Claude's Inner Workings: A Solo Builder's Guide

Anthropic says it found Claude’s “internal thoughts.” The internet freaked out. But when MIT Technology Review put a senior editor with a PhD in computer science on the story, his response was basically: slow down. We covered Claude’s “inner life” and OpenAI’s super app when the news first broke, and before that, how Sonnet 5 made agents affordable for solo builders. This week’s MIT Tech Review interview surfaces something the hype cycle buried — and if you’re running a solo business on Claude, it changes how you should evaluate everything your AI tells you.

What the MIT Tech Review editor actually said

The interview is worth reading in full, but here’s the part that matters most. When asked whether it’s fair to use “brain-like” terms when describing how Claude works, Will Douglas Heaven — the senior editor, who holds a PhD in computer science — said plainly: “I don’t love using those kinds of terms. LLMs are not brains.”

This isn’t academic hairsplitting. Anthropic’s own framing — “internal thoughts,” “puzzling over concepts,” “a window into Claude’s mind” — borrows heavily from neuroscience and psychology. That framing builds trust. It makes you think of Claude as a careful, deliberate reasoner with hidden depths. But the research shows something more mechanical.

We covered the J-Space mechanics in detail when the discovery first dropped. The short version: Anthropic’s interpretability team used a technique called the “Jacobian lens” to find a zone inside Claude’s neural network where abstract concepts float around during processing. Things like “the user wants a comparison” or “this fact needs verification.” These concepts shape Claude’s answer, but they never appear in the output.

The MIT Tech Review editor described J-Space as functioning like what neuroscientists call a “global workspace” — a shared processing area where specialized subsystems contribute to decision-making. Claude’s architecture developed this structure on its own during training. It’s not consciousness. It’s something more like a complex but ultimately mechanical process that can be confidently wrong in ways completely invisible to you.

The trust problem nobody’s talking about

Here’s where this gets practical. When Anthropic describes Claude as having “thoughts” it doesn’t show you, it creates a specific mental model — one where the AI is reasoning carefully behind the scenes. But Stanford research on faithful reasoning in LLMs found that large language models often generate plausible-sounding explanations for their answers that don’t actually reflect their internal decision-making process. The explanation sounds right. The internal process might be doing something entirely different.

The spider-ant experiment from Anthropic’s own paper proves the point. Researchers swapped the internal representation of “spider” with “ant” inside J-Space. Claude then confidently stated the creature had six legs. No hesitation. No uncertainty. Just a wrong answer delivered with full conviction.

If you’re using Claude for customer messages, automated workflows, or daily research, that’s the core issue: Claude can be maximally confident and completely wrong, and its explanations for why it’s right might be theater.

World models: the thing that makes this bigger

While J-Space dominated the discourse, a quieter revolution is happening. World models are AI systems that build predictive representations of physical environments — they don’t just process language, they understand how objects move, how spaces work, and how actions create consequences in the real world.

Companies like 1X Technologies are building world models that let robots learn from internet videos instead of requiring millions of physical trial-and-error runs. The MIT Tech Review covered this in the same newsletter edition as the J-Space interview — and the connection matters.

The same underlying architecture — models that build internal representations of environments — powers both J-Space reasoning and world models. The technology that lets Claude silently reason through your prompt is a cousin of the technology that will let AI tools understand physical spaces, navigate real-world contexts, and make decisions about things that aren’t text.

We explored how J-Space and world models converge in depth last week. The practical takeaway: your AI tools are evolving from text processors into systems that build hidden models of reality. That’s powerful. And it means the trust problem gets more complex, not less.

The skeptic’s playbook: what to actually change

The MIT Tech Review editor’s skepticism isn’t about dismissing AI tools. It’s about using them with clear eyes. Here’s what I’d adjust in your solo builder workflow right now:

Stop treating Claude’s first answer as final. For anything that matters — client deliverables, financial analysis, technical specs — run a second prompt asking Claude to challenge its own answer. “What assumptions are you making here?” or “What would make this answer wrong?” forces the model to surface reasoning it might otherwise keep hidden.

Use structured prompts to expose the hidden layer. The MIT Tech Review editor suggested prompting with: “Before you answer, explain the key concepts you’re considering and how they relate.” You won’t see J-Space directly, but you’ll get a more transparent reasoning chain. This is especially useful for complex automations where a silent wrong assumption cascades through multiple steps.

Build verification into your workflows, not just your prompts. If you’re using Claude for multi-step tasks — data analysis, content pipelines, customer responses — don’t just check the final output. Add checkpoints where the AI must present its reasoning before proceeding. ChatGPT Work’s approval checkpoints are a model for this, but you can build the same pattern into any AI workflow.

Compare models on the same task. Run the same prompt through Claude and other tools you trust. When the outputs diverge, that’s your signal to dig deeper. The divergence often reveals where one model’s internal reasoning went sideways.

Keep humans in the loop on anything with real consequences. This sounds obvious, but the J-Space discovery makes it more urgent. Claude’s hidden reasoning can be confidently wrong. For AI agents handling business tasks, design workflows where the AI proposes and the human approves — not the other way around.

The bottom line

The MIT Tech Review editor’s pushback matters more than the hype cycle. Claude’s J-Space is a real discovery, but it’s not consciousness — it’s a complex mechanical process that can fail silently. World models are coming, and they’ll make AI tools more capable and more opaque at the same time. The solo builders who thrive won’t be the ones who trust their AI the most. They’ll be the ones who verify the best.

Want to build AI workflows with verification baked in? Start at /start-here/.