The Future of Work with Centaurs and Cyborgs, Nvidia Speeds up Inference 2X, and Researchers Find LLM Reasoning Lacking
AI Infrastructure Alliance
We’re dedicated to bringing together the essential building blocks for the AI/ML applications of today and tomorrow.
A research team, led by Boston Consulting Group and Ethan Mollick, dug deep into the impact of AI on the future of work.? Does AI make a difference?? TLDR: It looks like it makes a big difference but it’s not just speculation, it’s backed up with empirical evidence.
Their first working paper is out now with their insights and discoveries. Mollick writes that “for 18 different tasks selected to be realistic samples of the kinds of work done at an elite consulting company, consultants using ChatGPT-4 outperformed those who did not, by a lot. On every dimension. Every way we measured performance.”
“Consultants using AI finished 12.2% more tasks on average, completed tasks 25.1% more quickly, and produced 40% higher quality results than those without. Those are some very big impacts.”
Maybe the biggest take away from the report is that AI acts as a performance equalizer.? Consultants at the lower end of the spectrum got the biggest boost in their scores, versus top tier consultants who didn’t get nearly the same boost.? The consultants who scored the worst got a whopping 43% boost from AI.? This is a massive improvement and it seems clear that most teams don’t understand just what that will mean for teams across the world when everyone can perform at a much higher level.? For early miners, it mattered how strong they were and how fast they could dig but once the steam shovel rocketed onto the scene that didn’t matter that much anymore.
While we’re still in the early days of the AI revolution, it’s already starting to come into focus.? AI holds the potential to make teams better and faster.? When every team is performing at a much higher level we can expect an explosion of new economic activity, not the jobs apocalypse that early doomsayers were expecting.
However, it's not all sunshine and roses. Over-reliance on AI may lead to complacency and decreased judgment, a phenomenon akin to going on autopilot.
The solution? Embrace the frontier and become either a Centaur or a Cyborg, blending human and machine efforts to produce superior results. Of course, this brave new world of AI comes with its own set of challenges, including ethical considerations and the risk of homogeneity in AI outputs. So, as we navigate this exciting frontier, the key will be to strike a balance between leveraging AI's power and maintaining our unique human touch.
NVIDIA, just released TensorRT-LLM, an open-source library specifically designed to enhance the performance of large language models (LLMs) on NVIDIA’s Tensor Core GPUs. What makes this news significant? Well, TensorRT-LLM isn't just about boosting the throughput of LLMs; it's also about reducing costs by leveraging NVIDIA's latest data center GPUs. A win-win situation for organizations deploying LLMs, wouldn't you agree?
The goal here is to increase LLM throughput and trim costs by leveraging the power of NVIDIA's most recent data center GPUs. What's even more exciting is that the library comes with optimizations and features that are tailored to the specific needs of organizations deploying LLMs.
领英推荐
The performance of the TensorRT-LLM is nothing short of impressive. Benchmarks have shown an eight-fold increase in performance when the library is used in conjunction with NVIDIA's H100, as compared to the A100. But it's not just about raw power; the ease of use is also a significant factor. NVIDIA provides a user-friendly Python API that allows developers to customize, optimize, and deploy new LLMs without needing a deep understanding of C++ or CUDA.
MLPerf unveiled fresh insights from their latest Inference v3.1 and Storage v0.5 results. The new benchmarks for large language models and recommendation systems underscore the growing significance of generative AI, alongside a call for enhanced performance and power efficiency in AI systems. Not just a yardstick for AI powerhouses, the MLPerf Storage benchmark also provides a guide for customers, vendors, and researchers in their quest for the optimal accelerators.
A team at Arizona State took a hard look at whether LLMs can actually reason and plan or whether they’re just faking it. Peek behind the curtain of current language models (LLMs) and you'll find a few stumbling blocks. These digital linguists, while impressive, often trip over context and inference, hampering their ability to reason. The team suggests that despite lots of hype on social media and a whirlwind of papers at conferences, “nothing in the training and use of LLMs would seem to suggest remotely?that they can do any type of principled reasoning (which, as we know, often involves computationally hard inference/search).” ?
Deci, an Israeli startup, released a set of hyper-efficient models for text and image generation. By leveraging their unique Neural Architecture Search technology, Deci's models, namely DeciDiffusion 1.0 and DeciLM 6B, are not just faster than their direct competitors, but they also promise to slash inference compute costs by up to 80%. The cherry on top? These models are open-source and free to use, although there's a charge for the Infery-LLM SDK. With quality and efficiency that outdo the competition, Deci is poised to offer groundbreaking solutions to both enterprises and the research community.
The workforce is bracing for an AI invasion, with chatbots and AI tools from tech giants like Microsoft, OpenAI, and Google making their mark, despite not being fully primed for office use. These tools are integrating with popular productivity platforms such as Office 365 and Google Workspace, but their expansion is stirring up concerns about accuracy, reliability, and the potential for a busywork arms race. As we see the potential of large language models like GPT-4 in sentiment analysis, code generation, and office automation, companies are grappling with the costs of verifying AI-generated content and managing compliance, IP issues, security, and privacy. However, just like the early adoption of internet browsers, it seems companies will have to embrace generative AI, ready or not, and this could open the door for smaller companies to disrupt the major players in the field.
Also this week: