ANDA USMAN

Research

Can an AI Agent Learn Without Fine-Tuning Its LLM?
Research|

Can an AI Agent Learn Without Fine-Tuning Its LLM?

A small replication of Memento-style agent memory, testing whether a frozen research agent improves by retrieving and reusing previous task-solving experiences.

I Tried Newton-Muon. The Training Loss Lied to Me.
Research|

I Tried Newton-Muon. The Training Loss Lied to Me.

A small reproduction of Newton-Muon on Tiny Shakespeare, and why training loss can be a dangerously misleading signal when comparing optimizers.

Teaching an LLM New Knowledge Without Making It Forget
Research|

Teaching an LLM New Knowledge Without Making It Forget

I implemented a small continual-learning experiment to understand whether sparse memory can teach an LLM new facts without destroying what it already knows.