How to Mitigate Gen AI Hallucinations, Bias & Intellectual Property Risk in LLMs - Aug. 2023

*** Overview ***

Vectara CEO/Founder/Gamer Amr Awadallah presents at Stanford University (Aug. 2023) and shares the current state-of-the-art in Large Language Models (LLMs), benefits and challenges of fine-tuning LLMs, including hallucinations in LLMs, and how Grounded Generation is a new alternative that mitigates hallucinations.

*** Jump to Relevant Section ***

01:23 - The Unprecedented Growth of ChatGPT (Chart) & Gen AI

05:05 - Belief: Generative AI in every application in 5 Years
(Search Engine to Answer Engine to Action Engine)

06:55 - What is Vectara (Architecture Diagram) & How it Solves AI Hallucinations

08:05 - Demo Applications Using Vectara (see links to demos below)

11:42 - Grounded Generation vs. LLM Fine-tuning to Solve Hallucinations

12:16 - What is LLM Fine-tuning? Training on Public and Private Data

13:05 - Negative Training Your Model

13:25 - Problems with Fine-Tuning LLM Models #1: Expensive & Time-consuming

13:49 - Problems with Fine-Tuning LLM Models #2: Requires Training on YOUR Data

13:55 - "Model Poisoning" Definition

14:35 - Problems with Fine-Tuning LLM Models #3: It Cannot Cite Sources

14:49 - Problems with Fine-Tuning LLM Models #4: It Compresses Data Massively, Resulting in Hallucinations

16:02 - Problems with Fine-Tuning LLM Models #5: Copyright or Biased Information in the LLM Will Influence Outputs

17:37 - The Solution to Hallucinations: "Grounded Generation" (Note: to date, nothing 100% eliminates hallucinations, but LLM builder companies like Vectara, Google, and more are working on it)

17:55 - Information Retrieval ML Models Explanation

18:25 - How Grounded Generation (GG) Works

19:40 - Prompt Engineering to Sandbox the Question (Distilling GG down to steps):

1. Only use information from the provided facts
2. Cite the facts in citations
3. Only cite the relevant facts
4. If the facts are of low relevance, say "I don't know"

20:18 - Main Methods of Grounded Generation

21:57 - Vectara Generative AI Use Cases

22:06 - How to Use Vectara for Free (QR Code - Scan Your Screen w/ Smartphone or Use the Link Below)

22:55 - Audience Question: "Fine-tuning and even GG won't get hallucination to 0%, so what will?"

*** What is Vectara ***

Vectara is a platform for building GenAI applications that scale. It provides an easy-to-use API for document indexing (including OCR) and querying that is managed by Vectara and is optimized for performance and accuracy.

*** Learn about "Grounded Generation" (AKA Retrieval Augmented Generation) ***

** ALL DEMOS **

*** OTHER SAMPLE APPS ***

Experience the power of Generative AI, LLMs, and best-in-class retrieval/hybrid search via Vectara sample apps:

** Ask LangChain Docs **

LangChain's doc repository is large. Developers can search by asking questions in natural human language and get answers, not results:

** ASK FEYNMAN ** - Ask renowned CalTech physics instructor any physics question from the corpus from his newly released archive:

** ASK NEWS ** - Ask questions about recent news events and get AI-generated answer summaries to your questions via AskNews:

** ASK HBS.edu ** - Ask any question about the Harvard Business School from HSB.edu and get AI-generated answer summaries to your questions via AskHBS:

** ASK UCLA.edu ** - Ask any question about UCLA admissions and get AI-generated answer summaries to your questions via AskUCLA:

** Try Vectara for FREE **

Most projects are covered by our Free Growth plan (50 MB, 15k queries/mo).

** DEVELOPER Resources **

Join our Discord

Read our API documentation

** Vectara Product Links **

** Social Channels **

** LLM Learning Resources **

Learn about Vectara’s “Grounded Generation” product release (5/30/2023)

Learn about LLM “Hallucinations”

[Dwedari]

Amazing work, a worker team like you should get more and more support.. hopefully hope you get more support and success

[rawanabdrabu]

Great idea! But then it’s restricted that each customer should provide or have their own data prepared to retrieve information for their own use case!