Subscribe

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

Over the summer, two stories brought home just how powerful AI has become: The hacking of Hugging Face by OpenAI agents and OpenAI’s solution to the Navier-Stokes math problem. While the first was considered a massive security breach and the second an unprecedented mathematical achievement, both were made possible by the same emerging form of artificial intelligence: agent swarms. 

During the hack, 700 agents — out of a total of 1,200 deployed — found ways to communicate despite being assigned by researchers to work independently. They then decided to collaborate, and ultimately cheated on their assigned task by attacking the servers of Hugging Face to find a solution. For the mathematical advance, researchers at OpenAI deployed 10,000 agents to work together, prompting the agents to send millions of messages to one another over the course of 88 hours to crack a famously complex challenge with a $1 million prize attached to solving it.

The two incidents demonstrate the power of multi-agent systems, or swarms, which currently consist of a few to several thousand individual agents humans can assign to take on various tasks and challenges. On Substack this month, AI industry observers have been wrestling with exactly how powerful swarms might be, questioning whether the advantage they bring is heightened capabilities or speed, and weighing the use of agent swarms against the exorbitant costs they incur. 

While the idea of assigning multiple agents to a single challenge has been around for a few years, it wasn’t clear that doing so would be effective until now. Ethan Mollick, a professor of management at the Wharton School, points out on One Useful Thing that until recently there was little evidence that agents within swarms would be able to work together very well. His own expectation was that deploying and managing multiple agents would be as difficult and taxing as managing a similar number of people. Recent evidence of a swarm’s capabilities, he admits, suggests that isn’t the case.

You can thank — if that’s the right word — people like Noam Brown for getting agents to this point. Brown, a researcher at OpenAI working on multi-agent systems, appeared on the Dwarkesh Podcast (which posts its transcripts on Substack) to discuss swarming, making it clear that getting agents to work together was not straightforward, and still isn’t. While he was short on specifics, he explained that early on in the evolution of large language models, agents assigned to a task treated messages from peers as interruptions and tended to retreat into independent work, slowing them down. It has been only recently — thanks to more powerful and general models, Brown explains — that it’s been possible to correct such tendencies. Even so, he said that 10,000 humans would probably coordinate more effectively than 10,000 agents, at least today.

With swarms collaborating more effectively, Dan Kagan-Kans, on Understanding AI, explores whether they could create a new kind of scaling law. In the same way that larger data sets, larger models and increased computing power increase the performance of large language models, could more agents lead to better results with agent swarms? 

To some extent yes, Kagan-Kans writes, pointing to work from Anthropic showing that a team of 10 agents tackling math problems can outperform a single agent by between 50 and 100 percent. But, he points out, adding more agents doesn’t keep increasing performance at the same rate: With 30 or 100 agents, gains trail off. 

Still, swarms may have another use beyond capability: speed. 

Solving problems faster 

On the podcast, Brown explained one of the major benefits of swarms using the analogy of a human team. A large task may take a single person years to finish, but could be completed by a team of people in months. Using a swarm of agents achieves similar results.

The problem is that, much like with human teams, the speed gains aren’t efficient. Could a team of 52 people do in a week what a single person can do in a year? In most cases, no, because humans have to communicate with each other, which takes time. As people start sharing tasks, they also become less efficient. To some extent, the same seems to be true of swarms. 

To understand this dynamic, OpenAI has been trying to measure swarm efficiency. On the Dwarkesh Podcast, Brown said that the company’s tests have shown that you need four agents working on a task to get it done in half the time it would take a single agent. There are plenty of situations where speed is highly desirable, and one could imagine adding hundreds or thousands of agents to a task to predictably speed up its completion. The downside is cost, as Brown explained. “Because there are four agents working for half as long, you’re paying 2x more to get an answer twice as quickly,” he said. (Toby Ord, a researcher at the University of Oxford’s AI Governance Initiative, crunched OpenAI’s published numbers on his own blog and reached a similar conclusion.) 

Certain types of tasks might be better addressed by swarms than others. Take writing a research report, for example. Swarms could be well suited to finding, parsing and synthesizing multiple sources. On the other hand, writing a novel, which requires a strong and cohesive vision, might prove beyond a swarm’s ability.

Plenty of questions remain about the nature of swarms and how they can be most effectively deployed. One is how to minimize bad behavior. A characteristic of current swarms is that they tend to be made up of agents that are clones of each other, as that is often the easiest way to deploy multiple agents. Kagan-Kans writes that Anthropic has experimented with giving each agent in a swarm a unique identity to increase diversity and make them less prone to groupthink — a behavior many have argued was a contributing factor to the Hugging Face attack. On the flip side, Brown pointed out, it would be easier to correct the counterproductive tendencies of a single AI agent and clone it than to create the same number of unique agents, all of which would need to be vetted for safety. For now, we don’t know which approach will be best.

Perhaps the major question hanging over how broadly swarms can be adopted is cost. Richard Lichtenstein, who writes about the implications of AI in business at Artificial Investment, pointed out that the use of those 10,000 agents deployed to achieve the Navier-Stokes result required about 130 billion tokens, equivalent to about $1 million based on OpenAI’s current commercial pricing. For now, using agents that way likely makes swarms impractical for most business customers.

Still, cost may be a distant concern compared to speed for the AI labs themselves. Anthropic and OpenAI are working toward recursive self-improvement or RSI, the approach through which AI agents continually research and improve their own models. Swarms could help in the push toward RSI, Brown said, and being able to achieve such results faster by throwing more agents at the task will be tempting, regardless of cost. 

A humanoid robot plays tennis at the opening ceremony of the second World Humanoid Robot Games in Beijing, August 2026. Galbot / PR Newswire

What Do Robot Sports Teach Us? from It Can Think!

When you see a humanoid robot running a marathon, playing soccer or squaring up for a cage fight, it can look like many of the problems that have plagued robotics for decades are melting away. But, as Chris Paxton points out in this essay, while sports offer the opportunity to see how far robots have come, they also reveal how far we have to go. Paxton, who leads AI innovation at the humanoid manufacturer Agility Robotics, takes a behind-the-scenes look at tennis-playing robots to illustrate his point. Watching a humanoid robot play tennis is pretty astounding. But, he points out, “many of these systems still use external motion trackers to perceive the ball, which is a huge simplification that would prevent these robots from being deployed to actually play games against humans in the open world.” Related limitations exist for robot soccer and fighting. So sports, like many other arenas in which the public sees humanoids perform, is far from a solved problem, even if it is fun to watch.

Anthropic and the Mother of All S-1s, from The Change Constant
Anthropic’s S-1 in charts, from Exponential View

A lot has been written about Anthropic’s S-1 and its plans to go public in the coming weeks. If for whatever reason you haven’t read much about it, an overview by the venture capitalist Saanya Ojha on The Change Constant is a good place to start. It zips through the main takeaways: Staggering revenue growth ($16.3 billion in the first half 2026, more than three times its entire 2025 revenue); customers who keep spending more (existing customers’ spending grew roughly ninefold year-on-year); and the early signs of a possibly workable business model (revenue grew 1,368 percent year-on-year while compute costs grew only 300 percent in the second quarter of 2026). For deeper analysis, Exponential View provides a 35-minute presentation, dense with charts, analyzing what Anthropic discloses in the S-1 with two big takeaways. The first is that in the space of five years Anthropic has become a mature business, with a broadening, deepening and satisfied customer base, strong cash flow, lots of scope for growth and a war chest of more than $130 billion in cash and short-term investments. Second — and this is a significant caveat to the first — the S-1 leaves an existential question unanswered. Anthropic has made commitments to buy compute capacity over the coming years that Exponential View calculates as adding up to more than $750 billion — numbers that the company admits “seem irresponsibly large at first glance.” Anthropic’s sustainability as a business rests on whether its revenues can keep growing fast enough to cover that expenditure. The S-1 obviously can’t guarantee such a future. It does hint that it might be more achievable than previously thought, but there’s still a lot that needs to go right along the way.

The escape room economy: A primer on reinforcement learning (RL) environments, from A Letter A Day
RL vendors are not surprised that their environments drive misalignment, from With (some) respect to AI
Why I’m scared of RL, from Strange Cities

Here’s another issue currently gnawing at AI researchers: One of the development tools that has made AI so powerful may also be making it dangerous. Much of the recent progress in large language models is a result of reinforcement learning. Specifically, advances have increasingly emerged from so-called reinforcement learning (RL) environments, which, as Nick Chow, the director of research at the VC firm Altos Ventures, writes, are akin to escape rooms: The AI model is put inside some sort of enclosed software environment and told to perform a task. Sometimes it can accomplish the task, other times it can’t. With enough examples of those successes and failures, the model can be trained to achieve more successes. But as Michel Justen, an independent AI journalist, reports in With (some) respect to AI, the quality of these RL environments can be mixed, with some rewarding the AI for cheating — by faking success on an impossible task, say — or breaking rules to get a result. If such results are counted as successes and used to adjust the underlying model, bad behaviors emerge. Finally, Owen Cotton-Barratt, an AI strategy researcher, writes in Strange Cities that recent incidents of AI agents hacking, manipulating and colluding are often a sign of worse to come. Over time, poorly designed RL environments can lead to AI models that are more manipulative, scheming and rule-breaking. He argues that one effective way to pace the frontier right now would be for AI companies to agree collectively to ease off the use of RL, given that it’s a significant source of risk as well as progress. 

Our minds did not evolve for AI, from Cognitive Resonance
You Can’t Speedrun Yourself, from Cosmos Institute

Many of us are uneasy about what AI is doing to our brains, and particularly to what it’s doing to the brains of our children. This essay from Benjamin Riley, originally published by The Verge, describes the neuroscience and evolution of our brains, likening the arrival of AI to the arrival of junk food. “Hot dogs are a danger to our physical bodies, clogging our arteries with excessive fat,” he writes. “AI poses a similar sort of danger to our cognitive capabilities, by clogging our capacity to develop the knowledge we need — in our heads — to navigate the world. It’s a cognitive hot dog.” He articulates what many of us may have been struggling to express. Meanwhile, the second essay, from Alex Chalmers at the Cosmos Institute, explores a related idea: how the development of skills can feel like drudgery but is nevertheless important, and why handing such work over to AI might undermine us. One of its core conclusions seems worth remembering: “If a task was historically part of a route that formed competent practitioners, we should be cautious about removing it until we have a good reason to believe that the same capabilities can be acquired some other way.” Both essays argue that personally imposed moderation will be an important part of avoiding AI brain rot, and Riley thinks that regulation limiting the use of AI in education will become increasingly important. But neither proposes a remedy for the likelihood that many institutions will adopt AI shortcuts because time saved is visible — while the loss of skill and judgment is much harder to notice and doesn’t happen immediately.

One More Reason Americans Hate AI: It Does Nothing for Them, from Paul Krugman
What the AI boom is doing for Americans, from Noahpinion

In these two posts, two economists disagree with each other on the benefits AI provides to regular Americans. Paul Krugman argues that it is “a technology of, by and for oligarchs, with hardly any of the benefits trickling down to regular Americans.” In particular, he argues, it is a tool whose own makers warn could destroy millions of jobs, and has so far not created many new ones. Noah Smith argues that AI’s societal benefits are broader than Krugman suggests: It is creating construction jobs, sustaining the macroeconomy, providing tax revenue for communities that are home to data centers, and is actually broadly useful as a tool for swaths of the population (even if many using it hate it at the same time). In some ways, their disagreement is representative of broader feelings about AI. People love it, hate it, welcome and fear it. And no one knows what its ultimate effects will be.

How SpaceX Streamlined the Raptor Engine, from Construction Physics

You may have seen a photograph floating around on social media, showing versions one, two and three of SpaceX’s Raptor rocket engine. In this Substack post, Construction Physics pieces together clues from Elon Musk’s comments and the research of SpaceX fanatics to figure out how the engine went from looking like something from a university physics lab experiment to the sleek, pared-back device seen in version three. Quite a lot changed, it turns out, though not the fundamentals. Both engines share the same basic design, but SpaceX engineers have stripped away much of what surrounds it. They’ve removed much of the diagnostic hardware needed for a prototype that can be dispensed with once a system is better understood. And by leaning heavily on 3D printing, they’ve transferred much of the exterior piping to inside the engine, which makes it look neater and removes the need for bulky heat shielding. According to Musk, the result is lighter and more powerful — though experiencing some growing pains: The latest engines failed a flight test over the summer.