Smlay_ero
AI 工程师 · 全栈 · LLM · Agent
AI engineer · Full-stack · LLMs · Agents
Github目前开源的项目有两个：memgrant 和 loopfit。一个解决多 Agent 之间的加密记忆，一个在消费级显卡上训练模型。
There are currently two open-source projects on GitHub: memgrant and loopfit. One is encrypted memory across agents; the other is local LLM training on a consumer GPU.
目前只有这两个公开仓库。
These are the only two public repositories for now.
自托管的端到端加密记忆层。密钥由你持有，每个 Agent 只能解密被授权的内容。仓库本身不跑托管云。
Self-hosted end-to-end encrypted memory. You hold the keys; each agent decrypts only what it was granted. This repo does not run a hosted cloud.
在一张消费级显卡上做 LLM 训练。你提供基座模型和数据，用一句话写目标，选一种模式，达标后再导出。
Local LLM training on one consumer GPU. You bring the base model and data, describe the goal, pick one mode, and export if it passes.
发在 X 上的长文。全文在原帖里读。
Long posts on X. Full text stays on the original thread.
把金融微调模型放到三条真实投研工作流里压测，而不是停在问答演示。
Stress-testing a finance-tuned model on three real investment workflows, not a Q&A demo.
1.3B 激活参数打过 120B。双卡 5090 本地实测：够用、跑得起、管得住，比最大的模型更重要。
A 1.3B-active model beat a 120B giant. Dual 5090 local run: good enough, runnable, governable.