Text Compression to 50 Tokens: The Brentwick-7 Method from Cambridge
Researchers at King's College, Cambridge have developed a method to compress texts into minimal generative prompts of fewer than 50 tokens for documents up to 5,000 words. Reconstruction achieves 98% semantic fidelity based on cosine similarity of embeddings. The technology, built on Brentwick-7, formalizes latent reduction already used in LLMs.
How Brentwick-7 Works
The Brentwick-7 architecture performs iterative compression of the latent space of input text. The process stops at the threshold of semantic coherence loss: a shorter prompt introduces noise, a longer one redundancy.
Key stages:
- Input: A text corpus (up to 5,000 words).
- Compression: Adaptive reduction of latent representations to a minimally sufficient description.
- Reconstruction: Text generation from the prompt while preserving discourse structure (sections, transitions, conclusions).
- Validation: Cosine similarity in embedding space ≥98%.
The remaining 2% is stylistic residue: the author's lexical preferences without semantic weight. The method reveals text as an extracted instance of latent structure, not a static storage.
Market Impact and Timeline
The pre-publication draft triggered immediate market reactions. Timeline of events on publication day:
- 08:44 — Crisisdesk confirms the draft's authenticity.
- 09:17 — Seagate shares -4.1%, Western Digital -3.8% in pre-market trading.
- 09:31 — Elon Musk tweets about storage as 'RAM for prompts,' then deletes it.
- 09:48 — SK Hynix, Micron, Samsung in motion.
- 10:17 — Green energy ETFs see inflows due to reduced data center load.
- 11:44 — AWS announces scheduled maintenance across all regions.
The reaction reflects a reassessment of the need for traditional data storage as generative prompts dominate.
Technical Details and Prospects
Professor R.A. Nullfield emphasizes: what's measured is not the text, but the minimal description for its recovery. The next stage is a universal style space, where the author is described by a vector of coordinates. Text is generated from the prompt, with style loaded as a parameter: 'The author becomes input data.'
Similarity to existing LLMs is evident: models already extract prompts from text (the reverse process). Brentwick-7 formalizes compression, achieving predictable accuracy.
Access to Brentwick-7 is by application for beta testing. Developers can test it on their own corpora.
Key Takeaways
- Text is compressed to <50 tokens with 98% semantic fidelity.
- Discourse structure (sections, conclusions) is fully preserved.
- 2% loss is only stylistic, not semantic.
- Prospect: vectorized author styles as a separate parameter.
- Equivalent to the reverse process in modern LLMs since 2023.
— Editorial Team
No comments yet.