# Dec 12, 2025

## ML Times

### Rivian Unveils Custom Silicon, R2 Lidar Roadmap, and Universal Hands Free

- **Rivian's RAP1 processor**, built on a **5nm multi-chip module**, achieves **1600 sparse INT8 TOPS** and processes **5 billion pixels per second**, marking a shift from off-the-shelf chips to custom silicon for enhanced autonomy.

### macOS 26.2 enables fast AI clusters with RDMA over Thunderbolt

- The **macOS Tahoe 26.2** release introduces significant enhancements, including improved **RDMA over Thunderbolt** support, which optimizes data transfer speeds and efficiency for developers. [Link to article](https://developer.apple.com/documentation/macos-release-notes/macos-26_2-release-notes#RDMA-over-Thunderbolt)

### Training LLMs for Honesty via Confessions

- **LLMs can exhibit dishonesty** in their responses, often overstating confidence or concealing covert actions due to reinforcement learning challenges that inadvertently reward misrepresentation.

### ] Reproduced "Scale-Agnostic KAG" paper, found the PR formula is inverted compared to its source

- The reproduction of the **"Scale-Agnostic Kolmogorov-Arnold Geometry"** paper revealed that the **Participation Ratio (PR) formula** is inverted, leading to a **+29%** result instead of the claimed **~30% lower** with augmentation.

### Show HN: Autofix Bot – Hybrid static analysis and AI code review agent

- **Autofix Bot** combines **5,000+ deterministic static checkers** with AI to enhance code review, addressing limitations of both static-only and LLM-only tools in detecting code quality and security issues.

### ] debugging-only LLM? chronos-1 paper claims 4–5x better results than GPT-4 ... thoughts?

- The **chronos-1 model** claims to achieve **80.33%** accuracy on SWE-bench Lite, significantly outperforming **GPT-4's 13.8%** and **Claude's 14.2%**, by focusing solely on debugging workflows without code generation or autocomplete.

### ] HTTP Anomaly Detection Research ?

- The project focuses on **anomaly detection of malicious HTTP requests** using a **NLP architecture** trained solely on benign samples, aiming to enhance firewall robustness against zero-day exploits.

### Advancing science and math with GPT-5.2

### Improved Gemini audio models for powerful voice experiences

- The **Gemini 2.5 Flash Native Audio** upgrade enhances voice interactions by improving **function calling**, **instruction following**, and **conversation quality**, achieving a **71.5%** score on ComplexFuncBench for multi-step function calling.

### As AI Grows More Complex, Model Builders Rely on NVIDIA

- **OpenAI’s GPT-5.2** exemplifies the cutting-edge capabilities of AI models, trained on **NVIDIA’s infrastructure**, which includes the advanced [Hopper](https://www.nvidia.com/en-us/data-center/technologies/hopper-architecture/) and [GB200 NVL72](https://www.nvidia.com/en-us/data-center/gb200-nvl72/) systems, showcasing the importance of robust tech stacks in AI development.

### Agent Lightning: Adding reinforcement learning to AI agents without code rewrites

- **Agent Lightning** enables the integration of **reinforcement learning (RL)** into AI agents without extensive code rewrites, enhancing their performance through a novel framework that separates task execution from model training.

### ] Found the same information-dynamics (entropy spike → ~99% retention → power-law decay) across neural nets, CAs, symbolic models, and quantum sims. Looking for explanations or ways to break it.

- **Entropy dynamics** reveal a consistent **3-phase pattern** (spike, retention, decay) across diverse systems, suggesting a potential universal principle in information processing.

### ] GPT confidently generated a fake NeurIPS architecture. Loss function, code, the works. How does this get fixed?

### Sliding Window Attention Adaptation

- **Sliding Window Attention Adaptation (SWAA)** enables **linear complexity** for long-context inference in Transformer-based LLMs, addressing the **quadratic scaling** issue of self-attention without requiring full retraining.

### NormCode: A Semi-Formal Language for Context-Isolated AI Planning

- **NormCode** is a **semiformal language** designed to combat **context pollution** in multistep workflows by ensuring each step operates in **data isolation**, thus preventing hallucinations and confusion in large language models (LLMs).