How do you keep large language models in line?
In this InTechnology video, Camille talks with Sanjay Rajagopalan, Chief Design and Strategy Officer at Vianai Systems. They get into how LLMs operate, uses for LLMs in enterprise spaces, and how to prevent LLMs from going astray.
Understanding How LLMs Operate
Camille and Sanjay delve into the mechanisms of large language models, focusing on their strong suits and limitations. Sanjay conveys that while users often find LLMs impressive in the initial 15 minutes of interaction, they soon recognize the models’ limitations and inaccuracies. He labels these inaccuracies as “hallucinations,” a topic that is explored further in the conversation. Overall, Sanjay posits that LLMs serve as a useful starting point and can guide users close to the desired answer. However, their capabilities are limited to their training, and introducing a randomness factor may lead to errors, even though it can also render responses more conversational and natural-sounding. Tweaking elements like randomness through parameters and maintaining alignment with techniques such as reinforcement learning from human feedback (RLHF) are essential for regulating LLMs.
LLMs in the Enterprise Organizations
Sanjay highlights the conversational UI as one of the advantages of LLMs, making interaction with language models more user-friendly across various levels of an enterprise organization. LLMs are adept at certain tasks such as code generation and excel at qualitative comparison and navigating extensive textual databases. He cites an instance where an LLM is employed to align discounts with client contracts, revealing potential matches that might have otherwise remained undiscovered.
Preventing LLMs from Going Astray
Revisiting the subjects of hallucinations and LLM regulation, Camille and Sanjay explore various strategies to prevent models from deviating. Techniques such as prompt engineering, which involves refining prompts before they are input into a language model, and output processing to eliminate any contentious elements or append disclaimers about potential inaccuracies, are discussed. Enterprises can implement a combination of automated software methods and human supervision to ensure the models remain on track. Furthermore, Sanjay advises utilizing diverse LLMs specialized for distinct tasks instead of depending solely on a singular model.
Sanjay Rajagopalan, Chief Design and Strategy Officer at Vianai Systems
Since 2019, Sanjay Rajagopalan has held the position of Chief Design and Strategy Officer at Vianai Systems, a startup specializing in providing an enterprise AI platform and AI solutions. Before this, he served as the SVP and Head of Design and Research at Infosys and has fulfilled several other leadership roles in the tech sector. With a fervor for technology and business, Sanjay is a distinguished leader in the realms of design, innovation, and technology strategy. He earned his Ph.D. from Stanford University and an M.S. from The University of Texas at Austin, with both degrees in Mechanical Engineering.
Check it out. For more information, previous podcasts, and full versions, visit our homepage.
To read more about cybersecurity topics, visit our blog.
#LLM #largelanguagemodels #languagemodels
The views and opinions expressed are those of the guests and author and do not necessarily reflect the official policy or position of Intel Corporation.
—–
If you are interested in emerging threats, new technologies, or best tips and practices in cybersecurity, please follow the InTechnology podcast on your favorite podcast platforms: Apple Podcast and Spotify.
Follow our hosts Tom Garrison @tommgarrison and Camille @morhardt.
Learn more about Intel Cybersecurity and Intel Compute Life Cycle (CLA).