🎧 Prefer to listen?

Two things happened this week that, taken together, tell you exactly where AI is heading. Anthropic published research showing Claude has a hidden internal reasoning layer — words it thinks but never says. And OpenAI launched ChatGPT Work, their most ambitious product yet — a tool that wants to do your job alongside you. I already wrote a deep dive on what Anthropic found inside Claude (and it’s wild), but today I want to connect both stories. Because the gap between these two moves tells you everything about what’s coming next.

What Anthropic found inside Claude

If you missed my earlier post, here’s the short version: Anthropic built a tool called J-Lens and used it to discover “J-Space” — an internal working memory inside Claude that contains words and concepts the model engages with but never outputs. When they swapped the J-Space representation of “spider” for “ant,” Claude changed its answer about leg count from eight to six. It’s not just autocomplete anymore — there’s structured reasoning happening under the hood.

This matters because it challenges the way most of us think about AI safety. If a model is maintaining internal representations that influence its output but remain invisible, how do you audit that? How do you trust it? I covered the full implications in my deep dive on Claude’s hidden monologue, but the takeaway for solo builders is simple: the models you’re using are more sophisticated than you think, and that sophistication cuts both ways.

OpenAI’s ChatGPT Work: the “super app” arrives

While Anthropic was peering inside Claude’s brain, OpenAI shipped something very different. ChatGPT Work is their answer to the question: “What if AI didn’t just answer questions, but actually did your job?”

Here’s what’s new:

  • Persistent project work. Unlike previous ChatGPT sessions that timed out after a few minutes, ChatGPT Work can “stay with a project for hours.” You give it a goal — analyzing a budget, preparing a sales meeting, building a campaign — and it works through it step by step, asking for approval on important actions.
  • Scheduled Tasks. Think of it as AI-powered cron jobs. You can set up recurring workflows that run even when you’re away, monitored from your phone.
  • Deep integrations. ChatGPT Work connects to Slack, Microsoft Teams, Google Drive, SharePoint, and your desktop files. It also has a built-in browser for web tasks.
  • Codex merger. OpenAI’s coding tool Codex is now built into ChatGPT Work, making it a single environment for both coding and general work.

The Ars Technica review noted that OpenAI is positioning this as a tool that does your work “for you and with you” — a subtle but important distinction from the AI agents we’ve been hearing about. This isn’t an autonomous bot running in the background. It’s a collaborative tool that needs your sign-off.

Why these two stories together matter

Here’s what I find interesting. Anthropic is spending its time understanding how AI thinks. OpenAI is spending its time making AI do more. Neither approach is wrong, but they represent fundamentally different strategies.

Anthropic’s research into J-Space is about transparency and safety. If we’re going to trust AI models with important decisions, we need to understand their internal processes. That’s a long game — it won’t change what you can do with Claude today, but it might determine whether you trust it tomorrow.

OpenAI’s ChatGPT Work is about capability and integration. They’re not waiting for perfect interpretability — they’re shipping tools that connect to your existing workflow and make AI more useful right now. For solo builders and small teams, that’s immediately actionable.

The gap between these approaches is where the opportunity lives. It’s not about Anthropic vs OpenAI anymore — it’s about which tool fits your workflow. Claude for deep reasoning and careful analysis. ChatGPT Work for persistent, integrated task automation. And increasingly, both.

What this means if you’re building with AI

If you’re running a business or building automations, this week’s news gives you three things to think about:

1. The models are getting more capable faster than most people realize. Claude’s J-Space suggests these systems are developing structured internal reasoning. ChatGPT Work suggests they’re ready to take on sustained, multi-step projects. The gap between “AI assistant” and “AI colleague” is closing.

2. Integration is becoming the differentiator. ChatGPT Work’s ability to connect to Slack, Google Drive, and your desktop files is a big deal. The tool that fits into your existing workflow wins, regardless of which model is “smarter.” If you’re still copy-pasting between apps, you’re leaving productivity on the table.

3. Safety and capability need to coexist. Anthropic’s research shows that the models we’re using have internal processes we don’t fully understand. OpenAI is connecting those same models to our work tools and letting them run for hours. Both are necessary. Neither is sufficient alone.

The bottom line

This was a big week for AI, and the headlines don’t capture the full picture. Anthropic is asking “how does this work?” while OpenAI is asking “what can this do?” For solo builders and no-code creators, the answer is: pay attention to both. The tools you use are about to get a lot more capable — and a lot more opaque.

If you want the full breakdown on Claude’s J-Space, check out my earlier post. And if you’re evaluating which AI tools to invest in for your workflow, my tool comparison guide is a good place to start.

New to no-code? Start here.