# Oct 9, 2025

## A small number of samples can poison LLMs of any size
- **As few as 250 malicious documents can create a "backdoor" vulnerability in large language models (LLMs), regardless of their size or training data volume**, challenging the notion that attackers need a percentage of training data to succeed.

## Introducing Figure 03
- **Figure 03** is a **third-generation humanoid robot** engineered for versatility, featuring a redesigned sensory suite and hand system that enhances its ability to learn and perform human-like tasks in various environments, including homes and commercial settings.

## Representation Engineering (2024)
- **Control vectors** enable manipulation of AI model behavior during inference by modifying hidden states, allowing for nuanced control without the need for prompt engineering or finetuning, as demonstrated in the recent paper on [Representation Engineering](https://arxiv.org/abs/2310.01405).

## A History of Large Language Models
- **Large Language Models (LLMs)** have evolved through key innovations like the **attention mechanism** and the **transformer architecture**, which replaced traditional recurrent models, enabling efficient training and improved performance on various tasks.

## Neutts-air – open-source, on device TTS
- **NeuTTS Air** is the first on-device TTS model that offers **instant voice cloning** and **real-time performance**, enabling applications like voice assistants and toys without relying on web APIs.

## Microsoft Azure Unveils World’s First NVIDIA GB300 NVL72 Supercomputing Cluster for OpenAI
- **Microsoft Azure's new NDv6 GB300 VM series** introduces the world's first supercomputing-scale cluster of **NVIDIA GB300 NVL72 systems**, designed specifically for OpenAI's advanced AI workloads, enhancing the capabilities for model development and deployment.

## Artificial Hippocampus Networks for Efficient Long-Context Modeling
- The **Artificial Hippocampus Network (AHN)** framework enhances long-sequence modeling by combining **lossless short-term memory** from Transformers with a **fixed-size long-term memory**, leading to significant efficiency gains.

## Native Hybrid Attention for Efficient Sequence Modeling
- **Native Hybrid Attention (NHA)** combines **linear** and **full attention** in a unified architecture, enhancing efficiency without sacrificing recall accuracy in long contexts.

## The Markovian Thinker
- **Markovian Thinking** revolutionizes reinforcement learning (RL) for reasoning LLMs by maintaining a **constant-size state**, allowing for linear compute and efficient memory usage, as demonstrated by the Delethink environment.

## Introducing the Gemini 2.5 Computer Use model
- The **Gemini 2.5 Computer Use model** is now available via the Gemini API, enabling developers to create agents that interact with user interfaces, outperforming competitors in web and mobile tasks with **lower latency**.

## Defining and evaluating political bias in LLMs
- **AI-native 6G networks** will revolutionize telecommunications by enabling **real-time AI traffic management**, supporting applications like autonomous vehicles and smart agriculture, thus positioning the U.S. as a leader in the global AI economy.

## How AI-Powered Wireless Networks Will Revitalize US Global Leadership in Communications
- **AI web agents face significant challenges with bot detection**, as real websites employ sophisticated systems that flag agents for unnatural behavior, such as pixel-perfect clicks and instant actions.

## GenPilot: A Multi-Agent System for Test-Time Prompt Optimization in Image Generation
- **GenPilot** introduces a **multi-agent system** for **test-time prompt optimization**, enhancing text-to-image synthesis by addressing semantic inconsistencies and improving interpretability of complex prompts.

## ELMUR: External Layer Memory with Update/Rewrite for Long-Horizon RL
- **ELMUR** (External Layer Memory with Update/Rewrite) introduces a **transformer architecture** that effectively manages long-term dependencies in decision-making by utilizing structured external memory, significantly enhancing performance in complex environments.