Read Online | Sign Up | Advertise |
|
|
|
Good morning, AI enthusiasts. Anthropic has been criticized for its AI consciousness talk — and while its latest research is careful to avoid any conclusions, it definitely has neuroscientists talking.
The researchers say a newly found “J-space” holds Claude’s unspoken thoughts and “emerged on its own” during training — a structure that mirrors leading theories of how the human brain handles conscious thought. |
|
|
In today’s AI rundown: |
Anthropic finds a hidden workspace inside Claude Tencent open-sources small, powerful Hy3 Build any mobile app idea in 15 minutes with Replit Kyutai’s AI dreams up a playable Rocket League 4 new AI tools, community workflows, and more
|
|
|
|
|
|
|
ANTHROPIC |
|
 |
Image source: Anthropic |
|
The Rundown: Anthropic just published research showing that Claude does its thinking in a small, internal “workspace” resembling what neuroscientists credit for human conscious access, with researchers saying the area wasn’t programmed but “emerged on its own.” |
The details: |
Anthropic found ‘J-space,’ an internal area acting like Claude’s internal notepad, which holds active concepts the model uses and steers while thinking. J-Space is different from chain-of-thought text, with these internal hints staying off-screen even when they help direct Claude’s next answer. Editing the unspoken thoughts changes answers: swapping an internal “spider” pattern for “ant” during a question about legs flipped its response from 8 to 6. When J-space was deleted, Claude still chatted and recalled facts, but multi-step problem completion collapsed, showing where harder thinking happens.
|
Why it matters: Anthropic has been criticized (most vocally by Microsoft AI head Mustafa Suleyman) for its AI consciousness talk, and while the researchers note this doesn’t reveal “whether Claude is conscious… or feels anything at all,” finding an undesigned, brain-like workspace inside a model is exactly why the lab keeps asking. |
|
|
|
TOGETHER WITH SLACK FROM SALESFORCE |
|
|
|
The Rundown: Slack is the coordination layer where AI agents access conversational history, workflow signals, and business context — so they expand capacity instead of creating risk. New IDC research shows 66% of developers have already moved AI agents into production. |
In the white paper, you’ll discover: |
Why context determines if AI agents succeed or fail How Slack reduces coordination overhead for developers Why Slack is built to orchestrate multi-agent work
|
Download the free IDC white paper. |
|
|
|
TENCENT |
|
 |
Image source: Tencent |
|
The Rundown: Chinese tech giant Tencent’s Hunyuan just promoted its Hy3 model out of an April preview and into a full open-source release, with benchmark results that claim to rival “flagship open-source models with 2-5x parameters.” |
The details: |
Hy3 only uses a small slice of its parameters to work on each request, letting it run on under half the hardware needed for Zhipu’s much larger GLM-5.2. GLM-5.2 still beats Hy3 across coding tests like SWE-bench, though Hy3 tops open rivals on web research and tool use in Tencent’s own testing. The release features a permissive Apache 2.0 license, avoiding regional bans of earlier Chinese models that left those in the EU, UK, or S. Korea locked out.
|
Why it matters: Dario Amodei recently said that Chinese models are just 6–12 months behind the frontier, and Hy3 doesn’t appear to challenge the top models or that timeline. But Tencent does make some major efficiency gains that, combined with its extremely open license, make for a very compelling release. |
|
|
|
AI TRAINING |
|
|
|
The Rundown: Learn how to turn an app idea into a Replit prototype to test on your phone. Define one user flow, give Replit Agent a focused prompt, preview with Expo Go, then improve the weakest part. |
Step-by-step: |
Install the desktop app from replit.com/desktop, and download the Expo Go app to your phone (for app testing). Before building, define the smallest version that proves the idea. Here is the example prompt from the video: “Help me write a PRD for a playlist-themed Pomodoro app. I dictate tasks, AI turns them into Pomodoro sessions, and I can skip tasks like songs. Include design requirements, core features, and five simple names. Do not choose the tech stack.” Copy the PRD into Replit, choose the mobile app option, and ask Agent to build the first version. Keep it focused on the main flow, not login or payments. When the app is ready, open preview, scan the QR code, and run it through Expo Go. Tap the main screens and note what feels unfinished. Ask Replit for one targeted improvement, such as clearer confirmation copy or stronger trust signals. Preview again before adding features.
|
Going further: Reuse the same prompt structure for other app ideas, then add storage, auth, or App Store steps only after the first loop feels right. |
|
|
|
PRESENTED BY GLEAN |
|
|
|
The Rundown: 87% use AI at work. Only 13% see a significant impact. Guest Lectures at Glean Academy returns for a third season, turning the Work AI Index 2026 into a practical, three-part series for leaders who want to close the performance gap. Hear AI researchers and experts at MasterCard, Target, Emory University, and more. |
In the series, you’ll learn: |
How to avoid the hidden labor that AI creates How the best teams design effective human + AI routines What high-performance AI leaders are doing to set themselves apart
|
Enroll now to save your spot and win one of 10 free Glean:GO 2026 tickets. |
|
|
|
AI RESEARCH |
|
 |
Image source: Kyutai and General Intuition |
|
The Rundown: Kyutai and General Intuition just released MIRA, an open-source world model built with Epic Games that runs live 2v2 Rocket League for four players — generated entirely inside a neural network with no game engine underneath. |
The details: |
MIRA picked up the game from 10K hours of footage of AI bots playing each other, with zero human gameplay or player data anywhere in the training. Despite not having a physics engine or graphics code, MIRA renders details like boost meters and crashes, while keeping all screens in sync on a single GPU. Its memory only spans about four seconds, so when a goal replay rolls, MIRA hallucinates convincing footage of a goal that never actually happened. The game runs at 20 frames per second on a single Nvidia GPU, with the teams open-sourcing the code, training data, and a playable demo.
|
Why it matters: We recently covered Odyssey’s Agora-1 world model generating a multiplayer version of Goldeneye, and now we have another model handling Rocket League. But as the companies note, the point isn’t to replace video games, but to simulate the physical world for a future full of robots that need data to train on. |
|
|
|
|
|
|
|
🚀 Hy3 - Tencent’s new open-source model for cheap, reliable agent deployment 🍌 Nano Banana 2 Lite - Google's high-volume image model that generates pictures in just four seconds 🔈 Seed Audio 1.0 - ByteDance's model for generating speech, music, and SFX in one pass ⚙️ ZCode- Z AI’s agentic coding environment tuned for GLM-5.2
|
|
|
|
|
Elon Musk’s xAI officially rebranded to SpaceXAI, following the company’s February merger that valued the AI lab at $250B. |
Illinois Gov. JB Pritzker signed the first U.S. state law forcing big AI developers to undergo annual third-party safety audits and reporting, with backing from both Anthropic and OpenAI. |
OpenAI’s Tibo Sottiaux revealed that the company’s coming GPT 5.6 ‘Ultra’ variant will be available in Codex, with the model family expected to be released this week. |
ByteDance and Alibaba are ending AI companion features in Doubao and Qwen, coming ahead of a July Beijing rule that targets chatbots and emotional dependency. |
Sotheby’s is auctioning off the “Jensen Jacket,” an autographed leather jacket previously worn by Nvidia CEO Jensen Huang, with an estimated final bid of $40–60K. |
|
|
|
|
|
|
|
Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier. |
Today’s workflow comes from reader Anonymous: |
“I've been studying Spanish and am always looking for ways to help me remember new vocabulary and phrases. One fun method I've been using lately: I use MacWhisper to make a transcript of a Pimsleur language lesson from the MP3 file, drop that into Claude with instructions to craft the lyrics to a catchy Latin pop song using only the vocabulary in the transcript, and then paste those lyrics into Suno to make the song. |
I now have a growing playlist of catchy Spanish-language songs that remind me of past lessons, help me with pronunciation, and prepare me for learning even more.” |
How do you use AI? Tell us here. |
|
|
|
|
|
|
That's it for today!Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you. |
|
|
See you soon, |
Rowan, Zach, Shubham, and Jennifer — the humans behind The Rundown |
|
|
|
|
|
|
