Benjamin Todd’s Post

MBZUAI has contributed to one of the most cited AI papers of 2023, according to Zeta Alpha, which recently compiled a list of the 100 most cited AI papers of 2023: https://lnkd.in/gGnuaQ-4 Citations are critical to determine which innovations influence the future of AI and the most cited research papers offer a glimpse into what lies ahead, as they present foundational research from which peers build up and advance progress in the field: https://lnkd.in/d4i_FXjK The paper MBZUAI contributed to is titled, “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena". This paper explores using powerful LLMs as judges to evaluate other models, looking at challenges like biases (position, verbosity, self-enhancement) and limited reasoning. Co-authored by Professor Eric Xing, the paper also suggests ways to reduce these issues and shows how new benchmarks complement traditional ones by testing various LLaMA and Vicuna models: https://lnkd.in/eSB_jcdF #MBZUAI #LLM #AI

The most intriguing part of my research on AI has been the role that emotion plays in how we adopt technology. Our survey results about how government leaders are thinking about Gen AI bear this out, and it's revealing where their feeling diverge from commercial counterparts. Check out more in the report: https://lnkd.in/d5t5tj2P

As AI becomes more integrated into labs, how do we balance the risks along with the benefits? Scientists argue it will no doubt redefine how we do research, but we need to make safety a priority. Read their opinion: https://ow.ly/8e3650UtaBF

A recent study from researchers at Cornell shows that even the most advanced AI models, like GPT-4o and Claude 3, still hallucinate frequently. Here are the key findings: 🔹 Inconsistent accuracy: even the best models only produce fully factual responses 35% of the time. 🔹 Topic sensitivity: models struggle most with finance and celebrity questions but perform better in geography and computer science. 🔹 Trade-off strategy: some models, like Claude 3 Haiku, lie less because they skip questions they’re unsure of. This has major implications for businesses using AI. While #ai tools are powerful, they aren’t flawless, and errors can be costly. Besides, many employees feel unprepared to use AI tools freely. At Honeycomb, we focus on implementing AI responsibly by integrating oversight systems and fact-checking processes to reduce risks and enhance accuracy. What’s more, we often team up with in-house folks to ensure you can confidently and effectively use AI in your daily workflows. Chat about how we can help you safely take advantage of AI’s potential? #aicommunity

When are combinations of humans and AI useful? Here’s two surprising findings from an MIT meta-analysis of over 100 recent experimental studies of human-AI systems: 1) On average, human-AI combinations performed significantly worse than the best of humans or AI alone. 2) There were performance losses in tasks that involved making decisions but gains in tasks that involved creating content. https://lnkd.in/eeSBg7eq

The use of AI in laboratories brings with it massive benefits, however, it also brings risks. Many of these risks we may not currently realise, or may not be apparent. This piece, recently published in #AdvancedIntelligentSystems Wiley, takes a deeper look at both and provides some perspective on what should potentially be considered. Check it out 👇

What research are AI companies doing into safe AI development? What research might they do in the future? To answer these questions, Oscar Delaney, Oliver Guest, and Zoe Williams looked at papers published by AI companies and the incentives of these companies. They found that enhancing human feedback, mechanistic interpretability, robustness, and safety evaluations are key focuses of recently published research. They also identified several topics with few or no publications, and where AI companies may have weak incentives to research the topic in the future: model organisms of misalignment, multiagent safety, and safety by design. (This report is an updated version that includes some extra papers omitted from the initial publication on September 12th.) https://lnkd.in/g_6DnqgX

Relying on AI to provide you with better and more likely greater quantity of information is a good idea, but not to make decisions for you. Large data aggregator and analyzer will not replace you as a manager, who knows the best way to make a decision to include all factors. including human. This MIT researcher shows it. https://lnkd.in/dwqEcyTm

A recent 2020 report from Statista reveals that the global AI software market is expected to grow approximately 54% year-on-year and is expected to reach a forecast size of $22.6 billion. According to recent data, the average person spends 3 hours and 15 minutes on their phone daily. In the future, they might spend these hours on AI conversations. Check out my latest article on Artificial Intelligence, which has been featured once again in The Print! You can find the article here :[https://lnkd.in/gg4FDnwT ] #theprint #artificialintelligence #humans

I have started writing about AI policy, impacts of AI on politics, and general musings on the state of the world on my substack. Check out my most recent essay here, focused on emerging trends in AI policy globally. https://lnkd.in/gEuuQkpb The TL;DR - AI is in way too early a stage to have hard and fast rules. Governments should learn from and work with industry partners, provide clear guidelines, and only regulate the deployment of AI in high risk scenarios like healthcare or defence.