Goodfire Researcher’s 26,000-Word Post Sparks Debate Over Whether Large Language Models Are Far More Interpretable Than Assumed

This edition of The Download arrives with a jolt of optimism—and a wave of debate—after a lengthy new post from Nick Cammarata, a researcher at the AI startup Goodfire, began ricocheting through the AI community. In a blog entry that runs roughly 26,000 words, Cammarata argues that large language models may be substantially more interpretable than many researchers have assumed. If his central claim holds up, it would reshape expectations about whether advanced AI systems can be meaningfully understood, audited, and ultimately governed.

Cammarata’s post has prompted everything from exuberant declarations that “interpretability is solved” to more cautious readings that treat the work as intriguing but far from definitive. In an interview conducted the day after his write-up went viral, Cammarata described what he believes the field has been missing, what his results do—and do not—demonstrate, and why some experts remain unconvinced. The stakes are clear: if the internal machinery of LLMs can be mapped in a reliable way, it could improve monitoring for harmful behaviors and make it easier to build controls that actually work. But skeptics warn that early successes at explaining pieces of model behavior don’t necessarily scale to the messy, emergent capabilities people worry about most.

While that argument plays out, another shift is quietly changing how the web functions. For years, online traffic could be sorted into two buckets: humans and bots. That neat split is breaking down as personal AI agents increasingly perform everyday tasks—booking flights, reserving hotels, and navigating services on users’ behalf. Meta’s Muse, for example, was downloaded 5 million times in its first month, underscoring the speed at which agent-style products can spread.

Yet the internet wasn’t built for a third category of visitor that behaves like a user but moves like automation. As these agents mingle with human traffic, friction has become the norm. Many sites have responded by blocking agents outright, a sign that today’s web infrastructure and business incentives are poorly aligned with automated “assistants” that consume services, scrape content, or interact at scale. If demand for such tools continues to climb, the underlying design of the internet could begin shifting—away from prioritizing human browsing patterns and toward interfaces, authentication, and standards optimized for software agents. That transition would have consequences for everyone: consumers, publishers, e-commerce platforms, and the companies building the agents.

The broader technology landscape is shifting in parallel in climate tech. Moment Energy, named to MIT Technology Review’s list of 10 Climate Tech Companies to Watch for 2026, is betting on a domestic supply-chain shortcut: repurposing used electric-vehicle batteries into grid-connected energy storage systems. Demand for storage is rising sharply, with global capacity projected to increase eightfold over the coming decade. Moment aims to capitalize by turning retired EV packs into new stationary storage—an approach that could ease constraints in a market where manufacturing and supply chains often run heavily through China. Ongoing US–China trade tensions have boosted the company’s prospects, and it expects to soon qualify its repurposed batteries as domestic products in both Canada and the United States, where it is building its next facility.

Meanwhile, the day’s headlines highlight how quickly “agent” technology is spreading beyond consumer convenience. Documents report that the US Army is exploring AI agents designed to identify and disable adversary targets, even as experts caution about unintended impacts on non-target systems and question whether “humans in the loop” can remain meaningfully in control. Separate reporting says a suspect in the alleged OpenAI bugging case had apparent Saudi-government support, including contacts linked to the royal family and a military-funded flight; investigators reportedly found a folder labeled “Saudi Gov Docs,” and a judge denied bail despite an offer from a Saudi consulate to cover costs. The US has also banned DeepSeek from civilian government systems, extending the restriction to products built using DeepSeek’s models. In Europe, major telecom carriers are exploring a joint push for satellite connectivity that could resemble Starlink, as experiments in high-altitude and stratospheric connectivity continue.

Why it matters

If Cammarata is right that LLMs are easier to interpret than assumed, regulators and developers could gain new tools to audit powerful models before they cause harm—but the field will need replication and proof that the approach scales. At the same time, AI agents are forcing a redesign conversation about the web itself, with potential knock-on effects for access, security, and who captures value online. Watch for escalating site-level blocks, emerging “agent-friendly” standards, and policy fights over how automated users should be treated.