Aurelien Grosdidier’s Post

When CEOs from medium enterprises seek my counsel on #generativeAI, I will tell them that hallucinations are features and not bugs within the system, but we should not let them roam free because it might create unintended consequences both good and bad. Of course, this interesting research paper on #AI with a catchy title clearly explains why this is the case. Do read if you are as academic as I am. Hallucination is Inevitable: An Innate Limitation of Large Language Models https://lnkd.in/gUbPKq7N

A must read paper, to understand the problem of AI hallucinations and the potential implications. “Hallucination has been widely recognized to be a significant drawback for large language models (LLMs). There have been many works that attempt to reduce the extent of hallucination. These efforts have mostly been empirical so far, which cannot answer the fundamental question whether it can be completely eliminated. In this paper, we formalize the problem and show that it is impossible to eliminate hallucination in LLMs.” #artificialintelligence #aichallenges #ainews #airegulation #euronews https://lnkd.in/dmt_JBAr

Hallucination is Inevitable - An Innate Limitation of Large Language Models New research by Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli explores fundamental limitations of Large Language Models (LLMs): hallucination. The study reveals that regardless of architecture or training data, LLMs will inevitably generate factually incorrect or nonsensical outputs. The researchers' theoretical framework defines hallucination and proves that any computable LLM will hallucinate with respect to some ground truth function. Key findings include that LLMs will hallucinate on infinitely many inputs, emphasizing the impossibility of completely eliminating this issue. Empirical studies on state-of-the-art LLMs validate these theoretical results, highlighting failures on tasks predicted to induce hallucination. The implications are significant, calling for further research on the safety boundaries of LLMs and the development of appropriate guardrails. Read the full paper for more insights: https://lnkd.in/ey_Q_rPB #LLM #MachineLearning #AI #Hallucination #TheoryOfComputation

Hallucination is Inevitable: An Innate Limitation of Large Language Models Ziwei Xu et al (National University of Singapore) on Arxiv: Nice read: https://lnkd.in/gBayqFBi Hallucination has been widely recognized to be a significant drawback for large language models (LLMs). There have been many works that attempt to reduce the extent of hallucination. These efforts have mostly been empirical so far, which cannot answer the fundamental question whether it can be completely eliminated. In this paper, Arxiv formalize the problem and show that it is impossible to eliminate hallucination in LLMs. Specifically, we define a formal world where hallucination is defined as inconsistencies between a computable LLM and a computable ground truth function. By employing results from learning theory, they show that LLMs cannot learn all of the computable functions and will therefore always hallucinate. Since the formal world is a part of the real world, which is much more complicated, hallucinations are also inevitable for real world LLMs. Furthermore, for real world LLMs constrained by provable time complexity, they describe the hallucination-prone tasks and empirically validate our claims. Finally, using the formal world framework, they discuss the possible mechanisms and efficacies of existing hallucination mitigators as well as the practical implications on the safe deployment of LLMs. #llm #hallucinations #genai #genaitrends #arxiv

Maximizing Efficiency: The KIVI Approach to Memory Optimization in Large Language Models #accuracy #AI #AItechnology #artificialintelligence #developers #Finetuning #keyvaluecaches #KIVI #Largelanguagemodels #llm #machinelearning #memoryoptimization #memoryusage #performance #quantizationalgorithm #realworldscenarios #researchers #Scalability #throughputimprovements

Researchers have put together a hallucination leader board for popular large language models based on summarization of input https://lnkd.in/eYR_aJGB.  #generativeai #largelanguagemodel #hallucinations #leaderboard #artificialintelligence

"Our study found that large language models can predict how linguistic information is encoded in the human brain, providing a new tool to interpret human brain activity. The similarity between the human brain’s and the large language model’s linguistic code has enabled us, for the first time, to track how information in the speaker’s brain is encoded into words and transferred, word by word, to the listener’s brain during face-to-face conversations." https://lnkd.in/eYF_7XAz

Being able to interpret an #ML model’s hidden representations is key to understanding its behavior. Introducing Patchscopes, an approach that trains #LLMs to provide natural language explanations of their own hidden representations. Paper: https://lnkd.in/gR_4VwS7

There's been a growing focus on effective benchmarking methods for large language models. Came across an interesting study introducing a new approach that utilizes LLM debates to assess model performance. The method involves creating an LLM debate framework and using an LLM judge that analyzes debate arguments based on criteria like clarity, factual accuracy, counter-arguments, and persuasiveness. This eliminates the need for human evaluation, making it more scalable for LLM development and evaluation. Could this be the step towards LLMs acting as judges in the future? Are we ready for AI referees? 🙂 #AI #machinelearning #LLMs #naturallanguageprocessing https://lnkd.in/g7A8pEUW

How do we go about evaluating Large Language Models? Most approaches can be categorized into four groups: 🎯 Task-Specific Metrics 🔬 Research Benchmarks 🤖 LLM Self-Evaluation 👤 Human Evaluation Read more from Michał Oleszak's post: