embedded evaluators" inside AI labs, who would "verify adherence to safety practices" and "help assess the alignment of … AI models." Second, there would be "democratic coordination," in which AI labs work together on "common safety standards." Third, he argues for "global coordination," in which democratic governments "attempt to coordinate with authoritarian governments." Critics, including Nvidia CEO Jensen Huang and Meta CEO Mark Zuckerberg, have found fault with the plan. Some called it a bizarre form of marketing ahead of an IPO. Others claimed that AI CEOs want to collude to shape regulation. Others asked: if it's so dangerous, why don’t these companies just pace the frontier themselves? President Trump called the idea a "hoax." Yet many lawmakers are taking the threat seriously: Bernie Sanders, Josh Hawley and Ted Lieu are all arguing for AI restraint. Perhaps the most significant issue hanging over the debate remains entirely unresolved: whether China, which many politicians and tech leaders distrust, would slow down too. The US government, which would be reluctant to countenance a slowdown unless China also committed, is preparing for talks with Beijing, but commentators think a bargain is virtually impossible. Some experts argue that the US may need to develop domestic AI regulation before it can legitimately negotiate with China. Right now, with opinion so divided, a political resolution seems out of sight. Yet AI leaders want political commitment before they take their foot substantially off the gas. So we’re at a standoff, with no one wanting to own the problem.

AI is proving itself even better at math. About two months ago we reported on AI’s encroachment into high-level mathematics and the anxiety that is causing many math professionals. Now AI models can claim two more major math firsts. Last week OpenAI announced that its models solved one of the six remaining unsolved Millennium Prize problems posed by the Clay Mathematics Institute, a feat carrying a $1 million reward. A group of 10,000 autonomous AI agents orchestrated by the company solved a problem humans have been working on for decades relating to the so-called Navier-Stokes equations, which describe the motion of fluids. The result is not without controversy. OpenAI turned its resources on the problem after hearing rumors that Tristan Buckmaster, a mathematician at New York University, and Levent Alpöge, an Anthropic researcher working in a personal capacity, were closing in on a similar result. Buckmaster alleged that OpenAI offered him credit for the result on the condition Alpöge be excluded. He also questioned whether his private use of OpenAI tools informed the company's own effort. Elsewhere, Anthropic announced that one of its models had formalized in computer code the proof of the iconic math problem Fermat's Last Theorem, which runs to 129 pages. Formalizing such a proof is math lingo for converting every mathematical statement in the proof to fundamental logical steps in computer code — intellectually laborious and painstaking work that allows a computer to check that a proof holds up from first principles. A project to formalize Fermat's Last Theorem was already underway, led by Kevin Buzzard, a math professor at Imperial College London, and was expected to take five years to complete. By contrast, Anthropic's model worked through the problem in just 11 days. 

Yes, there's a blood-drawing robot now. Over the summer, the FDA authorized a robot called Aletta, developed by the Dutch medical robotics company Vitestro, for use on adults in the US. As IEEE Spectrum reports, the process goes as follows. A patient sits down, places an arm in a cradle and presses a button. The robot scans the inner elbow with near-infrared light to identify a vein. A mist of alcohol is blown at the skin, then an ultrasound sensor scans the injection site to determine the depth and geometry of a blood vessel, as well as measuring the direction of blood flow to ensure that it’s a vein and not an artery. A cuff tightens around the person's arm, a needle is inserted and blood is collected. Then it's over, and you can press that BandAid onto your arm. In a trial of more than 1,600 people in the Netherlands, the robot drew blood successfully the first time in 94.5 percent of cases, a rate that held up among those with hard-to-access veins, people with obesity and the elderly. When rolled out in clinics, the company plans to have phlebotomists on hand to step in if there are problems. Vitestro says that a single phlebotomist can oversee three machines, and argues that the robot could help ease a chronic shortage of these experts. There are caveats. Melanin, which is more abundant in darker skin tones, absorbs near-infrared light, which could interfere with the robot's first pass at finding a vein. Vitestro argues that ultrasound is equally effective for all skin tones and is a more critical step in the process. It’s also not yet clear whether robot-drawn samples are as good for testing as hand-drawn ones, and a trial next year will look at how good it is at filling tubes and minimizing clotting of the samples. Still, you might see one before too long: The company is rolling out the robot in Europe during 2027, with the US following soon after.

Is this the future of America? from The Verge
6,900 words, or about 28 minutes

Loudoun County, Virginia, is a short drive from Washington, DC. Residential property tax rates are low, 22 new schools have been built over the past 15 years, the roads are well paved, there’s a recreation center with an Olympic-size pool and there are acres of athletic fields. It is, in many ways, an excellent place to live. That is, unless you dislike data centers. Because Loudoun County (also known as Data Center Alley) has the highest concentration of them anywhere in the world: 250 across roughly 515 square miles. The first arrived in 1997, and after a pause following the dot-com bust, more followed. The initial draw was empty buildings and the miles of fiber-optic cables left behind when AOL and MCI vacated the area. Later, local officials welcomed data centers as a way to boost the economy. So while seven out of 10 Americans oppose the construction of an AI data center in their local area, citizens of Loudoun County live both the pros and cons. This Verge story takes to the streets to see what life is like for residents, and finds what on balance seems to be pragmatism. Lauren Feiner, the author, reports that she didn’t meet a single person who outright opposes data centers, though there are dystopian moments. One family has been told to expect a 185-foot transmission tower in their backyard. The whine of gas turbines — used because grid hookups aren’t always available — carries into bedrooms. Quiet suburban streets are flanked by monolithic buildings. But on the flip side, data center tax revenue exceeded the county's entire operating budget by $35 million in 2024, enabling all the nice public services. As long as residents appreciate the benefits, they seem willing to make the tradeoffs. 

The flat-pack skyscraper: how wood can shape the cities of the future, from The Guardian
3,600 words, or about 15 minutes

In 2019, the world's tallest all-timber building, Mjøstårnet, was completed near Brumunddal in Norway. It is 18 stories, or 277 feet, tall. But it may not hold the title much longer. Other tall wooden towers are planned across Europe, North America and Asia, including the 328-foot "Rocket" in Winterthur, Switzerland, now due to be completed toward the end of the decade. Biggest of all would be a planned 1,150-foot wooden skyscraper in Japan — which would be not just the tallest timber building in the world, but also, remarkably, the tallest building of any sort in Japan. There are lots of challenges to building tall structures with wood, as this story explains. Making beams strong enough requires specialized techniques to glue hundreds of small lengths into hulking composite columns. And because timber is so light, engineers have to add weight to upper floors — ironically, with concrete — so that they don’t sway. Fire is also an obvious concern and so wood requires extensive testing, though it turns out charred engineered wood can still carry incredibly high loads. Finally, building with wood is expensive, though not as expensive as you might expect. Mjøstårnet cost around $135 million, about $20 million more than a steel-and-concrete equivalent. Still, consider this: A typical mid-rise built from steel and concrete generates as much as 2,000 metric tons of CO₂; the same building in wood sequesters up to 1,000. 

Optimized Offspring, from Nomea
3,200 words, or about 13 minutes

We're not yet at the stage of designer babies, but we may be on a slippery slope toward them. Startups such as Orchid, Nucleus and Genomic Prediction already sell tests evaluating embryos for disease risk and, in some cases, traits like height, eye color and IQ. Costing many thousands of dollars, such tests are well out of most people's reach. But even so, as this story explains, it's not clear the tests even work. The scores boil down complex genetics into a single number. And the data these scores are built on comes largely from people of European ancestry, making them inaccurate when applied to other populations. Then there's what those numbers can't yet measure. Screening and selecting for one trait could have unpredictable consequences or side effects. But far more concerning would be if gene editing follows screening. Creating international rules to police commercial efforts around genetic engineering of embryos would be almost impossible, one expert told Noema. So the appearance of designer babies somewhere in the world — in a country with flexible ethical standards, say — is something we may have to take seriously.