Chain-of-Verification: How to fight AI Hallucination

[@jmirodg7094]

Excellent, Happy to see that Bert is not dead and still manage to be useful

[@ytpah9823]

🎯 Key Takeaways for quick navigation: 00:00 📚 The video discusses an MIT Technology Review article and examines an AI solution to combat hallucinations. 00:53 🚀 The host encourages viewers to challenge traditional research and come up with innovative solutions. 01:21 🤯 Discusses how interacting with AI can sometimes lead to unexpected or surreal answers, described as "AI hallucinations". 02:14 🤔 The importance of trust in AI systems is highlighted, mentioning issues of "digital deception". 03:05 ⚠️ MIT Technology Review mentions the problem of AI language models making things up, which poses a security and privacy concern. 03:59 🤓 Explains how auto-aggressive large language models, like GPT-4, predict the next word in a sentence. 05:24 🔍 Introduces a survey of hallucination in NLP from April 2022 to understand the challenges better. 06:21 📖 Talks about the chain of verification methodology developed by META AI and ETH Zurich to reduce hallucinations in LLMs. 07:41 🤖 Uses the GPT-4 system to demonstrate METE's four methodologies to combat hallucinations: the joint method, the two-step method, the factored method, and the factor and revised method. 08:25 🔄 Explains the two-step method where GPT-4 generates verification questions to fact-check information, acting as a proof run. 10:37 🧠 Describes the factored method, where each verification question is presented in a separate prompt to GPT-4. 11:32 🎉 Introduces the factor plus the revised method, which combines the previously discussed methods with an added cross-check prompt. 11:54 🔄 Cross-checking GPT-4's responses helps in reducing the probability of hallucination by distributing risk across multiple prompts. 12:48 🔬 The latest AI research focuses on preventing hallucinations, with METAI introducing ways to do so. 13:12 💡 Suggestions to enhance METAI's method include hierarchical prompting and increased complexity in verification questions. 14:02 🎚️ Hierarchical prompting begins with basic questions and advances to more complex, nuanced queries. 14:28 🤔 Conditional prompting guides the model's reasoning process by asking in-depth, contextual questions. 15:43 📊 Users can request GPT-4 to provide a confidence interval for its responses. 16:08 😅 There's a humorous note on the irony if GPT-4 hallucinates while giving a confidence score. 16:37 🤖 The idea of a multi-modal collaboration involves connecting various AI systems to validate information. 17:05 🧩 Ensemble verification uses multiple AI models to mitigate the risk of a single model's bias or error. 18:00 🔍 Utilizing domain-specific models for verification allows for specialized, tailored accuracy. 19:14 🖼️ Multi-modal systems offer benefits as some information is better represented visually. 20:04 💻 Critique of Bing's performance on the OpenAI platform versus its native browser. 21:27 🌐 WebPilot is recommended as a reliable tool to connect GPT-4 to the internet for validation. 22:22 🧪 An expert system, updated with the latest data, aids in the scientific domain for verification. 23:07 📊 Differentiating between GPT-4's autoregressive model and a sentence transformer system for validation. 24:27 📜 Sentence embeddings in domain-specific knowledge allow for semantic topic comparison and clustering. 24:49 🔍 Using SBERT, the speaker can extract top sentences from over 800 publications on specific quantum field theory subtopics. 25:14 🎥 The speaker has over 50 videos on pre-training and fine-tuning the SBERT system for encoding sentences, paragraphs, pages, and even images. 25:44 📚 The system encodes existing scientific literature for comparison with GPT-4 answers. 26:10 🔢 The system operates in a vector space, allowing for mathematical semantic comparison and topic clustering. 26:33 🌐 The speaker's approach connects GPT-4 to other AI systems, including local ones and web verifiers like Webpilot. 27:00 💡 Instead of relying on expensive databases, the speaker built a sentence transformer system trained on domain-specific literature. 27:29 🤝 An ensemble of specialized models is used to compare GPT-4 generated answers to trusted sources. 27:54 📉 This system significantly reduces the risk of AI hallucinations by comparing with trusted sources.

[@VaibhavPatil-rx7pc]

Excellent information thanks

[@BoominGame]

I would like to point out that hallucinations are never syntactic but semantic.

[@BoominGame]

@@RoulDukeGonzo It doesn't screw up the grammar, or rarely, but the concepts.

[@BoominGame]

@@RoulDukeGonzo We already got the grammar somehow right in the mid 90ies, Trados would not miss a beat, but it was just a normal database in the back not a neuronal network, so you needed a human to supervise the semantics and it was subject -specific. Worked reasonably well for technical and legal.

[@sumitmamoria]

At 9:00, Model output says that it needs to test for 3 laws. What happens when the model get this wrong? How do I fact check my fact-checker? Do I trust the LLM generated fact-check questions? This is also a real problem that I have faced myself while trying to build production ready applications. Isnt it like "trusting a ruler to correctly measure its own length?". The fact checker system needs to be significantly more advanced than the system generating the facts. No?

[@empyrionin]

You are highly limited. I guess a check for logic consistency of the statements, a check for common sense truths... Ultimately even a human "baseline" fact checker can be wrong. The problem is a lot harder to solve and I don't think we have a thorough solution yet.

[@TheRebelDane]

beautiful!

[@loicbaconnier9150]

Really well done thanks

[@moh.syafrianabie8899]

8:00

[@MeanChefNe]

Yep

[@ghostwhowalks2324]

The author has knowledge but needs to work on his teaching skills.