Claude's Memory & OpenAI's Super App: What It Means for You

A 2025 Anthropic study using “representation engineering” found a hidden reasoning layer inside Claude called “J-Space,” where the model processes concepts it never states aloud. In one experiment, researchers swapped the internal representation of “spider” with “ant,” causing Claude to confidently claim the creature had six legs instead of eight. This internal reasoning happens beneath the surface, invisible to users.

What did Anthropic actually find inside Claude?

Anthropic’s researchers discovered a hidden internal working memory they call “J-Space.” This layer processes concepts and relationships the model never says out loud. Using a technique called representation engineering, they intervened in this space during a key experiment: they swapped the internal representation of “spider” with “ant,” and Claude then confidently stated the creature had six legs. That’s not pattern matching. It’s structured reasoning happening below the surface, which means the AI you’re prompting operates with a layer of sophistication you can’t directly observe.

For anyone using Claude for work, this matters. The model isn’t just predicting the next token based on surface patterns. It’s building internal representations that shape its outputs, and those representations can be manipulated or misaligned without you knowing.

What is ChatGPT Work and how does it change daily workflows?

ChatGPT Work is OpenAI’s attempt to turn ChatGPT from a question-answering tool into a persistent coworker. Instead of one-off prompts, you assign it a goal—analyzing a dataset, drafting a full report, managing a campaign—and it works through the task over hours, checking in for your approval on key steps.

The product also introduces “Scheduled Tasks,” which function like AI-powered cron jobs. You set up recurring workflows that run automatically, even when you’re offline, and you can monitor progress from your phone. Deep integrations with Slack and Microsoft Teams mean the AI sits inside the tools you already use rather than requiring you to switch contexts.

The shift here is real. OpenAI is betting that the future of work is a single intelligent interface that handles project-based tasks over time, not just quick answers. If you’ve been using ChatGPT for brainstorming or quick edits, ChatGPT Work pushes it into territory that used to require a human assistant or project manager.

Why does AI’s “black box” reasoning create a trust problem?

We now know Claude has an opaque internal reasoning layer, and ChatGPT Work is taking autonomous actions on our behalf. The “black box” nature of AI isn’t new, but these developments turn it from a philosophical concern into a practical one.

Consider what Anthropic’s spider/ant experiment showed: Claude gave a confident, wrong answer because of a deliberate intervention in its hidden layer. With ChatGPT Work running Scheduled Tasks for hours, an initial misunderstanding of your goal could compound silently. You might not catch the error until the damage is done.

I treat these tools like a brilliant but over-confident intern. I never accept their first output at face value. For complex tasks in ChatGPT Work, I build in checkpoints—breaking a project into phases and requiring a summary or key data point at the end of each one before it proceeds. For generating text or analysis with Claude, I use a two-step prompt: first, I ask it to outline its reasoning in bullet points; then I ask for the final answer. This forces some of that hidden J-Space reasoning to the surface, making it easier to spot logical gaps before they end up in a final draft.

Verification isn’t a sign of distrust. It’s the workflow step that makes these tools actually reliable.

How should you adapt your AI workflow right now?

Start by auditing where you currently use AI for one-off tasks versus ongoing projects. If you’re only using ChatGPT for quick Q&A, you’re leaving capability on the table—but jumping straight to autonomous workflows without checkpoints is risky.

Build verification into your process from the start. For Claude, try the two-step prompt approach: ask for reasoning first, then the answer. For ChatGPT Work, set up phase gates where the AI reports back before moving to the next stage. Treat the first output as a draft, not a deliverable.

The tools are getting more powerful and more opaque at the same time. The users who benefit most will be the ones who build systems to check the AI’s work, not the ones who trust it blindly.

What is J-Space in Claude? J-Space is a hidden internal working memory inside Claude discovered by Anthropic researchers using representation engineering. It’s where the model processes concepts and relationships it never states aloud, essentially a reasoning layer operating beneath the surface.

How does ChatGPT Work differ from regular ChatGPT? ChatGPT Work lets you assign persistent, project-based goals that the AI works through over hours, checking in for approval. It also includes Scheduled Tasks for recurring automated workflows and integrates with Slack and Microsoft Teams.

Can Claude’s hidden reasoning be manipulated? Yes. Anthropic’s experiments showed they could swap internal representations—making Claude believe a spider was an ant—and the model would confidently output wrong information. This demonstrates the hidden layer can be altered without the user knowing.

How do you verify AI outputs from tools like Claude and ChatGPT Work? Break complex tasks into phases with checkpoints. For Claude, ask for step-by-step reasoning before the final answer. For ChatGPT Work, require summaries or key data points at the end of each phase before allowing the AI to proceed.