# Dec 29, 2025

**ML Times**  Dec 29, 2025

## Daily

## Weekly

### Fast GPU Interconnect over Radio

- **Radio-based cables** from startups like Point2 and AttoTude promise to **outperform copper** by offering longer reach, lower power consumption, and reduced latency, crucial for scaling AI data centers efficiently.

### Doublespeak: In-Context Representation Hijacking

- **Doublespeak** introduces a **novel attack** on large language models (LLMs) by substituting harmful keywords with benign tokens, leading to a **semantic hijacking** that allows harmful instructions to bypass safety mechanisms.

### A bitwise reproducible deep learning framework

- **RepDL** ensures **bitwise identical outcomes** across hardware platforms for deep learning tasks, enhancing reproducibility in research and development.

### [R] Sophia: A Framework for Persistent LLM Agents with Narrative Identity and Self-Driven Task Management

- **Sophia introduces a third stratum, System 3**, which enhances AI agents by integrating a persistent meta-layer for identity maintenance and long-term adaptation, transforming them into self-improving entities.

### [P] Fine-tuned 8B model for Quantum Cryptography

- A **fine-tuned 8B model** for quantum cryptography demonstrates **domain accuracy** of **85-95%** on QKD tasks, significantly outperforming the base model, which fails in these areas.