The part of Claude's brain nobody built
Anthropic researchers discovered 'J-space,' an internal workspace in Claude that wasn't explicitly programmed but emerged during training, functioning similarly to human conscious thought. The newsletter also covers Tencent's efficient Hy3 model, Kyutai's Rocket League world model, and practical AI applications including mobile app development with Replit.
Summary
Anthropic published research revealing a hidden internal structure called 'J-space' within Claude that acts as an unspoken workspace holding active concepts the model uses while thinking. This structure emerged spontaneously during training rather than being explicitly designed, and mirrors neuroscientific theories about how the human brain handles conscious access. The researchers demonstrated that editing these internal patterns changes Claude's outputs (swapping 'spider' for 'ant' changed a legs-count answer from 8 to 6), and when J-space was deleted, Claude could still chat and recall facts but struggled with multi-step problem solving. While researchers explicitly state this doesn't prove Claude is conscious or can feel, the discovery of an undesigned, brain-like workspace motivates continued investigation into AI consciousness.
Tencent open-sourced Hy3, a model using only a fraction of its parameters per request, achieving performance comparable to models with 2-5x more parameters while requiring less than half the hardware of larger competitors. The Apache 2.0 license avoids regional restrictions that plagued earlier Chinese models. Kyutai and General Intuition released MIRA, a world model that generates live 2v2 Rocket League gameplay entirely within a neural network without a game engine, learned from 10,000 hours of AI gameplay footage. The application runs at 20 fps on a single GPU and even generates hallucinated replay footage when needed, demonstrating world models' potential for training robots rather than replacing games.
The newsletter featured a reader workflow using MacWhisper transcripts of language lessons fed to Claude to generate song lyrics, which are then converted to audio via Suno, creating a personalized vocabulary-reinforcement playlist. Additional news included Illinois becoming the first U.S. state to mandate annual third-party AI safety audits, OpenAI's upcoming GPT 5.6 Ultra variant, and ByteDance and Alibaba discontinuing AI companion features ahead of Beijing regulations targeting emotional dependency.
About this episode
PLUS: Build any mobile app idea in 15 minutes with Replit
Key Insights
- Anthropic found an unspoken internal workspace ('J-space') in Claude that emerged spontaneously during training without explicit programming, functioning analogously to human conscious thought structures.
- Modifying Claude's internal patterns directly changes its outputs, as demonstrated by swapping conceptual representations ('spider' to 'ant') to alter numerical answers in reasoning tasks.
- When J-space was deleted, Claude retained factual recall and basic conversational ability but experienced significant degradation in multi-step problem-solving, indicating this structure is critical for complex reasoning.
- Tencent's Hy3 achieves competitive performance with models containing 2-5x more parameters by using only a fraction of its parameters per request, requiring less than half the computational hardware.
- Kyutai's MIRA world model learned 2v2 Rocket League gameplay exclusively from 10,000 hours of AI-vs-AI footage without human gameplay data or a traditional game engine, achieving 20 fps on single-GPU inference.
Topics
Transcript
Good morning, {{ first_name | 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. 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 The Rundown: Anthropic just published research showing that Claude does its thinking in a…
Full transcript available for MurmurCast members
Sign Up to AccessMore from The Rundown AI
Meta sizes up GPT-5.5 with 'Watermelon'
Meta's upcoming 'Watermelon' AI model reportedly matches OpenAI's GPT-5.5 performance while still in training, though the AI frontier continues advancing with competing models like Mythos and Fable. The newsletter covers Meta's $145B AI investment progress, practical AI applications from users, and emerging AI tools across education, coding, and consumer devices.
Altman invites Washington inside the AI industry
Sam Altman proposes a U.S.-led international AI safety forum modeled after the IAEA, with reports suggesting OpenAI discussed offering the government a 5% equity stake. Meanwhile, Meta's smart glasses capture 80% of the market but still fall short of replacing phones, and specialized AI models trained on expert data outperform frontier models at significantly lower costs.
Anthropic's Fable returns worldwide
Anthropic's Fable 5 AI model has returned to public access after a 3-week government-mandated shutdown, now featuring enhanced safety filters and requiring U.S. government pre-release access to future models. The newsletter also covers Meta's cloud compute rental plans, AI design tools, and new benchmarks showing Fable 5 achieving record performance on freelance task automation at 16.1% professional quality.
Sonnet 5 ships as Washington frees Fable
Anthropic launched Sonnet 5, a mid-tier model with improvements in agentic coding and reasoning, though its release is overshadowed by the imminent return of the more powerful Fable 5 and Mythos 5 models after export control restrictions were lifted. The newsletter also covers new AI tools from Google, Anthropic's Claude Science platform, and emerging developments in AI hardware and infrastructure.
Meta's brain-reading AI leaves letters behind
Meta's Brain2Qwerty v2 achieves a major breakthrough in non-invasive brain-computer interfaces, reading full sentences with 61% average word accuracy—nearly matching surgical implant performance. The newsletter also covers advances in AI-powered coding tools, automation platforms, and research into AI's integration into daily routines.