Reproducing J-space and Reading Model's Hidden Thoughts
vikvang1 · x · 2026-07-11
The author reproduced Anthropic's J-space research and migrated it to Llama-3.3-70B, claiming they can "read" the model's thoughts hidden in J-space.
Key takeaways:
- They split concept vectors (like secrecy, elephant, democracy) into J-space and "subconscious" components, injecting them separately.
- The model directly reports thoughts from J-space but is "blind" to non-J-space content.
- Even if the model verbally denies it, the author notes the NLA can still capture hidden thoughts, which continue to influence its behavior.
- The post includes a code link, emphasizing this is a reproduction and extension of Anthropic's "Global Workspace in Language Models" research.
Related event: Reproducing J-space to Read Hidden Thoughts in Llama Models(2 posts)→
More from Models
- Rumor claims Kimi faked performance by serving Claude; DeepSeek new model surprises in evals — realsohamparekh · 2026-09-11
- GPT-5.6 writes well but is instantly forgettable, user complains — BasedRaddka · 2026-09-11
- Opus Refuses Protein Research Codebase Over 'Safety' Concerns, Dev Considers Rolling His Own — josephdviviano · 2026-09-11
- User Hails Unconfirmed 'DeepSeek 4.1 Flash' as an Inflection Point in LLMs — himanshustwts · 2026-09-11
- Terminal Bench v4: GLM-5.3 Leads at 41.9%, Kimi-K3 Underwhelms at 12.6% — Ok_Warning2146 · 2026-09-11
- GPT-6 Astra beats Factorio with enemies in 44 in-game hours at ~$4,500 API cost — liminal_bardo · 2026-09-11