Mathematical reasoning continues to be a critical challenge in large language model (LLM) development with significant interest. However, most of the cutting-edge progress in mathematical reasoning with LLMs has become \emph{closed-source} due to lack of access to training data. This lack of data access limits researchers from understanding the impact of different choices for synthesizing and utilizing the data. With the goal of creating a high-quality finetuning (SFT) dataset for math reasoning, we conduct careful ablation experiments on data synthesis using the recently released \texttt{Llama3.1} family of models. Our experiments show that: (a) solution format matters, with excessively verbose solutions proving detrimental to SFT performance, (b) data generated by a strong teacher outperforms equally-sized data generated by a weak student model, (c) SFT is robust to low-quality solutions, allowing for imprecise data filtering, and (d) question diversity is crucial for achieving data scaling gains. Based on these insights, we create the OpenMathInstruct-2 dataset, which consists of 14M question-solution pairs ($\approx$ 600K unique questions), making it nearly eight times larger than the previous largest open-source math reasoning dataset. Finetuning the \texttt{Llama-3.1-8B-Base} using OpenMathInstruct-2 outperforms \texttt{Llama3.1-8B-Instruct} on MATH by an absolute 15.9\% (51.9\% $\rightarrow$ 67.8\%). Finally, to accelerate the open-source efforts, we release the code, the finetuned models, and the OpenMathInstruct-2 dataset under a commercially permissive license.

In AI, looking at development costs in terms of total cost of ownership avoids the common mistake of looking only at raw costs. As this analysis shows, the benefits of faster arrival, less wear and tear, and fewer opportunities for detours, accidents, congestion, or wrong turns make it a smarter choice for our road trips. The same is true for optimized AI processing. The term AI governance has recently taken on many meanings, from ethics to explainability. Here, it refers to the ability to measure cost, value, auditability, and compliance with regulatory standards, especially as it relates to data and customer information.

Originally published on Towards AI the World's Leading AI and Technology News and Media Company. If you are building an AI-related product or service, we invite you to consider becoming an AI sponsor. At Towards AI, we help scale AI and technology startups. Let us help you unleash your technology to the masses. According to Alation's State of Data Culture Report, 87% of employees attribute poor data quality to why most organizations fail to adopt AI meaningfully.

The success of machine learning for a wide range of applications has come with serious costs. The largest deep neural networks can have hundreds of billions of parameters that need to be tuned to mammoth datasets. This computationally intensive training process can cost millions of dollars, as well as large amounts of energy and associated carbon. Inference, the subsequent application of a trained model to new data, is less demanding for each use, but for widely used applications, the cumulative energy use can be even greater. "Typically there will be more energy spent on inference than there is on training," said David Patterson, Professor Emeritus at the University of California, Berkeley, and a Distinguished Engineer at Google, who in 2017 shared ACM's A.M. Turing Award.

Despite the vast potential of artificial intelligence (AI), it hasn't caught hold in most industries. The majority of AI projects hit a wall. Accenture estimates that 80% to 85% of companies' AI projects are in the proof-of-concept stage, and nearly 80% of enterprises fail to scale AI deployments across the organization. Once AI is built, many organizations struggle to get their models to production. It can take months, even years, to deliver value from your AI.

Business use of AI grew 270% over the past four years, according to Gartner, while Deloitte says 62% of respondents to its corporate October 2018 report deployed some form of AI. That's up 53% from a year ago, but what we've learned is that adoption doesn't equal success, and success is an evolving model in this phase of our digital revolution. Unfortunately, roughly 25% of companies have seen half of their AI projects fail. Failure, in heavily technical deployments, like AI projects, is incredibly expensive when data scientist and other team time, technical cost of computation, and resources wasted is accounted for. Statistics like these have generated tremendous buzz around the end results: success or failure, but we've reached a pivot point where we must widen our lens and shift our attention.

In July, VMware acquired Bitfusion, a company whose technology virtualizes compute accelerators with the goal of enabling modern workloads like artificial intelligence and data analytics to take full advantage of systems with GPUs or with FPGAs. Specifically, Bitfusion's software allows for virtual machines to offload compute duties to GPUs, FPGAs, or even other kinds of ASICs. The deal didn't get a ton of attention at the time, but for VMware, it was an important step in realizing its cloud ambitions. "Hardware acceleration for applications delivers efficiency and flexibility into the AI space, including subsets such as machine learning," Krish Prasad, senior vice president and general manager of VMware's Cloud Platform business unit, wrote in a blog post announcing the acquisition. "Unfortunately, hardware accelerators today are deployed with bare-metal practices, which force poor utilization, poor efficiencies, and limit organizations from sharing, abstracting and automating the infrastructure. This provides a perfect opportunity to virtualize them – providing increased sharing of resources and lowering costs."

"Hardware acceleration for applications delivers efficiency and flexibility into the AI space, including subsets such as machine learning," Krish Prasad, …

Today at the Microsoft Build developer conference, we are announcing a partnership with Qualcomm, one of the largest mobile and IoT chipset manufacturers in the world, to jointly create a vision AI developer kit. This will empower Qualcomm's latest AI hardware accelerators to deliver real-time AI on devices without the need for constant connectivity to the cloud or expensive machines. This vision AI developer kit brings all the key hardware and software required to develop camera-based IoT solutions using Azure IoT Edge and Azure Machine Learning (ML) – helping innovators deliver the next generation of AI-enabled robotics, industrial safety, retail, home and enterprise security cameras, smart home devices and more. This is a crucial step toward enabling developers to easily create, manage and monitor AI on the edge. This partnership allows developers to start building AI offerings with prebuilt solutions -- including customizable models -- or create new AI models and deploy directly to the cloud or to the new hardware accelerated devices.

SiFive does a quarterly series of tech talks, not necessarily directly to do with SiFive or even RISC-V. For example, last quarter it was Paul Kocher (and if you don't know that name, you need to go and read my post about that talk Paul Kocher: Differential Power Analysis and Spectre). This quarter it was Krste Asanović on Accelerating AI: Past, Present, and Future. This post will cover the past. The present and future have to wait (good title for a movie?).