Fugatto, a new generalist audio synthesis and transformation model developed by NVIDIA, and ComposableART, an inference-time technique designed to enhance its capabilities. Fugatto distinguishes itself by its ability to follow free-form text instructions, often with optional audio inputs, addressing the challenge that audio data, unlike text, typically lacks inherent instructional information. The document details a comprehensive data and instruction generation strategy that leverages large language models (LLMs) and audio understanding models to create diverse and rich datasets, enabling Fugatto to handle a wide array of tasks including text-to-speech, text-to-audio, and audio transformations. Furthermore, ComposableART allows for compositional abilities, such as combining, interpolating, or negating instructions, providing fine-grained control over audio outputs beyond the training distribution. The text presents experimental evaluations demonstrating Fugatto's competitive performance against specialised models and highlights its emergent capabilities, such as synthesising novel sounds or performing tasks not explicitly trained for.

link: https://d1qx31qr3h6wln.cloudfront.net/publications/FUGATTO.pdf

The provided text argues that small language models (SLMs) are the future of agentic AI, positioning them as more economical and operationally suitable than large language models (LLMs) for the majority of tasks within AI agents. While LLMs excel at general conversations, agentic systems frequently involve repetitive, specialised tasks where SLMs offer advantages like lower latency, reduced computational requirements, and significant cost savings. The authors propose a shift to heterogeneous systems, where SLMs handle routine functions and LLMs are used sparingly for complex reasoning. The document also addresses common barriers to SLM adoption, such as existing infrastructure investments and popular misconceptions, and outlines a conversion algorithm for migrating agentic applications from LLMs to SLMs.

Link: https://arxiv.org/pdf/2506.02153

Welcome to PodXiv! In this episode, we dive into groundbreaking research from METR that introduces a novel metric for understanding AI capabilities: the 50%-task-completion time horizon. This unique measure quantifies how long humans typically take to complete tasks that AI models can achieve with a 50% success rate, offering intuitive insight into real-world performance.

The study reveals a staggering trend: frontier AI's time horizon has been doubling approximately every seven months since 2019, driven by improvements in reliability, mistake adaptation, logical reasoning, and tool use. This rapid progress has profound implications, with extrapolations suggesting AI could automate many month-long software tasks within five years, a critical insight for responsible AI governance and safety guardrails.

However, the research acknowledges crucial limitations. Current AI systems perform less effectively on "messier," less structured tasks and those requiring complex human-like context or interaction. These factors highlight that while impressive, the generalisation of these trends to all real-world intellectual labour requires further investigation. Tune in to explore the future of AI autonomy and its societal impact!

Paper: https://arxiv.org/pdf/2503.14499