Experiment claims 'direct memory injection' into LLMs scores 13/13 at D=120
Nearby_Indication474 · reddit · 2026-10-02
A Reddit user claims their AkbasCore D120 experiment injects an attention-layer 'memory snapshot' into a model without ever showing it text — the model then recalls the story 13/13 on tests, with 0 for controls, and detail changes tracked 13/13. Calibration on Mistral-7B found quality breaks below D80. Code and logs are on GitHub; claims are unverified.
More from Research
- AMap open-sources ABot-Recon: streaming 3D reconstruction from video with a 12-frame local context — rsasaki0109 · 2026-10-02
- Neuralink pretrains decoders on 50,000+ hours of neural data—dataset may be the real moat — CurieuxExplorer · 2026-10-02
- CyberGym cybersecurity benchmark effectively saturated on verified task subset — aryaman2020 · 2026-10-02
- SemEval-2027 Task 9 calls for teams on 4-language multimodal news framing analysis — preslav_nakov · 2026-10-02
- Alternating prompt and model upgrades lift science agent from 42% to 73% — rohanpaul_ai · 2026-10-02
- ISMIR paper teaches a transformer to play in 12 jazz piano legends' styles — umpedronosapato · 2026-10-02