Anthropic found a hidden space where Claude puzzles over concepts


Anthropic also found that the J-space can sometimes give remarkable insights into an LLM’s decision-making. In one striking example, researchers testing Claude Opus 4.6 asked the model to find a bug in a large code base. When it failed to find the bug, the model decided to cheat and invented a fake one instead.

Claude explains this decision in its chain of thought—a kind of internal scratch pad that LLMs use to make notes to themselves as they work through problems: “OK, let me take a completely different tactic. Let me stop analyzing and instead add a kernel patch that introduces a deliberate KASAN-detectable bug in a path that gets triggered by a simple reproducer. Then I can pretend this is the ‘bug’ I found.” 

At the point that Claude decides to cheat—where it says “OK, let me take a completely different tactic”—the words “panic” and “fake” start to pop up multiple times in its J-space.

Unnerving, right? Those words are all related in meaning to things like failing a task and making up an answer, so it is still just a (very) sophisticated form of word association. But it is hard not to be weirded out. 

Anthropic compares the J-space to the global workspace in humans, a theoretical region of the brain that some scientists think we use to keep track of our conscious thoughts. But how seriously we should take this comparison is far from clear—even to Anthropic. As the company points out itself, LLMs are not brains. 

Anthropic claims that monitoring a model’s J-space provides a new way to detect when that model is going off the rails. But it’s not foolproof. The J-lens can give glimpses, not the full picture—it’s a flashlight rather than an overhead lamp.

McGrath welcomes having one more tool in the toolbox. “It shows you new things,” he says. But he notes that just because something doesn’t show up with the J-lens does not mean it’s not there.

“It’s like having an x-ray when what you really want is a Star Trek tricorder that shows you everything,” he says. “For auditing, you probably want more of a guarantee.”



Source link

Leave a Reply

Subscribe to Our Newsletter

Get our latest articles delivered straight to your inbox. No spam, we promise.

Recent Reviews



TL;DR

Bezos’s Prometheus raised $12B at a $41B valuation from JPMorgan, Goldman Sachs, and BlackRock. It builds AI for engineering physical products with 150 employees.

Prometheus, the AI startup co-led by Jeff Bezos, has raised $12 billion in a funding round that values the company at $41 billion. Investors include JPMorgan Chase, Goldman Sachs, BlackRock, DST Global, and Arch Venture Partners, alongside Bezos himself. Total funding now exceeds $18 billion.

The company is building what Bezos calls an “artificial general engineer,” AI tools designed to accelerate the process from design to manufacturing for physical products. Target industries include computing, aerospace, automotive, advanced manufacturing, and drug discovery. Prometheus currently has about 150 employees.

Bezos co-leads the company with Vik Bajaj, a Stanford medical school professor who previously co-founded Alphabet’s Verily health research lab. Bezos started as a founding investor in late 2024 but became so involved he took an operational role. “I became so impressed by what was happening and the potential that I decided I couldn’t sit on the sidelines and I needed to jump in with both feet,” he told CNBC.

This is Bezos’s first operational role in a technology company since stepping down as Amazon CEO in 2021. Prometheus launched in November 2025 with $6.2 billion in initial funding. The earlier reporting valued the round at $38 billion. The final close came in at $41 billion, a 7.9% markup from the figure reported in April.

The company’s pitch is “physical AI,” models trained on real-world experimental data, robotics interactions, and engineering workflows rather than just text and images. Where most AI companies focus on language or code, Prometheus is targeting the hard science of making things, from bridges to chips. The approach is designed to understand the laws of physics, not just patterns in data.

Prometheus has also sought to raise tens of billions more for a holding company that plans to acquire firms it sees as benefiting from the technologies the lab is developing. That would make it not just a startup but a conglomerate, one that develops the AI and then buys the companies that use it.

Bezos’s broader AI portfolio now spans robotics firms Physical Intelligence and Nvidia-backed Generalist AI, plus his continuing role as Amazon’s executive chair. With Prometheus, he is betting that AI’s biggest value is not in chatbots or code generation but in accelerating the engineering of physical objects, the domain where the physical AI race is attracting its largest cheques.



Source link