[untitled]1 points·by msukhareva·vor 6 Monaten·0 comments1 commentsPost comment[–]msukhareva·vor 6 MonatenreplyI built an architecture that enables continuous learning for open-weight models.The "curse" of LLMs is the knowledge cut-off. RAG is just a band-aid. Google's "Titans" memory solves this, but costs millions to train.I built a "Grafted" version on a single desk-side GPU (€4.4k). I attached a trainable memory adapter to a frozen Qwen-2.5-0.5B model.The Result: It beat the vanilla model even when the model had the full context on the challenging Babilong benchmark.Vanilla Qwen (seeing context): 34.0% AccuracyGrafted Titans (memory retrieval): 44.7% AccuracyMy memory module effectively acts as a denoising filter, outperforming the model's native attention mechanisms.The Specs:Hardware: Single Nvidia DGX Spark Blackwell.Training Time: ~7 days.Output: Plug-and-play adapter for open weights.This proves we don't need industrial clusters to build AI that remembers.
The "curse" of LLMs is the knowledge cut-off. RAG is just a band-aid. Google's "Titans" memory solves this, but costs millions to train.
I built a "Grafted" version on a single desk-side GPU (€4.4k). I attached a trainable memory adapter to a frozen Qwen-2.5-0.5B model.
The Result: It beat the vanilla model even when the model had the full context on the challenging Babilong benchmark.
Vanilla Qwen (seeing context): 34.0% Accuracy
Grafted Titans (memory retrieval): 44.7% Accuracy
My memory module effectively acts as a denoising filter, outperforming the model's native attention mechanisms.
The Specs:
Hardware: Single Nvidia DGX Spark Blackwell.
Training Time: ~7 days.
Output: Plug-and-play adapter for open weights.
This proves we don't need industrial clusters to build AI that remembers.