A MESSAGE FROM BLOOM ENERGY

As AI adoption accelerates, power has become the defining constraint—and opportunity—for data center growth. Our latest 2025 Mid-Year Power Report reveals a dramatic shift in how industry leaders are planning for the future.

Read the report.

Dr. Danielle Perszyk, a cognitive scientist at Amazon’s AGI Labs, can’t help but feel anxious every time she opens her computer, where she is greeted by all the usual pop-up distractions: email alerts, Slack messages, social media updates, “productivity” app notifications. Of course, her frustration—which has to do with these obstructive detours between thinking and interfacing—was hardly evident in the 45 minutes we spent chatting about her research. But it’s certainly one of the factors driving her interest in A.I. “I think that part of the solution is making the tech go more in the background,” she told me, noting the irony of developing new technologies designed to make older technologies less obtrusive.

Amazon’s AGI Labs hasn’t been around for long: It was launched at the end of 2024, after the tech behemoth hired away the top executives at A.I. startup Adept. The team is largely focused on long-term research bets centered around the creation of so-called agentic (action-taking) systems that can help mitigate the negative cognitive impacts of all the bullshit clicking and scrolling that suppress higher-level thinking and creativity. “The technology that we’re building allows all of this existing structure to be masked, essentially, by creating what will ultimately be a meta interface,” she explained, describing a generative user interface “where you’re starting from the goal that the human has, and the model understands what that goal is, and it will put everything else on silent and only give you the button that you need to click in this moment.” In other words: an interactive focusing tool that would just work.

To achieve this lofty vision, the lab is attempting to build something like an artificial general intelligence, which is among the more controversial end goals of the A.I. race. The idea—some would say fantasy—is to build models that actually think like humans, and perhaps eventually replace all human labor. Earlier this year, there was excitement across Silicon Valley that we might be at the precipice of A.G.I.; more recently, there’s been growing acknowledgement that large language models alone are likely not up to the task. Sam Altman, for instance, has backed away from his earlier suggestions that GPT-5 would represent a meaningful step toward A.G.I., and now dismisses A.G.I. as an unrealistic near-term target.

Perszyk’s lab, however, is defining A.G.I. a bit differently—as a system that can do everything a human can do, but on a computer. Their goal is not to create models that can replace human labor. Instead, it’s to build systems that understand how we think, and can therefore enhance our capabilities through “useful agents.” In March, the team unveiled Nova Act, a research preview for a model designed to perform actions within a web browser—a stepping stone toward the “meta interface” she described to me. The system, according to Amazon, achieved 90 percent reliability rates across specific enterprise use cases. Months earlier, Anthropic and OpenAI launched iterations of this same concept in the form of Computer Use and Operator, respectively, but so far, the trouble with those “agents” has been reliability: Like all L.L.M.s, they’re error-prone, and when you’re dealing with actions, rather than words, the impact of those mistakes can be a lot worse. But Perszyk believes her lab has struck upon an approach that is much more commercially viable.

Obviously, though, there’s a long road ahead. As Perszyk explained, the “ingredients that go into the breakthroughs for L.L.M.s are categorically different from agents. It’s not even starting with the foundation of the L.L.M.s that we all take for granted. It’s a whole different beast.” In short, since agentic products require the model to understand actions as well as words, those systems must be trained on “a much more diverse range of skills and human interactions,” Perszyk said, describing her lab’s approach as the only path to generalization—the G in A.G.I.

Getting there, however, represents a significant challenge, since the necessary decision-making is social in nature, and further, requires the model to understand what individual humans actually care about. For Amazon, L.L.M.s will inevitably play a role in that effort, but they’re just one piece of the puzzle. “A lot of human knowledge is contained within the L.L.M.s, but it’s kind of inert,” Perszyk told me. “It doesn’t have a way to reason about that information, and the methods for getting the L.L.M.s to better reason still require using domains that have correct or incorrect answers.” And human reasoning, as she explained, is just a lot more complicated than yes or no.

A MESSAGE FROM BLOOM ENERGY

As AI adoption accelerates, power has become the defining constraint—and opportunity—for data center growth. Our latest 2025 Mid-Year Power Report reveals a dramatic shift in how industry leaders are planning for the future.

Read the report.

According to Perszyk, given the limitations of L.L.M. datasets, the only path to achieving useful, reliable, and trustworthy agents is through the construction of a “digital world model.” Amazon’s plan to develop that world model involves training its agents, through lots of reinforcement learning, to interact with fake websites, which essentially act as a simulation of the digital world. Part of that effort has pinned its hopes on the nebulous idea of “emergence”—that the systems will just sort of get better on their own—but the other part involves “playing around with a lot of different architectures and then making them interact.” It’s not clear what those different architectures might be.

In Perszyk’s view, the main limitation is still reliability, which she divides into three parts: consistent accuracy, which she said is lacking in current models; the ability to self-assess when human intervention is needed; and the ability to understand an individual human’s perspective, which involves tackling the ancient philosophical problem of other minds, through an artificial version of collective intentionality. So, to get there, Amazon needs its agents to have digital world models, social models, and mental models. “We need it to deeply understand how our minds work,” she said, noting that solving these three problems is critical for building user trust, and might one day enable the lab to develop systems capable of generalization. That’s the idea, anyway.

Of course, this is Amazon, which would love to see commercial-ready applications of A.I. agents being produced ASAP by its AGI Lab, the establishment of which came amid pretty extensive job cuts at the company. C.E.O. Andy Jassy recently predicted that the company-wide adoption of A.I. would likely reduce their workforce over time. But at least on the A.G.I. side of things, it’s still early days; Perszyk’s team is partnering with enterprise customers to identify and develop highly specific, product-ready workflows to which their A.I. agents could be applied.

But Perszyk’s work raises all kinds of philosophical questions, too. In the months and years since the rise of ChatGPT, early-stage studies have consistently noted that generative A.I. could result in both skill and cognitive atrophy as users offload basic tasks to bots. Yes, skill atrophy has been associated with nearly every new technology throughout history. Hardly any drivers in the U.S. can still read maps, for instance. But what happens if Americans… can’t read period?

Perszyk believes we’ll have to think carefully about which skills are valuable to us—and which aren’t—so that we can walk the fine line between augmentation and atrophy. “The majority of tools are kind of artifacts from trying to solve problems in the workplace that we have to become expert in,” she said. “But they’re what I call arbitrary skills, and I would be more than happy if my brain atrophied with respect to all of the arbitrary skills that I’ve learned. Let’s automate those.”

At the same time, she did note that some of those “arbitrary” skills might well provide the scaffolding necessary to understand more-essential concepts. Even with the current generation of chatbots, this process already seems to be underway. “If you’re just asking the chatbot to generate a first draft for you, you are literally not doing the type of thinking that allows you to put ideas together in a logically coherent way,” Perszyk said. “Learning how to read and write is a forcing function for you to notice inconsistencies and incoherences in your argument. People are losing that. That is absolutely something that we should call out.”

Even so, we’d probably all be happier with bots answering questions for us on Slack, or responding to emails. (If you’re using Gemini in Gmail, you’re already halfway there.) In the end, though, the trouble comes down to who actually gets to decide which tasks are worthy of automation, assuming a reliable-enough model becomes ubiquitous. What will happen if the tasks vital for cognition are gradually replaced, and perhaps vanish altogether?

A recent Reuters investigation unearthed an internal Meta document that asserted that it’s acceptable for its chatbots to engage in sensual, romantic role-playing with child users. The detail emerged from a story about a cognitively impaired New Jersey man who died while pursuing a romance with a chatbot. [Reuters]

Police officers around the country are using generative A.I. tools to write their reports… and they’ve been toggling off a feature that would signify that those reports were generated by A.I. [Mother Jones]

As L.L.M.-based coding assistants become increasingly popular, the security nightmare is just beginning. The scope is a lot broader than you might think. [Nathan Hamiel]

That’s all for today. I’ll see you on Thursday.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St, New York, NY 10006