Goodfire researcher says AI models may be easier to control

Goodfire researcher argues scientists have been looking at it wrong; the AI community is split between celebration and caution

By LineZotpaper

Published20 October 2026

Updated20 October 2026

Read Time2 min

Nick Cammarata, a researcher at AI startup Goodfire, has published a 26,000-word blog post arguing that large language models may be far easier to understand, monitor and control than previously thought, a claim that has thrown the AI world into debate and drawn both celebration and skepticism.

The post opens with the words "I think I have good news about the nature of neural networks," and argues that researchers have been looking at it wrong. Cammarata contends that LLMs might be far more transparent than the prevailing view suggests, which, if true, would be a huge deal for the monitoring and governance of AI systems.

In an interview with MIT Technology Review conducted the day after the post spread, Cammarata discussed what he thinks researchers are getting wrong, what his work actually shows, and why skeptics are not yet willing to celebrate. While the post has sparked wild speculation and both cautious and unabashed enthusiasm, not everyone is convinced, and the source reports that skeptics are not ready to celebrate just yet.

Agents and the internet

The newsletter also highlights the growing friction between personal AI agents and the human-centred internet. AI agents now handle tasks from booking flights to reserving hotel rooms, with Meta's Muse app downloaded 5 million times in its first month. Many websites are blocking these agents, and if demand continues to rise, the internet could be redesigned for agents rather than people.

A note on other stories in the issue

Other items flagged include the reported US Army plan to use AI agents to hack enemy systems, the US ban on DeepSeek products in civilian government systems, Europe's four major carriers exploring a joint bid for spectrum to build a rival to Starlink, reports that Trump officials asked a federal cyber agency to hack voting machines, and Moment Energy's work repurposing used EV batteries into grid storage, aided by US-China trade tensions.

Analysis

Why This Matters

If true, the claim would make large language models substantially easier to monitor and control, with direct implications for AI safety and regulation.
It challenges a long-held assumption that neural networks are opaque "black boxes" whose behaviour resists explanation.
The reaction demonstrates how quickly a single research claim can shift the AI field's mood and priorities.

Background

Neural networks, including the large language models behind modern AI chatbots, have generally been treated as difficult or impossible to interpret from the inside. Researchers have spent years trying to reverse-engineer why models produce the outputs they do, with limited success. Cammarata's post argues this difficulty may be largely a misconception and that the internal structure of these models could be far more legible than assumed. The claim is notable for coming from a researcher at a startup, Goodfire, rather than an academic lab, and for arriving via a long-form blog post rather than a peer-reviewed paper.

Key Perspectives

Nick Cammarata and Goodfire: The researcher presents the work as good news, arguing that his findings point to models being easier to understand, track and control, which would in turn make managing AI systems more feasible. Skeptics: Reports note that not everyone is convinced and that skeptics are not ready to celebrate. Their concerns centre on whether the findings generalise beyond the specific cases examined and whether the optimism is warranted. The wider AI community: Reactions range from wild speculation to cautious welcome, reflecting both the potential stakes of the claim and the difficulty of verifying it quickly.

What to Watch

Whether the findings are subjected to independent replication and peer review, and how they hold up under scrutiny.
How other interpretability researchers respond in coming weeks, since their acceptance or rejection will carry significant weight.
Whether the claim translates into practical tools for monitoring deployed models, which would test its real-world value.

Sources

The Download: the claim shaking up AI and the internet’s agent problem · MIT Technology Review

Checked

12 claims in this story checked against 1 source: 10 supported.

Show each claim

SupportedGoodfire researcher Nick Cammarata says AI models may be easier to control than previously thought.MIT Technology Review: “Cammarata claims that LLMs might be far easier to understand, and thus to monitor and control, than people previously thought.”
Partly supportedThe AI community is split between celebration and caution over Cammarata’s claim.MIT Technology Review: “sparking wild speculation as well as cautious (and not-so-cautious) celebration. ... But not everyone is convinced.”
SupportedCammarata’s 26,000-word blog post argues that LLMs may be easier to understand, monitor and control than people previously thought.MIT Technology Review: “Nick Cammarata, a researcher at AI startup Goodfire, kicked off a 26,000-word blog post ... Cammarata claims that LLMs might be far easier to understand, and thus to monitor and control, than people previously thought.”
Partly supportedThe post argues that researchers have been looking at the issue incorrectly.MIT Technology Review: ““We were looking at it wrong”: The explosive claim that’s upended the AI world”
SupportedThe post opens with the words “I think I have good news about the nature of neural networks.”MIT Technology Review: ““I think I have good news about the nature of neural networks.” With those simple words of optimism, Nick Cammarata”
SupportedCammarata spoke with MIT Technology Review the day after the post spread.MIT Technology Review: “I talked to Cammarata the day after his post blew up.”
SupportedThe post sparked wild speculation and cautious and unabashed celebration.MIT Technology Review: “sparking wild speculation as well as cautious (and not-so-cautious) celebration.”
SupportedMany websites are blocking AI agents.MIT Technology Review: “The mingling of these agents with human visitors—on an internet designed largely to prioritize people—has gone poorly, with many websites blocking AI agents.”
SupportedMeta’s Muse app was downloaded 5 million times in its first month.MIT Technology Review: “Meta’s Muse alone was downloaded 5 million times in its first month.”
SupportedThe US Army has a reported plan to use AI agents to hack enemy systems.MIT Technology Review: “The US Army is trying to get AI agents to hack its enemies”
SupportedFour major European carriers are exploring a joint spectrum bid to build a rival to Starlink.MIT Technology Review: “Europe is trying to build its own version of Starlink Four major carriers are exploring a joint bid for spectrum.”
SupportedUS-China trade tensions have aided Moment Energy’s work repurposing used EV batteries into grid storage.MIT Technology Review: “Ongoing trade battles between the US and China have been a shot in the arm for the company.”