# ML Times

## Google CEO says more than a quarter of the company's new code is created by AI

- **Over 25% of new code at Google is generated by AI**, significantly enhancing productivity and efficiency, as confirmed by CEO Sundar Pichai during the Q3 earnings call.

## Chain-of-Thought Can Hurt Performance on Tasks Where Thinking Makes Humans Worse

- **Chain-of-thought (CoT) prompting can lead to a significant drop in model performance**, with reductions up to **36.3% absolute accuracy** in certain tasks, highlighting the need for careful application in specific contexts.

## AI Flame Graphs

- **AI Flame Graphs** are a new tool from Intel designed to visualize AI accelerator performance, potentially reducing resource costs and contributing to a **10% decrease in US power usage by 2030**.

## Pushing the frontiers of audio generation

- **Google DeepMind's latest audio generation technology** enables the creation of **2-minute dialogues** with improved naturalness and speaker consistency, utilizing advanced models that process dialogue scripts and speaker markers in under **3 seconds** on a TPU v5e chip.

## DeepSeek v2.5 – open-source LLM comparable to GPT-4, but 95% less expensive

- **DeepSeek-V2.5** excels in performance, ranking in the **top 3** on AlignBench and rivaling **GPT-4-Turbo**, showcasing its advanced capabilities in math, coding, and reasoning.

## Cerebras Trains Llama Models to Leap over GPUs

- **Cerebras has achieved a remarkable 3.5X increase in AI inference performance with its Llama 3.2 models, significantly outperforming Nvidia's H100 GPUs.**

## Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro

- **Qodo now integrates Claude Sonnet 3.5, OpenAI o1, and Google Gemini 1.5 Pro**, enhancing its platform with advanced models that improve coding tasks and problem-solving capabilities, with full access set to launch next week.

## The National EUV Accelerator comes to Albany

- The **NSTC EUV Accelerator** will be established at the **Albany NanoTech Complex**, enhancing North America's semiconductor research and manufacturing capabilities, crucial for maintaining technological leadership.

## Physical Intelligence's first generalist robotic model

- **π0** is a groundbreaking generalist robot foundation model designed to enable robots to perform a wide range of tasks by learning from embodied experiences, akin to how large language models (LLMs) operate with text and images.

## ] Experimental Design for Multi-Channel Imaging via Task-Driven Feature Selection (ICLR)

- The paper introduces a **novel method for supervised feature selection** that enhances task-driven image channel selection, significantly impacting MRI acquisition times and multispectral image reconstruction.

## ] Our results experimenting with different training objectives for an AI evaluator

- **Preference optimisation techniques** like DPO and RPO outperform traditional supervised fine-tuning (SFT) in training LLM-as-a-judge models, suggesting a shift in effective training objectives for AI evaluators.

## ] using LLMs to power novelty-seeking adaptive learning agents ?

- The proposed **novelty-seeking machine learning algorithm** leverages Gödel's completeness/incompleteness theorems, aiming to enhance efficiency by integrating **LLMs** for data generation and parsing, thus reducing the resource burden of training each component independently.

## Unified Triplet-Level Hallucination Evaluation for Large Vision-Language Models

- **Unified framework** for evaluating hallucinations in Large Vision-Language Models (LVLMs) addresses both **object** and **relation hallucination**, revealing a critical gap in existing benchmarks that focus solely on object-related issues.

## Introducing DRIFT Search: Combining global and local search methods to improve quality and efficiency

- **DRIFT Search** enhances local query responses by integrating community insights, allowing for a broader range of facts and more nuanced answers compared to traditional methods.

## BUZZ: Beehive-structured Sparse KV Cache with Segmented Heavy Hitters for Efficient LLM Inference

- **BUZZ** introduces a **beehive-structured sparse KV cache** that optimizes memory usage and enhances inference speed for large language models (LLMs) by dynamically segmenting historical tokens and utilizing a sliding window for recent information capture.