Claude’s hidden thinking space

PLUS: Meta’s Watermelon catches GPT-5.5, and the AI jobs story just flipped

Author

Gen AI Art
July 08, 2026 • Estimated Reading Time: 3 minutes

In partnership with

Anthropic cracked open Claude and found something nobody built: an internal workspace where the model holds and works through ideas before it says a word.

Researchers can now watch concepts move through that space, and even swap them mid-thought. It raises a fun question: what else is sitting in there that nobody has looked at yet?

Today in AI:

  • Anthropic finds Claude’s hidden thinking space

  • Meta says Watermelon has caught GPT-5.5

  • The AI jobs story just flipped

Cut Lead Review From Hours To Minutes

Sign up for a free trial of Attio, the agentic CRM.

Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.

Review your pipeline in Claude, synced live from Attio via MCP.

That's it.

What’s new? Anthropic researchers discovered an internal workspace inside Claude where the model holds concepts and reasons through them before producing any output. They call it J-space, and it behaves a lot like the “global workspace” theory of how human thinking works.

What matters?

  • Using a technique called the Jacobian lens, researchers can watch ideas sitting in the workspace. In one test, swapping “spider” for “ant” inside J-space changed Claude’s answer from 8 legs to 6.

  • When researchers suppressed the workspace, complex reasoning fell apart while the writing stayed fluent. The model kept talking. It just stopped thinking.

  • You can try this yourself on Neuronpedia, which lets you send a message to an open model and watch what pops up in its workspace.

Why it matters?

Safety teams finally get a way to inspect what a model is actually considering before it answers, including things it noticed but never said, like a prompt injection. It also changes the mental model for the rest of us: these systems hold organized ideas, not just probable next words.

GUIDE

What’s new? Meta’s superintelligence chief Alexandr Wang told employees that Watermelon, the company’s still-in-training flagship model, now matches OpenAI’s GPT-5.5 on closely watched benchmarks.

What matters?

  • Watermelon is trained on roughly 10x the compute of its predecessor Muse Spark, and Meta is pouring up to $145 billion into AI infrastructure this year.

  • Wang also teased a Muse Spark update focused on coding, aimed squarely at closing the gap with Claude Opus.

  • One caveat worth keeping: the benchmarks are unnamed and unverified, so this is an internal claim until the model actually ships.

Why it matters?

Meta got written off after Muse Spark’s lukewarm reception, and a genuine GPT-5.5-class model would put it right back in the frontier race. More competition at the top has a habit of making tools cheaper and features ship faster for everyone.

SPONSORED BY LINDY.AI

“Who is this person again?”

You’ve had that moment. Walking into a call, scrolling through old emails, trying to remember what you promised. Lindy texts you a brief 15 minutes before: attendee context, past discussions, open items, talking points. All pulled automatically. Try Lindy free.

What’s new? New data from Ramp shows companies that adopted AI aggressively grew their white-collar headcount 10.2% after adoption. The jobs didn’t vanish. They moved.

What matters?

  • 32% of managers say they’re now rehiring for roles they had previously cut in favor of AI.

  • Freelancers are getting paid to fix “AI slop”, cleaning up rushed AI content that businesses shipped too fast.

  • AI labs are actively recruiting philosophy majors for alignment and reasoning work. Unusual backgrounds are suddenly assets.

Why it matters?

The simple “AI replaces workers” story isn’t holding up. The market is paying for people who can direct AI and repair its output, which is exactly the skill set yesterday’s issue was about.

Everything else in AI

Meta released Pocket, a free app for making and sharing AI-generated minigames. Describe a game and it builds something playable.

Google is training its AI on your saved images, files, and audio unless you opt out. Two settings changes turn it off.

Tencent open-sourced Hy3, a model with a 262K-token context window that you can run and fine-tune yourself.

Jensen Huang’s famous leather jacket is up for auction at Sotheby’s, photo-matched and authenticated, with proceeds going to charity.

Essential AI Guides - Reading List:

Let us know!

Work with us

Reach 100k+ engaged Tech Professionals, Engineers, Managers and decision makers. Join brands like MorningBrew, HubSpot, Prezi, Nike, Ahref, Roku, 1440, Superhuman, and others in showcasing your product to our audience. Get in touch now →