Welcome back to AI Researches! 👋

This week, Anthropic revealed something unexpected inside Claude: a hidden reasoning space called J-space, where the model appears to hold ideas before producing an answer. It's a step toward understanding how AI thinks internally and could help researchers make future models more transparent and reliable.

We've also rounded up the biggest AI developments from the past week, practical AI guides, useful tools, new research, expert predictions, and other updates worth knowing.

Let's dive in. 🚀

In today’s AI researches:

• Anthropic Found Claude’s Hidden AI Thinking Space

• Meta says its next AI model has already reached GPT-5.5-level performance.

• Sam Altman is calling for a U.S.-led AI safety forum

• Learn How to create on-brand ads with Claude and GooseWorks

• AI Researches case studies

• AI tutorial

• wild predictions from Google CEO, Naval Ravikant

• Best performing Ai tools of the week

• More Ai news and developments

Read time: 5 minutes

 NEW 

🧠 Anthropic Found Claude’s Hidden AI Thinking Space

AI Researches: Anthropic says it found a hidden reasoning area inside Claude, called J-space, where the AI appears to hold ideas it does not directly say out loud.

The Details

• NResearchers found J-space inside Claude Sonnet 4.5 and Opus 4.6, where the model keeps small internal clues while solving harder tasks.

• This hidden space can hold things like middle math steps, bug clues, fake content signals, or prompt tricks before Claude gives the final answer.

• Anthropic used a new tool called Jacobian Lens to look at these hidden patterns and see which ideas Claude had “on its mind.”

• When Claude could not use this J-space well, it still handled simple chats, but struggled more with multi-step reasoning and harder problem-solving.

why this is important

This gives researchers a better way to peek inside AI reasoning, which could help catch hidden mistakes, risky behavior, or dishonest answers before they reach users.

Keep up with tech in 5 minutes

TLDR is the free daily email with summaries of the most interesting stories in startups, tech, and programming. The stuff worth knowing, minus the doomscrolling.

Issues are curated by ex-Google and Anthropic engineers and land in your inbox before your morning coffee. A 5-minute read, and you walk into the day already knowing what your team is still catching up on.

Tech is just the start. We also cover AI, marketing, dev, and more. Pick the briefs that match your work.

Free, daily, and read by 7M+ subscribers. Subscribe and let the experts do the digging for the tech news that matters.

Subscribe for Free

 iMajor AI News that Happened in this week i

Meta says its next AI model has already reached GPT-5.5-level performance.
Alexandr Wang revealed that Watermelon is still training, uses 10× more compute than Muse Spark, and is expected to power Meta's next major AI upgrade.

Sam Altman is calling for a U.S.-led AI safety forum. He says governments should set the rules for frontier models, while OpenAI has reportedly discussed giving the U.S. a 5% stake.

Fable 5 is back online, but under stricter rules. Anthropic restored access after the Commerce Department lifted restrictions, adding stronger cybersecurity filters and fallback responses from Opus 4.8.

Google added cheaper image generation and stronger video editing to its AI API.
Nano Banana 2 Lite costs just 3.4 cents per image, while Gemini Omni Flash creates and edits 10-second videos at 10 cents per second.

🎯 Create on-brand ads with Claude and GooseWorks

Learn how to use Claude with GooseWorks to scan your website, extract brand details, choose ad templates, estimate credits, and generate ready-to-test ad creatives.

1. Install GooseWorks
Open Terminal and run npx gooseworks install --all, then complete the login and approval steps.

2. Start Claude Code
Type claude, press Return, then run: /goose-ads create ads for my brand [brand name] [website link].

3. Review the brand profile
Open the Goose Ads project link and check the logo, colors, product details, messaging, and brand voice.

4. Approve templates
Prompt Claude: “Pick 3 Goose Ads templates, estimate credit usage, and wait for approval.”

5. Generate and refine
Approve the batch, review the ad outputs, then improve the strongest concepts before using them.

Tutorial link: Learn from the official GooseWorks Creating Ads guide, which explains templates, generation settings, remix modes, and credit usage.

Researchers from Kyutai and General Intuition released MIRA, an open-source AI that generates a fully playable 2v2 Rocket League match in real time without using a traditional game engine. Trained on 10,000 hours of AI gameplay, it creates the entire game inside a neural network, showing how AI world models could eventually simulate real-world environments for robotics and training.

A new Remote Labor Index tested AI models on real freelance jobs like design, animation, and architecture, with humans grading the results against professional work. Anthropic’s Fable 5 achieved the best score, matching or beating professionals on 16.1% of tasks, showing that AI is improving quickly but still works best as a tool that helps freelancers rather than replacing them.

🔊 Create Custom Sound Effects with AI in Seconds

1. Open Sound Effects
Sign in to ElevenLabs for free and select Sound Effects from the left sidebar.

2. Generate your sound
Describe the sound effect you want in the prompt box, then press Enter or click Generate. In a few seconds, you'll get four different variations.

3. Download and use
Pick your favorite version and download it for videos, podcasts, games, presentations, or any other creative project.

🔗 Learn more→

Speaking on his podcast, Naval Ravikant said some of the people building frontier AI believe the technology could eventually solve problems across science, engineering, and medicine while leaving far fewer roles for humans. He illustrated the idea with a provocative joke about Anthropic becoming one of the only employers left.

Sundar Pichai said the AI boom is not a simple bubble, because real demand exists, but he admitted the industry is also overinvesting in infrastructure and some companies may struggle to justify those costs later.

• 📐 MySpec — Interviews you like a senior architect, then turns your idea into a structured spec bundle for coding agents.

• 📊 CorpusIQ — Connects ChatGPT, Claude, and Perplexity to live business data, returning source-linked answers from your tools.

• 🔗 Lynx AI — Maps your niche, studies competitor backlinks, and places your links across 2,000+ sites with 90-day monitoring.

• 🎬 Sea Imagine — Turns text or images into AI videos across multiple models, with motion and camera transitions in one interface.

• 🐣 Hatchable — Builds and hosts real apps from ChatGPT, Claude, or Cursor through an MCP connector.

• 🎨 Axey — Lets you work in an AI canvas with multiple files, chats, custom specialists, and your choice of AI models side by side.

• 📱 NiuNiu — Turns a plain-language app idea into an installable Android APK, creating a build plan and credit estimate before generating the app.

xAI has officially rebranded as SpaceXAI. The new identity brings Grok, Colossus supercomputers, and SpaceX’s AI satellite plans under one brand as Musk folds AI deeper into SpaceX.

Lenovo launched a $44 AI phone for students in China. The device keeps only the basics, like calls, GPS tracking, classroom mode, payments, and an AI button for homework help.

Midjourney is asking the court to make Disney, Universal, and Warner Bros. reveal how they use AI internally, arguing that the same studios accusing Midjourney of training on unlicensed content may also be using copyrighted material to train or test their own AI systems.

Anthropic launched Sonnet 5, its most capable mid-tier AI model yet.
The upgrade brings stronger coding, reasoning, and agentic abilities, arriving just days before the return of the more powerful Fable 5 and Mythos 5.

Claude works best when the system prompts itself. An Anthropic engineer shows how to build a self-improving agent that writes, tests, and improves its own prompts instead of relying on one perfect manual prompt.

Google is turning NotebookLM into an educational doomscrolling app. Short Video Overviews turn complex sources into 60-second vertical explainers, rolling out first to AI Ultra and Pro users.

SpaceX’s Colossus data centers are becoming an AI moneymaker.
The company signed a $6.3B compute deal with Reflection AI, adding another tenant alongside Anthropic and Google.

Meta is reportedly preparing to rent out extra AI compute. The plan could let outside companies pay for Meta’s spare data center capacity or use models hosted on Meta’s infrastructure.