# Dec 17, 2025

## Gemini 3 Flash: frontier intelligence built for speed
- **Gemini 3 Flash** is a new AI model that combines **frontier intelligence** with **high speed** and **cost efficiency**, enabling users to learn and build faster than ever before.

## AI will make formal verification go mainstream
- **AI is poised to revolutionize formal verification**, transitioning it from a niche practice to a mainstream necessity in software engineering, driven by advancements in LLM-based coding assistants that can automate proof script generation.

## Nvidia Nemotron 3 Family of Models
- **NVIDIA Nemotron 3** introduces a family of models— **Nano, Super, and Ultra**—that excel in agentic AI applications, with Nano achieving superior accuracy while being cost-efficient for inference.

## [P] Eigenvalues as models
- **Neurons may require enhanced computational capabilities**, as suggested by Sutskever, prompting exploration of **eigenvalues** in optimization within neural networks.

## The State of AI Coding Report 2025
- **AI coding tools have significantly boosted developer productivity**, with lines of code per developer increasing by **76%**, from 4,450 to 7,839, as reported in Greptile's internal data.

## Show HN: TheAuditor v2.0 – A ”Flight Computer“ for AI Coding Agents
- **TheAuditor** is an **AI-centric SAST** platform that not only detects security vulnerabilities but also tracks data flow and analyzes architecture, ensuring a reliable source of **ground truth** for AI-assisted development workflows.

## Show HN: High-Performance Wavelet Matrix for Python, Implemented in Rust
- **JavaScript loading issues** can hinder site functionality, often due to browser settings or extensions, impacting user experience significantly.

## [P] Cyreal - Yet Another Jax Dataloader
- **Cyreal** is a **fast, lightweight, and flexible JAX dataloader** that operates independently of other libraries, addressing the common issues of dependency conflicts and slow performance found in existing solutions like Grain.

## Denoising Language Models for Speech Recognition
- **Denoising language models (DLMs)** offer a **superior alternative** to traditional language models (LMs) for automatic speech recognition (ASR), particularly when trained with richer contextual information from ASR hypotheses, as demonstrated in a large-scale empirical study.

## [D] Recent research in training embedding models
- **Recent advancements** in training embedding models highlight **CodeT5+ by Salesforce** as a leading method, emphasizing its focus on understanding _source code_ effectively. [arXiv](https://arxiv.org/abs/2305.07922)

## UC San Diego Lab Advances Generative AI Research With NVIDIA DGX B200 System
- The **Hao AI Lab** at UC San Diego has integrated the **NVIDIA DGX B200**, a cutting-edge AI system, to enhance their research in **large language model (LLM) inference**, significantly accelerating projects like **FastVideo** and **Lmgame benchmark**.

## Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed
- **Diffusion language models (dLMs)** can achieve **superior speed** and efficiency through **AR-to-dLM conversion**, which transforms pretrained autoregressive models while maintaining their task accuracy.

## RePo: Language Models with Context Re-Positioning
- **RePo** introduces a **differentiable mechanism** for context re-positioning in LLMs, enhancing cognitive efficiency by reducing extraneous load and improving attention allocation to relevant information.

## Evaluating AI’s ability to perform scientific research tasks

## Measuring AI’s capability to accelerate biological research tasks