DIREC - Digital Research Centre Denmark’s Post

AI Talk Reflection From the AI Talk Live on Facebook that i recently attended on 13 November 2024, which was last week, the speaker covered a wide range of topics of AI such as the using of ChatGPT, other AI tools such as AI Generating Music, NoteBookLM, WolframAlpha and many more. The speaker demonstrated on how ChatGPT prompted, various techniques, the use case from ChatGPT for students and other features from ChatGPT. One of the keypoints discussed during the talk was underlying technology of ChatGPT, specifically the use of deep learning and natural language and processing. The speaker explained how the model is trained on vast amounts of text data, enabling it to understand and generate human-like text. However, the speaker also touched on some the challenges and limitations of ChatGPT, including issues like ensuring accuracy, preventing misuse and addressing biases in AI-generated responses. Overall attending the AI talk about ChatGPT has broadened my understanding of AI models and motivated me to explore futher more about the impact of AI in the modern world. Well, that's all from me, thank you for taking your time to read my reflection! Siti Kudnie Sahari

AI hallucinations can’t be stopped — but these techniques can limit their damage Developers have tricks to stop artificial intelligence from making things up, but large language models are still struggling to tell the truth, the whole truth and nothing but the truth. By Nicola Jones https://lnkd.in/gRviRPuT

“Automating public policy: a comparative study of conversational artificial intelligence models and human expertise in crafting briefing notes” This paper investigates the application of artificial intelligence (AI) language models in writing policy briefing notes within the context of public administration by juxtaposing the technologies’ performance against the traditional reliance on human expertise. Using a structured evaluation-by-experts methodology, the research scrutinizes and compares the output of three leading AI language models—OpenAI’s ChatGPT, Google’s Gemini, and Mistral’s Le Chat—across ten critical dimensions reflective of policy briefing notes quality. https://lnkd.in/eGX3jXn3 #policycapacity #publicadministration #technology

"As artificial intelligence gets better at giving humans what they want, it also could get better at giving malicious humans what they want. That's one of the concerns driving new research by University of Delaware researchers, published in March in the journal Scientific Reports. Xiao Fang, professor of management information systems and JPMorgan Chase Senior Fellow at the Alfred Lerner College of Business and Economics, and Ming Zhao, associate professor of operations management, collaborated with Minjia Mao, a doctoral student in UD's the Financial Services Analytics (FSAN) program, and researchers Hongzhe Zhang and Xiaohang Zhao, who are alumni of the FSAN program. Specifically, they were interested in whether AI large language models, like the groundbreaking and popular ChatGPT, would produce biased content toward certain groups of people." #biases #ml #ai #chatbots

Are #LLMs like ChatGPT limited to mimicking human language, or do they have the capacity for non-verbal reasoning? While LLMs are trained on text data to be predictive, researchers are exploring whether these models can tackle reasoning tasks typically unrelated to language—like spatial puzzles or abstract pattern recognition. This article from Ars Technica says there are suggestions that by using what’s called “latent space” outside the logical framework of language, LLMs might exhibit reasoning capabilities beyond what they were initially designed for - and explore multiple next steps simultaneously. As we explore these frontiers, it’s clear that the potential for AI to transcend its training is exciting… and maybe even a bit daunting. #ai #genai #tech #llm #machinelearning https://lnkd.in/eREcwesy

Nature writes "The AI revolution is running out of data. What can researchers do? - AI developers are rapidly picking the Internet clean to train large language models such as those behind ChatGPT. Here’s how they are trying to get around the problem." https://lnkd.in/eMjkNfrz. #artificialintelligence #trainingdata #shortage #alternatives #chatgpt #researchers #natures

“AI hallucination” is becoming a cause for concern, as generative AI (Gen-AI) is increasingly being used in real-world applications. In the pursuit of creating human-like chatbots, Large Language Models (LLMs) are generating fabricated scientific theories and false historical events. The root cause of this is, the LLMs are being trained on inaccurate datasets. As these models are trained on vast amounts of human-generated text, currently they are increasingly rely on synthetic data which are produced by those AI models themselves. As the quality of AI-generated content depends heavily on the quality of training data, the indiscriminate use of synthetic data is introducing errors into the resulting models. To make sure “hallucination “, does not derail the progress of Gen-AI, researchers are developing programs to identify these inaccuracies within the generated output. #AI #ArtificialIntelligence https://lnkd.in/gum66NuH

IS YOUR AI HALLUCINATING? NEW APPROACH CAN TELL WHEN CHATBOTS MAKE THINGS UP Discover how researchers are tackling AI hallucinations in large language models with new methods to detect false responses. Learn about the implications for AI applications in healthcare and education. https://lnkd.in/dgk6VcAW

AI Language Models Struggle with Precise Word Count I recently performed a simple experiment with four AI LLM-based systems. The challenge was to generate phrases with specific word counts corresponding to prime numbers in ascending order. All AIs have consistently failed the challenge. I also asked the AI possible reasons for this error, the implications for AI development, and strategies to overcome this limitation. You can read more about the generated answers here: https://lnkd.in/dt_PZ7-H This serves as a reminder that while AI has made remarkable progress, it still has specific limitations that need to be addressed for more comprehensive language processing capabilities.

**The Rise of deepSeek: A Powerful AI with a Catch** deepSeek has quickly emerged this week as one of the most advanced open-weight AI models, pushing the boundaries of what's possible in natural language processing. With impressive capabilities, it’s gaining attention as a serious competitor to current AI models. Its performance, particularly in coding and multilingual tasks, is already turning heads. Especially considering the budget used to develop it (6mil$?) and its open source platform (free to collaborate), which really commodifies the Ai large language models like chatGPT. But there’s a catch. deepSeek is developed in China, meaning it operates within the constraints of the country’s strict censorship policies. While the technology itself is impressive, users must be aware that topics deemed sensitive by Chinese authorities may be filtered or restricted. This raises concerns about intellectual freedom, bias, and the broader implications of relying on AI models developed under government oversight. I tried a simple task and I asked deepseek two questions and I got two very different answers: #1 What is wrong with Donald Trump? #2 What is wrong with Xi Jinping? See video below. What do you think? Can a model with censorship limitations still be globally competitive? Let’s discuss. #ai #deepseek #censorship