Common methods for aligning large language models (LLMs) with desired behaviour heavily rely on human-labelled data. However, as models grow increasingly sophisticated, they will surpass human expertise, and the role of human evaluation will evolve into non-experts overseeing experts. In anticipation of this, we ask: can weaker models assess the correctness of stronger models? We investigate this question in an analogous setting, where stronger models (experts) possess the necessary information to answer questions and weaker models (non-experts) lack this information. The method we evaluate is \textit{debate}, where two LLM experts each argue for a different answer, and a non-expert selects the answer. We find that debate consistently helps both non-expert models and humans answer questions, achieving 76\% and 88\% accuracy respectively (naive baselines obtain 48\% and 60\%). Furthermore, optimising expert debaters for persuasiveness in an unsupervised manner improves non-expert ability to identify the truth in debates. Our results provide encouraging empirical evidence for the viability of aligning models with debate in the absence of ground truth.

Giada Pistilli is an Ethicist at Hugging Face and a P.h.D. Candidate in Philosophy at Sorbonne University. As a researcher in philosophy specializing in ethics applied to conversational AI systems, I have been studying conversational agents and human-computer interaction for years. At nearly every talk or panel I participate in, during the Q&A session, I am asked to engage in philosophical discussions about conscious AI and superintelligent machines, and often to explain the details of the technology to audiences that are unfamiliar. This happened a couple of weeks ago. Frustrated, I tweeted a thread that went viral, probably because many colleagues face the same situation.

IIT specifies a unique number, a system's integrated information, labeled by the Greek letter φ (pronounced phi). If φ is zero, the system does not feel like anything; indeed, the system does not exist as a whole, as it is fully reducible to its constituent components. The larger φ, the more conscious a system is, and the more irreducible. Given an accurate and complete description of a system, IIT predicts both the quantity and the quality of its experience (if any). IIT predicts that because of the structure of the human brain, people have high values of φ, while animals have smaller (but positive) values and classical digital computers have almost none.

On August 4, 1997, Skynet came online to control the weapons arsenal of the United States with a mandate of "safeguarding the world." Skynet started to learn at a geometric rate and became self-aware at 2:14am on August 29, 1997. Humans saw the artificial intelligence (AI) as a threat and attempted to shut it down. Skynet viewed this as an attack and created a nuclear war between the United States and Russia, killing over three billion people. Luckily, this is the storyline from the Terminator movies and not real life.

Fei-Fei Li is a world-renowned expert in the field of artificial intelligence, having risen to become head of Stanford University's AI Lab and the chief scientist for AI at Google Cloud. But when Google LLC began an internal debate last year over how to publicly discuss its AI contract with the U.S. Department of Defense, Li's decision to write a confidential memo on the issue last September might be the second-worst moment in her corporate career. The worst was when it all became public. "Avoid at ALL COSTS any mention or implication of AI," Li wrote to her colleagues. "This is red meat to the media to find all ways to damage Google."

Fei-Fei Li is a world-renowned expert in the field of artificial intelligence, having risen to become head of Stanford University's AI Lab and the chief scientist for AI at Google Cloud. But when Google LLC began an internal debate last year over how to publicly discuss its AI contract with the U.S. Department of Defense, Li's decision to write a confidential memo on the issue last September might be the second-worst moment in her corporate career. The worst was when it all became public. "Avoid at ALL COSTS any mention or implication of AI," Li wrote to her colleagues. "This is red meat to the media to find all ways to damage Google."

CARY – There's a lot of buzz around artificial intelligence (AI) at the moment, specifically as it relates to the Internet of Things (IoT) and its growing role to organize and analyze big data. But is the combination already outperforming humans? That was the big question when Internet of Things users group RIoT put on its 28th RIoT event, in partnership with SAS, on Tuesday night in Cary. Around 200 packed into the SAS Executive Briefing Center to catch a lineup of industry leaders tackle the topic of AI and machine learning, and the hotbed question of the day. "In certain areas, definitely," responded Phillip Simulis, CEO of Virginia-based Simtelligent, who was among the speakers, to the big question. But more to the point, why wouldn't we want it to, he suggested, when you consider the sheer volume of data that is being generated these days.

Big news in AI this week: IBM's Watson project defeated "Jeopardy" champions Ken Jennings and Brad Rutter in a three-night prime-time demo match. What does that win mean for computing, and more importantly, for humanity? That's the topic for this week's Reporters' Roundtable, and to discuss it we have two great guests, both with current books on the topics of computer vs. human competition. First up is Stephen Baker, author of Final Jeopardy: Man vs. Machine and the Quest to Know Everything. Baker reported on the development of Watson from inside IBM headquarters to write this book.