- Horizon AI
- Posts
- How to Catch the Answer Even Your Best AI Model Gets Wrong 🎯
How to Catch the Answer Even Your Best AI Model Gets Wrong 🎯
Anthropic uncovers Claude's hidden 'J-Space,' offering a window into AI's inner workings 👀
Gina Acosta
July 07, 2026
In partnership with
Welcome to another edition of Horizon AI,
Even the best AI models still get things wrong, and a confident answer isn't always a correct one. Instead of relying on a single chatbot, you can compare responses across multiple leading AI models before using the information.
In today's issue, we'll show you a free tool that automatically checks your prompts across ChatGPT, Claude, Gemini, and other AI models, helping you spot differences, catch mistakes, and get more reliable answers without switching tabs.
Let’s jump into it!
Read Time: 4.5 min
Here's what's new today in the Horizon AI
Anthropic Unveils Claude's Hidden "J-Space" Internal Workspace
Tencent Releases Hy3 Under Apache 2.0 License
AI Tutorial: How to Get More Accurate Answers from ChatGPT, Claude, and Gemini
AI Tools to check out
AI Findings/Resources
The Latest in AI and Tech 💡
AI News
ANTHROPIC
Anthropic Unveils Claude's Hidden "J-Space" Internal Workspace
Anthropic has revealed a newly discovered internal mechanism inside its Claude models, called J-Space, offering researchers a closer look at how AI systems process information beyond the reasoning they show users.
Details:
J-Space is an internal computational workspace that operates separately from the model's visible reasoning, allowing Claude to process concepts that never appear in its final response.
In one demonstration, Claude was asked to copy a sentence while simultaneously thinking about the Golden Gate Bridge. Although the output only repeated the sentence, researchers observed internal activations related to concepts like "bridge" and "California."
The researchers note that the J-Space wasn't designed or programmed by them, but instead "emerged on its own" during Claude's training process.
While the research discusses concepts related to consciousness, Anthropic emphasizes that the findings do not suggest Claude is conscious. Instead, the company frames J-Space as an important step toward opening the AI "black box," with the long-term goal of making advanced AI systems safer, more understandable, and easier to monitor.
TOGETHER WITH LEVANTA
The Free Playbook Behind Millions in Off-Amazon Revenue
Most eCommerce brands running external traffic aren't scaling — they're just spending.
Wrong channels, no real attribution, and at the end of the month, still no clear answer to the only question that matters: what actually moved your BSR?
The brands getting it right aren't necessarily spending more. They've just stopped guessing. They know which channels pull weight on Amazon listings, which ones look good in a dashboard but bleed budget, and why creator traffic consistently outperforms paid social on ROI when it's set up correctly.
Levanta put together a free playbook breaking down 7 proven external traffic strategies. Inside you'll see how top brands are driving millions in off-Amazon revenue and why most channels underdeliver when brands don't know what to look for before they start spending.
If you're serious about growing outside of PPC, this is worth 5 minutes.
TENCENT
Tencent Releases Hy3 Under Apache 2.0 License
Tencent has officially released its new AI model, Hy3, an open-source model the company claims matches the performance of models two to five times its size.
Details:
Hy3 is a 295-billion-parameter mixture-of-experts (MoE) model with 21 billion active parameters, designed to deliver strong performance while requiring significantly less compute than larger frontier models.
Tencent says Hy3 performs particularly well on agentic search, tool use, long-context retrieval, and AI workflows, outperforming many competing open models in these areas.
While Hy3 trails GLM-5.2 on coding benchmarks, Tencent positions it as a more efficient alternative, offering competitive capabilities with less than half the total parameters (295B vs. 744B, with 21B vs. 40B active).
The company also reports major improvements in reliability, including lower hallucination rates, fewer reasoning errors, and stronger multi-turn conversations compared to the preview version released earlier this year.
The company is releasing Hy3 under the permissive Apache 2.0 license on Hugging Face, ModelScope, and GitHub.
AI Tutorial
In partnership with
How to Get More Accurate Answers from ChatGPT, Claude, and Gemini
You attach AI answers to real work every day. On Artificial Analysis's AA-Omniscience accuracy test, the top-ranked model, Claude Fable 5, still lands at just 61%.
Fluent does not mean correct. Cuey checks your prompt across ChatGPT, Claude, Gemini, and more in a single tab, so you catch the misses before they reach your work. Here is the free setup:
Install the free Cuey Chrome extension.
Open ChatGPT, Claude, Gemini, Grok, or your open source model.
Cuey runs quietly in Auto Mode, checking your prompt across top models.
When a stronger answer exists, Cuey notifies you.
Tap the Enhanced Answer to see every model synthesized, plus where they agree, where they diverge, and what yours missed.
No extra subscriptions. No tab switching. Your memory and history stay encrypted on your device, so nothing starts from scratch.
AI Tools to check out
✍ Typeahead: A writing assistant that writes with you, not for you. It works in every text field on your Mac, helping you type faster and smarter with inline suggestions that appear as you write.
✈ AirKaren: AI that fights airline customer service on your behalf.
🌐 Framer: AI website builder for professional sites and teams.
👉 Glaze by Raycast: Create your own Mac apps by chatting with AI.
💡 Humalike: Give your AI agents the social intelligence they're missing.
AI Findings/Resources
👀 Getting started with loops — different types of loops in Claude Code and when to use them
🔍 What is Mistral AI? Everything to know about the OpenAI competitor
The latest in AI and Tech
Elon Musk's AI company has officially rebranded from xAI to SpaceXAI, completing its integration into SpaceX after the companies merged earlier this year. The rebrand includes a new logo and comes after SpaceX's blockbuster IPO in June.
TeraWulf has signed a 20-year lease with Anthropic for a 401 MW AI data center campus at its Justified Data site in Hawesville, Kentucky. The AI infrastructure provider expects to earn $19 billion in revenue from the agreement.
DeepSeek is reportedly developing its own AI chip designed for inference, aiming to reduce its dependence on Nvidia and Huawei hardware. The project is still in its early stages, with the company recruiting chip engineers and working with manufacturing and memory partners despite U.S. export restrictions.
Cohere has launched Cohere Transcribe Arabic, a new open-source speech recognition model designed to handle Arabic dialects, Arabic-English code-switching, and specialized vocabulary.
Anthropic's Claude Cowork AI platform will now be available on mobile and web. The expanded access is rolling out first to Max subscribers and coming to Claude users on other plans "in the coming weeks."
That’s a wrap!
We'd love to hear your thoughts on today's email!Your feedback helps us improve our content |
Not subscribed yet? Sign up here and send it to a colleague or friend!
See you in our next edition!
Gina 👩🏻💻