GatorTronGPT joins a growing number of large language models (LLMs) trained on clinical data. Researchers trained the model using the GPT-3 framework, also used by ChatGPT.
They used a massive corpus of 277 billion words for this purpose. The training corpora included 82 billion words from de-identified clinical notes and 195 billion words from various English texts.
But there’s a twist: The research team also used GatorTronGPT to generate a synthetic clinical text corpus with over 20 billion words of synthetic clinical text, with carefully prepared prompts. The synthetic clinical text focuses on clinical factors and reads just like real clinical notes written by doctors.
This synthetic data was then used to train a BERT-based model called GatorTron-S.
In a comparative evaluation, GatorTron-S exhibited remarkable performance on clinical natural language understanding tasks like clinical concept extraction and medical relation extraction, beating the records set by the original BERT-based model, GatorTron-OG, which was trained on the 82-billion-word clinical dataset.
More impressively, it was able to do so using less data.
Both GatorTron-OG and GatorTron-S models were trained on 560 NVIDIA A100 Tensor Core GPUs running NVIDIA’s Megatron-LM package on the University of Florida’s HiPerGator supercomputer. Technology from the Megatron LM framework used in the project has since been incorporated with the NVIDIA NeMo framework, which has been central to more recent work on GatorTronGPT.
Using synthetic data created by LLMs addresses several challenges. LLMs require vast amounts of data, and there’s a limited availability of quality medical data.
In addition, synthetic data allows for model training that complies with medical privacy regulations, such as HIPAA.
The work with GatorTronGPT is just the latest example of how LLMs — which exploded onto the scene last year with the rapid adoption of ChatGPT — can be tailored to assist in a growing number of fields.
It’s also an example of the advances made possible by new AI techniques powered by accelerated computing.
The GatorTronGPT effort is the latest result of an ambitious collaboration announced in 2020, when the University of Florida and NVIDIA unveiled plans to erect the world’s fastest AI supercomputer in academia.
This initiative was driven by a $50 million gift, a fusion of contributions from NVIDIA founder Chris Malachowsky and NVIDIA itself.
Using AI to train more AI is just one example of HiPerGator’s impact, with the supercomputer promising to power more innovations in medical sciences and across disciplines throughout the University of Florida system.
Human analysts can no longer effectively defend against the increasing speed and complexity of cybersecurity attacks. The amount of data is simply too large to screen manually.
Generative AI, the most transformative tool of our time, enables a kind of digital jiu jitsu. It lets companies shift the force of data that threatens to overwhelm them into a force that makes their defenses stronger.
Business leaders seem ready for the opportunity at hand. In a recent survey, CEOs said cybersecurity is one of their top three concerns, and they see generative AI as a lead technology that will deliver competitive advantages.
Here are three ways generative AI can bolster cybersecurity.
Begin With Developers
First, give developers a security copilot.
Everyone plays a role in security, but not everyone is a security expert. So, this is one of the most strategic places to begin.
The best place to start bolstering security is on the front end, where developers are writing software. An AI-powered assistant, trained as a security expert, can help them ensure their code follows best practices in security.
The AI software assistant can get smarter every day if it’s fed previously reviewed code. It can learn from prior work to help guide developers on best practices.
Whether users customize their own models or use a commercial service, a security assistant is just the first step in applying generative AI to cybersecurity.
An Agent to Analyze Vulnerabilities
Second, let generative AI help navigate the sea of known software vulnerabilities.
At any moment, companies must choose among thousands of patches to mitigate known exploits. That’s because every piece of code can have roots in dozens if not thousands of different software branches and open-source projects.
An LLM focused on vulnerability analysis can help prioritize which patches a company should implement first. It’s a particularly powerful security assistant because it reads all the software libraries a company uses as well as its policies on the features and APIs it supports.
To test this concept, NVIDIA built a pipeline to analyze software containers for vulnerabilities. The agent identified areas that needed patching with high accuracy, speeding the work of human analysts up to 4x.
The takeaway is clear. It’s time to enlist generative AI as a first responder in vulnerability analysis.
Fill the Data Gap
Finally, use LLMs to help fill the growing data gap in cybersecurity.
Users rarely share information about data breaches because they’re so sensitive. That makes it difficult to anticipate exploits.
Enter LLMs. Generative AI models can create synthetic data to simulate never-before-seen attack patterns. Such synthetic data can also fill gaps in training data so machine-learning systems learn how to defend against exploits before they happen.
Staging Safe Simulations
Don’t wait for attackers to demonstrate what’s possible. Create safe simulations to learn how they might try to penetrate corporate defenses.
This kind of proactive defense is the hallmark of a strong security program. Adversaries are already using generative AI in their attacks. It’s time users harness this powerful technology for cybersecurity defense.
To show what’s possible, another AI workflow uses generative AI to defend against spear phishing — the carefully targeted bogus emails that cost companies an estimated $2.4 billion in 2021 alone.
This workflow generated synthetic emails to make sure it had plenty of good examples of spear phishing messages. The AI model trained on that data learned to understand the intent of incoming emails through natural language processing capabilities in NVIDIA Morpheus, a framework for AI-powered cybersecurity.
The resulting model caught 21% more spear phishing emails than existing tools. Check out our developer blog or watch the video below to learn more.
Wherever users choose to start this work, automation is crucial, given the shortage of cybersecurity experts and the thousands upon thousands of users and use cases that companies need to protect.
These three tools — software assistants, virtual vulnerability analysts and synthetic data simulations — are great starting points for applying generative AI to a security journey that continues every day.
But this is just the beginning. Companies need to integrate generative AI into all layers of their defenses.
Attend a webinar for more details on how to get started.
Editor’s note: This post is part of Into the Omniverse, a series focused on how artists, developers and enterprises can transform their workflows using the latest advances in OpenUSD and NVIDIA Omniverse.
In 3D art and design, efficient workflows are essential for quickly bringing creative visions to life.
Universal Scene Description, aka OpenUSD, is a framework that enhances these workflows by providing a unified, extensible ecosystem for describing, composing, simulating and collaborating within 3D worlds. OpenUSD is a key technology in Autodesk’s suite of products and solutions, across media and entertainment; architecture, engineering and construction; and product design and manufacturing.
Unveiled at the AU 2023 conference this week, the latest OpenUSD updates to Autodesk Maya enable artists and technical professionals to create and manipulate OpenUSD assets with greater control and efficiency, while also ensuring more efficient and accurate 3D workflows.
Bridging the Digital and Real Worlds With Maya and OpenUSD
Many creators are using Maya and OpenUSD to propel their 3D workflows.
Karol Osinski is a 3D artist at S20M, an architectural and design firm that specializes in tackling unique, bold and elegant projects. When it comes to creating architectural visualizations, Osinski says the biggest challenge is matching the digital world to the real one.
Using USD and creative tools such as Maya, SideFX Houdini and Epic Games’ Unreal Engine, Osinski creates high-quality visuals for clients while accelerating his architectural workflows.
Osinski’s panoramic view from the 20th floor terrace in the Upper East Side
“OpenUSD provides the possibility of bridging different tools like never before,” said Osinski. “I love how accessible USD is for first-time users and how it opens opportunities to make designs very complex.”
“Sir Wade” Neistadt, an animator and YouTube creator, aims to make animation and 3D education more accessible through his video tutorials and industry training. The first step of his unique animation workflow is to act out his animations on camera. He then translates them in Maya to begin his animation work before using USD to export them to other 3D software, including Blender, for finishing touches.
The making of Sir Wade’s VFX robot animation
3D artists at NVIDIA are also experiencing the power of Maya and OpenUSD. Technical specialist Lee Fraser led the “Ferret-Tale Project” to showcase character creation and animation workflows enabled by OpenUSD and generative AI.
To create the demo, Fraser and his team collaborated across 3D applications like Blender, Autodesk Maya and Reallusion Character Creator through OpenUSD Connectors. This allowed them to reduce the data prep and import and export time that’s usually required when working with multiple data sources.
“My favorite thing about using OpenUSD is not having to think about where the 3D files I use originated from,” Fraser said. “It was also easy to use USD layers to experiment with applying different animation clips with different characters.”
Members of the creative community joined a recent livestream to share their workflows using Autodesk tools, OpenUSD and NVIDIA Omniverse, a development platform for connecting and building OpenUSD-based tools and applications.
Whether adjusting lighting conditions in an environment or looking at building designs from the street view, designers in architecture, engineering, construction and operations are advancing their work with AI. Learn more by watching the replay:
Shaping the Future of 3D With More Efficient Workflows
AU 2023 attendees experienced how Autodesk is enhancing Maya with its new OpenUSD plug-in to provide additional practical workflows for various production processes. The software’s latest features include:
Simplified asset sharing: Designers can now use relative paths when creating OpenUSD stages, allowing for easy asset sharing between different systems. This includes support for sublayers, references, payloads and textures.
Enhanced control: Plug-in developers and technical directors can overwrite the default prim writers in Maya USD to gain complete control over their OpenUSD exports.
Plus, Autodesk introduced impressive capabilities to LookdevX in Maya, a look-development tool that lets users create OpenUSD shade graphs and custom materials in Maya. These new features include:
Streamlined shader creation: Users can employ a unified shader workflow, replacing the need for multiple shaders. They can select their desired shader type within the parameters panel, with intuitive error messages guiding them to the correct selection.
Efficient operations: Creators can copy, paste and duplicate shaders and materials using the Outliner and LookdevX tool sets, with the option to include or exclude connections.
Seamless color management: LookdevX in Maya integrates with color managers in other digital content creation apps to ensure accurate color representation. Color management data is precisely embedded in USD files for accurate reading.
Advanced graphing: Users can explore advanced graphing options with the integrated component workflow, supporting multichannel Extended Dynamic Range (EXR) workflows within USD, MaterialX or Arnold shading graphs.
Efficient troubleshooting: Solo nodes enable faster look-development workflows and efficient graph troubleshooting. Users can inspect renders of upstream nodes, supporting both Autodesk Arnold and MaterialX graphs, including materials, shaders and compounds.
Access to default prim options in Maya UI
Get Plugged Into the World of OpenUSD
Anyone can build their own Omniverse extension or Connector to enhance their 3D workflows and tools. Explore the Omniverse ecosystem’s growing catalog of connections, extensions, foundation applications and third-party tools.
Share your Autodesk Maya and Omniverse work through November as part of the #SeasonalArtChallenge. Use the hashtag to submit an autumn harvest-themed scene for a chance to be featured on the @NVIDIAStudio and @NVIDIAOmniverse social channels.
The fastest way to give the gift of cloud gaming starts this GFN Thursday: For a limited time, every six-month GeForce NOW Ultimate membership includes three months of PC Game Pass.
Also, the newest GeForce NOW app update is rolling out to members, including Xbox Game Syncing and more improvements.
Plus, take advantage of a heroic, new members-only Guild Wars 2 reward. It’s all topped off by support for 18 more games in the GeForce NOW library this week.
Give the Gift of Gaming
Pair PC Game Pass with a GeForce NOW Ultimate bundle for the ultimate gaming gift.
Unwrap the gift of gaming: For a limited time, gamers who sign up for the six-month GeForce NOW Ultimate membership will also receive three free months of PC Game Pass — a $30 value.
With it, Ultimate members can play a collection of high-quality Xbox PC titles with the power of a GeForce RTX 4080 rig in the cloud. Jump into the action in iconic franchises like Age of Empires, DOOM, Forza and more, with support for more titles added every GFN Thursday.
Seamlessly launch supported favorites across nearly any device at up to 4K and 120 frames per second or at up to 240 fps with NVIDIA Reflex technology in supported titles for lowest-latency streaming.
This special offer is only here for a limited time, so upgrade today.
Sync’d Up
Look who just joined the party!
With so many games ready to stream, it might be hard to decide what to play next. The latest GeForce NOW app update, currently rolling out to members, is here to help.
Members can now connect their Xbox accounts to GeForce NOW to sync the games they own to their GeForce NOW library. Game syncing lets members connect their digital game store accounts to GeForce NOW, so all of their supported games are part of their streaming library. Syncing an Xbox account will also add any supported titles a member has access to via PC Game Pass — perfect for members taking advantage of the latest Ultimate bundle.
The new update also adds benefits for Ubisoft+ subscribers. With a linked Ubisoft+ account, members can now launch supported Ubisoft+ games they already own from the GeForce NOW app, and the game will be automatically added to “My Library.” Get more details on Ubisoft account linking.
Version 2.0.58 also includes an expansion of the new game session diagnostic tools to help members ensure they’re streaming at optimal quality. It adds codec information to the in-stream statistics overlay and includes other miscellaneous bug fixes. The update should be available for all members soon.
A Heroic Offering
Rewards fit for a hero.
This week, members can receive Guild Wars 2 “Heroic Edition,” which includes a treasure trove of goodies, such as the base game, Legacy Armor, an 18-slot inventory expansion and four heroic Boosters. It’s the perfect way to jump into ArenaNet’s critically acclaimed, free-to-play, massively multiplayer online role-playing game.
It’s easy to get membership rewards for streaming games on the cloud. Visit the GeForce NOW Rewards portal and update the settings to receive special offers and in-game goodies.
Members can also sign up for the GeForce NOW newsletter, which includes reward notifications, by logging into their NVIDIA account and selecting “Preferences” from the header. Check the “Gaming & Entertainment” box and “GeForce NOW” under topic preferences.
Ready, Set, Go
A new DLC awakens.
The first downloadable content for Gearbox’s Remnant 2 arrives in the cloud. The Awakened King brings a new storyline, area, archetype and more to the dark fantasy co-op shooter — stream it today to experience the awakening of the One True King as he seeks revenge against all who oppose him.
Catch even more action with the 18 newly supported games in the cloud:
The telecommunications industry — the backbone of today’s interconnected world — is valued at a staggering $1.7 trillion globally, according to IDC.
It’s a massive operation, as telcos process hundreds of petabytes of data in their networks each day. That magnitude is only increasing, as the total amount of data transacted globally is forecast to grow to more than 180 zettabytes by 2025.
To meet this demand for data processing and analysis, telcos are turning to generative AI, which is improving efficiency and productivity across industries.
NVIDIA announced an AI foundry service — a collection of NVIDIA AI Foundation Models, NVIDIA NeMo framework and tools, and NVIDIA DGX Cloud AI supercomputing and services — that gives enterprises an end-to-end solution for creating and optimizing custom generative AI models.
Using the AI foundry service, Amdocs, a leading provider of software and services for communications and media providers, will optimize enterprise-grade large language models for the telco and media industries to efficiently deploy generative AI use cases across their businesses, from customer experiences to network operations and provisioning. The LLMs will run on NVIDIA accelerated computing as part of the Amdocs amAIz framework.
The collaboration builds on the previously announced Amdocs-Microsoft partnership, enabling service providers to adopt these applications in secure, trusted environments, including on premises and in the cloud.
Custom Models for Custom Results
While preliminary applications of generative AI used broad datasets, enterprises have become increasingly focused on developing custom models to perform specialized, industry-specific skills.
By training models on proprietary data, telcos can deliver tailored solutions that produce more accurate results for their use cases.
To simplify the development, tuning and deployment of such custom models, Amdocs is integrating the new NVIDIA AI foundry service.
Equipped with these new generative AI capabilities — including guardrail features — service providers can enhance performance, optimize resource utilization and flexibly scale to meet future needs.
Amdocs’ Global Telco Ecosystem Footprint
More than 350 of the world’s leading telecom and media companies across 90 countries take advantage of Amdocs services each day, including 27 of the world’s top 30 service providers, according to OMDIA.(1) Powering more than 1.7 billion daily digital journeys, Amdocs platforms impact more than 3 billion people around the world.
NVIDIA and Amdocs are exploring several generative AI use cases to simplify and improve operations by providing secure, cost-effective, and high-performance generative AI capabilities.
Initial use cases span customer care, including accelerating resolution of customer inquiries by drawing information from across company data.
And in network operations, the companies are exploring ways to generate solutions to address configuration, coverage or performance issues as they arise.
Automotive companies are transforming every phase of their product lifecycle — evolving their primarily physical, manual processes into software-driven, AI-enhanced digital systems.
To help them save costs and reduce lead times, NVIDIA is announcing two new simulation engines on Omniverse Cloud: the virtual factory simulation engine and the autonomous vehicle (AV) simulation engine.
Omniverse Cloud, a platform-as-a-service for developing and deploying applications for industrial digitalization, is hosted on Microsoft Azure. This one-stop shop enables automakers worldwide to unify digitalization across their core product and business processes. It allows enterprises to achieve faster production and more efficient operations, improving time to market and enhancing sustainability initiatives.
For design, engineering and manufacturing teams, digitalization streamlines their work, converting once primarily manual industrial processes into efficient systems for concept and styling; AV development, testing and validation; and factory planning.
Virtual Factory Simulation Engine
The Omniverse Cloud virtual factory simulation engine is a collection of customizable developer applications and services that enable factory planning teams to connect large-scale industrial datasets while collaborating, navigating and reviewing them in real time.
Design teams working with 3D data can assemble virtual factories and share their work with thousands of planners who can view, annotate and update the full-fidelity factory dataset from lightweight devices. By simulating virtual factories on Omniverse Cloud, automakers can increase throughput and production quality while saving years of effort and millions of dollars that would result from making changes once construction is underway.
On Omniverse Cloud, teams can create interoperability between existing software applications such as Autodesk Factory Planning, which supports the entire lifecycle for building, mechanical, electrical, and plumbing and factory lines, as well as Siemens’ NX, Process Simulate and Teamcenter Visualization software and the JT file format. They can share knowledge and data in real time in live, virtual factory reviews across 2D devices or in extended reality.
T-Systems, a leading IT solutions provider for Europe’s largest automotive manufacturers, is building and deploying a custom virtual factory application that its customers can deploy in Omniverse Cloud.
SoftServe, an elite member of the NVIDIA Service Delivery Partner program, is also developing custom factory simulation and visualization solutions on this Omniverse Cloud engine, covering factory design, production planning and control.
AV Simulation Engine
The AV simulation engine is a service that delivers physically based sensor simulation, enabling AV and robotics developers to run autonomous systems in a closed-loop virtual environment.
The next generation of AV architectures will be built on large, unified AI models that combine layers of the vehicle stack, including perception, planning and control. Such new architectures call for an integrated approach to development.
With previous architectures, developers could train and test these layers independently, as they were governed by different models. For example, simulation could be used to develop a vehicle’s planning and control system, which only needs basic information about objects in a scene — such as the speed and distance of surrounding vehicles — while perception networks could be trained and tested on recorded sensor data.
However, using simulation to develop an advanced unified AV architecture requires sensor data as the input. For a simulator to be effective, it must be able to simulate vehicle sensors, such as cameras, radars and lidars, with high fidelity.
To address this challenge, NVIDIA is bringing state-of-the-art sensor simulation pipelines used in DRIVE Sim and Isaac Sim to Omniverse Cloud on Microsoft Azure.
Omniverse Cloud sensor simulation provides AV and robotics workflows with high-fidelity, physically based simulation for cameras, radars, lidars and other types of sensors. It can be connected to existing simulation applications, whether developed in-house or provided by a third party, via Omniverse Cloud application programming interfaces for integration into workflows.
Fast Track to Digitalization
The factory simulation engine is now available to customers via an Omniverse Cloud enterprise private offer through the Azure Marketplace, which provides access to NVIDIA OVX systems and fully managed Omniverse software, reference applications and workflows. The sensor simulation engine is coming soon.
As NVIDIA continues to collaborate with Microsoft to build state-of-the-art AI infrastructure, Microsoft is introducing additional H100-based virtual machines to Microsoft Azure to accelerate demanding AI workloads.
At its Ignite conference in Seattle today, Microsoft announced its new NC H100 v5 VM series for Azure, the industry’s first cloud instances featuring NVIDIA H100 NVL GPUs.
This offering brings together a pair of PCIe-based H100 GPUs connected via NVIDIA NVLink, with nearly 4 petaflops of AI compute and 188GB of faster HBM3 memory. The NVIDIA H100 NVL GPU can deliver up to 12x higher performance on GPT-3 175B over the previous generation and is ideal for inference and mainstream training workloads.
Additionally, Microsoft announced plans to add the NVIDIA H200 Tensor Core GPU to its Azure fleet next year to support larger model inferencing with no increase in latency. This new offering is purpose-built to accelerate the largest AI workloads, including LLMs and generative AI models.
The H200 GPU brings dramatic increases both in memory capacity and bandwidth using the latest-generation HBM3e memory. Compared to the H100, this new GPU will offer 141GB of HBM3e memory (1.8x more) and 4.8 TB/s of peak memory bandwidth (a 1.4x increase).
Cloud Computing Gets Confidential
Further expanding availability of NVIDIA-accelerated generative AI computing for Azure customers, Microsoft announced another NVIDIA-powered instance: the NCC H100 v5.
These Azure confidential VMs with NVIDIA H100 Tensor Core GPUs allow customers to protect the confidentiality and integrity of their data and applications in use, in memory, while accessing the unsurpassed acceleration of H100 GPUs. These GPU-enhanced confidential VMs will be coming soon to private preview.
To learn more about the new confidential VMs with NVIDIA H100 Tensor Core GPUs, and sign up for the preview, read the blog.
Today’s landscape of free, open-source large language models (LLMs) is like an all-you-can-eat buffet for enterprises. This abundance can be overwhelming for developers building custom generative AI applications, as they need to navigate unique project and business requirements, including compatibility, security and the data used to train the models.
NVIDIA AI Foundation Models — a curated collection of enterprise-grade pretrained models — give developers a running start for bringing custom generative AI to their enterprise applications.
NVIDIA-Optimized Foundation Models Speed Up Innovation
NVIDIA AI Foundation Models can be experienced through a simple user interface or API, directly from a browser. Additionally, these models can be accessed from NVIDIA AI Foundation Endpoints to test model performance from within their enterprise applications.
Available models include leading community models such as Llama 2, Stable Diffusion XL and Mistral, which are formatted to help developers streamline customization with proprietary data. Additionally, models have been optimized with NVIDIA TensorRT-LLM to deliver the highest throughput and lowest latency and to run at scale on any NVIDIA GPU-accelerated stack. For instance, the Llama 2 model optimized with TensorRT-LLM runs nearly 2x faster on NVIDIA H100.
The new NVIDIA family of Nemotron-3 8B foundation models supports the creation of today’s most advanced enterprise chat and Q&A applications for a broad range of industries, including healthcare, telecommunications and financial services.
The models are a starting point for customers building secure, production-ready generative AI applications, are trained on responsibly sourced datasets and operate at comparable performance to much larger models. This makes them ideal for enterprise deployments.
Multilingual capabilities are a key differentiator of the Nemotron-3 8B models. Out of the box, the models are proficient in over 50 languages, including English, German, Russian, Spanish, French, Japanese, Chinese, Korean, Italian and Dutch.
Fast-Track Customization to Deployment
Enterprises leveraging generative AI across business functions need an AI foundry to customize models for their unique applications. NVIDIA’s AI foundry features three elements — NVIDIA AI Foundation Models, NVIDIA NeMo framework and tools, and NVIDIA DGX Cloud AI supercomputing services. Together, these provide an end-to-end enterprise offering for creating custom generative AI models.
Importantly, enterprises own their customized models and can deploy them virtually anywhere on accelerated computing with enterprise-grade security, stability and support using NVIDIA AI Enterprise software.
To understand the latest advance in generative AI, imagine a courtroom.
Judges hear and decide cases based on their general understanding of the law. Sometimes a case — like a malpractice suit or a labor dispute — requires special expertise, so judges send court clerks to a law library, looking for precedents and specific cases they can cite.
Like a good judge, large language models (LLMs) can respond to a wide variety of human queries. But to deliver authoritative answers that cite sources, the model needs an assistant to do some research.
The court clerk of AI is a process called retrieval-augmented generation, or RAG for short.
The Story of the Name
Patrick Lewis, lead author of the 2020 paper that coined the term, apologized for the unflattering acronym that now describes a growing family of methods across hundreds of papers and dozens of commercial services he believes represent the future of generative AI.
Patrick Lewis
“We definitely would have put more thought into the name had we known our work would become so widespread,” Lewis said in an interview from Singapore, where he was sharing his ideas with a regional conference of database developers.
“We always planned to have a nicer sounding name, but when it came time to write the paper, no one had a better idea,” said Lewis, who now leads a RAG team at AI startup Cohere.
So, What Is Retrieval-Augmented Generation?
Retrieval-augmented generation is a technique for enhancing the accuracy and reliability of generative AI models with facts fetched from external sources.
In other words, it fills a gap in how LLMs work. Under the hood, LLMs are neural networks, typically measured by how many parameters they contain. An LLM’s parameters essentially represent the general patterns of how humans use words to form sentences.
That deep understanding, sometimes called parameterized knowledge, makes LLMs useful in responding to general prompts at light speed. However, it does not serve users who want a deeper dive into a current or more specific topic.
Combining Internal, External Resources
Lewis and colleagues developed retrieval-augmented generation to link generative AI services to external resources, especially ones rich in the latest technical details.
The paper, with coauthors from the former Facebook AI Research (now Meta AI), University College London and New York University, called RAG “a general-purpose fine-tuning recipe” because it can be used by nearly any LLM to connect with practically any external resource.
Building User Trust
Retrieval-augmented generation gives models sources they can cite, like footnotes in a research paper, so users can check any claims. That builds trust.
What’s more, the technique can help models clear up ambiguity in a user query. It also reduces the possibility a model will make a wrong guess, a phenomenon sometimes called hallucination.
Another great advantage of RAG is it’s relatively easy. A blog by Lewis and three of the paper’s coauthors said developers can implement the process with as few as five lines of code.
That makes the method faster and less expensive than retraining a model with additional datasets. And it lets users hot-swap new sources on the fly.
How People Are Using Retrieval-Augmented Generation
With retrieval-augmented generation, users can essentially have conversations with data repositories, opening up new kinds of experiences. This means the applications for RAG could be multiple times the number of available datasets.
For example, a generative AI model supplemented with a medical index could be a great assistant for a doctor or nurse. Financial analysts would benefit from an assistant linked to market data.
In fact, almost any business can turn its technical or policy manuals, videos or logs into resources called knowledge bases that can enhance LLMs. These sources can enable use cases such as customer or field support, employee training and developer productivity.
The broad potential is why companies including AWS, IBM, Glean, Google, Microsoft, NVIDIA, Oracle and Pinecone are adopting RAG.
Getting Started With Retrieval-Augmented Generation
The software components are all part of NVIDIA AI Enterprise, a software platform that accelerates development and deployment of production-ready AI with the security, support and stability businesses need.
Getting the best performance for RAG workflows requires massive amounts of memory and compute to move and process data. The NVIDIA GH200 Grace Hopper Superchip, with its 288GB of fast HBM3e memory and 8 petaflops of compute, is ideal — it can deliver a 150x speedup over using a CPU.
Once companies get familiar with RAG, they can combine a variety of off-the-shelf or custom LLMs with internal or external knowledge bases to create a wide range of assistants that help their employees and customers.
RAG doesn’t require a data center. LLMs are debuting on Windows PCs, thanks to NVIDIA software that enables all sorts of applications users can access even on their laptops.
An example application for RAG on a PC.
PCs equipped with NVIDIA RTX GPUs can now run some AI models locally. By using RAG on a PC, users can link to a private knowledge source – whether that be emails, notes or articles – to improve responses. The user can then feel confident that their data source, prompts and response all remain private and secure.
A recent blog provides an example of RAG accelerated by TensorRT-LLM for Windows to get better results fast.
The History of Retrieval-Augmented Generation
The roots of the technique go back at least to the early 1970s. That’s when researchers in information retrieval prototyped what they called question-answering systems, apps that use natural language processing (NLP) to access text, initially in narrow topics such as baseball.
The concepts behind this kind of text mining have remained fairly constant over the years. But the machine learning engines driving them have grown significantly, increasing their usefulness and popularity.
In the mid-1990s, the Ask Jeeves service, now Ask.com, popularized question answering with its mascot of a well-dressed valet. IBM’s Watson became a TV celebrity in 2011 when it handily beat two human champions on the Jeopardy! game show.
Today, LLMs are taking question-answering systems to a whole new level.
Insights From a London Lab
The seminal 2020 paper arrived as Lewis was pursuing a doctorate in NLP at University College London and working for Meta at a new London AI lab. The team was searching for ways to pack more knowledge into an LLM’s parameters and using a benchmark it developed to measure its progress.
Building on earlier methods and inspired by a paper from Google researchers, the group “had this compelling vision of a trained system that had a retrieval index in the middle of it, so it could learn and generate any text output you wanted,” Lewis recalled.
The IBM Watson question-answering system became a celebrity when it won big on the TV game show Jeopardy!
When Lewis plugged into the work in progress a promising retrieval system from another Meta team, the first results were unexpectedly impressive.
“I showed my supervisor and he said, ‘Whoa, take the win. This sort of thing doesn’t happen very often,’ because these workflows can be hard to set up correctly the first time,” he said.
Lewis also credits major contributions from team members Ethan Perez and Douwe Kiela, then of New York University and Facebook AI Research, respectively.
When complete, the work, which ran on a cluster of NVIDIA GPUs, showed how to make generative AI models more authoritative and trustworthy. It’s since been cited by hundreds of papers that amplified and extended the concepts in what continues to be an active area of research.
When users ask an LLM a question, the AI model sends the query to another model that converts it into a numeric format so machines can read it. The numeric version of the query is sometimes called an embedding or a vector.
Retrieval-augmented generation combines LLMs with embedding models and vector databases.
The embedding model then compares these numeric values to vectors in a machine-readable index of an available knowledge base. When it finds a match or multiple matches, it retrieves the related data, converts it to human-readable words and passes it back to the LLM.
Finally, the LLM combines the retrieved words and its own response to the query into a final answer it presents to the user, potentially citing sources the embedding model found.
Keeping Sources Current
In the background, the embedding model continuously creates and updates machine-readable indices, sometimes called vector databases, for new and updated knowledge bases as they become available.
Retrieval-augmented generation combines LLMs with embedding models and vector databases.
Many developers find LangChain, an open-source library, can be particularly useful in chaining together LLMs, embedding models and knowledge bases. NVIDIA uses LangChain in its reference architecture for retrieval-augmented generation.
Looking forward, the future of generative AI lies in creatively chaining all sorts of LLMs and knowledge bases together to create new kinds of assistants that deliver authoritative results users can verify.
Get a hands on using retrieval-augmented generation with an AI chatbot in this NVIDIA LaunchPad lab.
Artificial intelligence on Windows 11 PCs marks a pivotal moment in tech history, revolutionizing experiences for gamers, creators, streamers, office workers, students and even casual PC users.
It offers unprecedented opportunities to enhance productivity for users of the more than 100 million Windows PCs and workstations that are powered by RTX GPUs. And NVIDIA RTX technology is making it even easier for developers to create AI applications to change the way people use computers.
New optimizations, models and resources announced at Microsoft Ignite will help developers deliver new end-user experiences, quicker.
An upcoming update to TensorRT-LLM — open-source software that increases AI inference performance — will add support for new large language models and make demanding AI workloads more accessible on desktops and laptops with RTX GPUs starting at 8GB of VRAM.
TensorRT-LLM for Windows will soon be compatible with OpenAI’s popular Chat API through a new wrapper. This will enable hundreds of developer projects and applications to run locally on a PC with RTX, instead of in the cloud — so users can keep private and proprietary data on Windows 11 PCs.
Custom generative AI requires time and energy to maintain projects. The process can become incredibly complex and time-consuming, especially when trying to collaborate and deploy across multiple environments and platforms.
AI Workbench is a unified, easy-to-use toolkit that allows developers to quickly create, test and customize pretrained generative AI models and LLMs on a PC or workstation. It provides developers a single platform to organize their AI projects and tune models to specific use cases.
This enables seamless collaboration and deployment for developers to create cost-effective, scalable generative AI models quickly. Join the early access list to be among the first to gain access to this growing initiative and to receive future updates.
To support AI developers, NVIDIA and Microsoft will release are releasing DirectML enhancements to accelerate one two of the most popular foundational AI models,: Llama 2 and Stable Diffusion. Developers now have more options for cross-vendor deployment, in addition to setting a new standard for performance.
Portable AI
Last month, NVIDIA announced TensorRT-LLM for Windows, a library for accelerating LLM inference.
The next TensorRT-LLM release, v0.6.0 coming later this month, will bring improved inference performance — up to 5x faster — and enable support for additional popular LLMs, including the new Mistral 7B and Nemotron-3 8B. Versions of these LLMs will run on any GeForce RTX 30 Series and 40 Series GPU with 8GB of RAM or more, making fast, accurate, local LLM capabilities accessible even in some of the most portable Windows devices.
Up to 5X performance with the new TensorRT-LLM v0.6.0.
The new release of TensorRT-LLM will be available for install on the /NVIDIA/TensorRT-LLM GitHub repo. New optimized models will be available on ngc.nvidia.com.
Conversing With Confidence
Developers and enthusiasts worldwide use OpenAI’s Chat API for a wide range of applications — from summarizing web content and drafting documents and emails to analyzing and visualizing data and creating presentations.
One challenge with such cloud-based AIs is that they require users to upload their input data, making them impractical for private or proprietary data or for working with large datasets.
To address this challenge, NVIDIA is soon enabling TensorRT-LLM for Windows to offer a similar API interface to OpenAI’s widely popular ChatAPI, through a new wrapper, offering a similar workflow to developers whether they are designing models and applications to run locally on a PC with RTX or in the cloud. By changing just one or two lines of code, hundreds of AI-powered developer projects and applications can now benefit from fast, local AI. Users can keep their data on their PCs and not worry about uploading datasets to the cloud.
Perhaps the best part is that many of these projects and applications are open source, making it easy for developers to leverage and extend their capabilities to fuel the adoption of generative AI on Windows, powered by RTX.
The wrapper will work with any LLM that’s been optimized for TensorRT-LLM (for example, Llama 2, Mistral and NV LLM) and is being released as a reference project on GitHub, alongside other developer resources for working with LLMs on RTX.
Model Acceleration
Developers can now leverage cutting-edge AI models and deploy with a cross-vendor API. As part of an ongoing commitment to empower developers, NVIDIA and Microsoft have been working together to accelerate Llama on RTX via the DirectML API.
Building on the announcements for the fastest inference performance for these models announced last month, this new option for cross-vendor deployment makes it easier than ever to bring AI capabilities to PC.
Developers and enthusiasts can experience the latest optimizations by downloading the latest ONNX runtime and following the installation instructions from Microsoft, and installing the latest driver from NVIDIA, which will be available on Nov. 21.
These new optimizations, models and resources will accelerate the development and deployment of AI features and applications to the 100 million RTX PCs worldwide, joining the more than 400 partners shipping AI-powered apps and games already accelerated by RTX GPUs.
As models become even more accessible and developers bring more generative AI-powered functionality to RTX-powered Windows PCs, RTX GPUs will be critical for enabling users to take advantage of this powerful technology.
Osinski’s panoramic view from the 20th floor terrace in the Upper East Side
The making of Sir Wade’s VFX robot animation
Access to default prim options in Maya UI
NVIDIA GeForce NOW (@NVIDIAGFN) 
