FAU Researchers Question ChatGPT's Understanding

Key Takeaways

• FAU researcher Elan Barenholtz found that simple word patterns can match large language models like ChatGPT on tasks involving geography, pain and emotions, questioning whether AI truly understands concepts
• UNC Chapel Hill launched a 15-credit-hour AI minor open to all majors, covering AI use in writing, research, data analysis, ethics and safety
• About 40 percent of students in one survey admitted to using AI for coursework in ways that violated class policies, with Harvard researchers pointing to lack of curiosity as the root cause
• OpenEvidence and Johns Hopkins Medicine partnered on October 13 to develop AI tools for physician training and clinical care while preserving core diagnostic skills
• Federal agencies including EXIM and the State Department manage AI with zero-trust and privacy controls, with leaders warning that quiet AI failures pose a bigger risk than loud ones
• The National Institute of Allergy and Infectious Diseases is building a shared platform to speed AI integration across enterprise government systems
• Open hardware specifications could reduce AI data center integration time from nine to twelve months down to three to six months
• SOFA, a State-aligned Order-Flow Allocation system, lowered execution costs by 0.040 tick on average when tested across 20 dealer populations in an ABIDES market
• AI is democratizing knowledge that once required years of training, making judgment and the ability to apply knowledge the key differentiators among professionals
• The University of Nevada, Reno hosted a cybersecurity conference on March 22, 2023, featuring a keynote by Heather Case-Hall titled 'AI didn't break in. We let it in.'

AI Understanding May Just Be a Word Game Say FAU Researchers

An FAU researcher named Elan Barenholtz found that simple word patterns can reveal much of the same information that large language models like ChatGPT appear to understand. He tested word vectors on tasks involving geography, history, pain and emotions. The simpler method matched or nearly matched the performance of LLMs on many tests. This suggests AI may not truly understand the world but simply relies on word associations.

FAU Researcher Questions AI Understanding of Space Time and Pain

Elan Barenholtz from Florida Atlantic University studied whether large language models truly understand concepts like location, time and emotion. He compared LLMs to simpler word-based mathematical tools and found both could predict geographic coordinates and identify pain descriptions with similar accuracy. His work challenges the idea that extracting information from AI proves it has an internal model of the world. The research was published through FAU's Charles E. Schmidt College of Science.

New SOFA Method Tests How Order-Flow Rules Work After Dealers Adapt

Researchers developed SOFA, a State-aligned Order-Flow Allocation system with three parts to test trading rules. It uses a State Scorer, Commitment Scheduler and Quote-to-Flow Allocator to route customer demand toward tighter quotes. Tested across 20 dealer populations in an ABIDES market, concentrating interventions in low-risk states lowered execution costs by 0.040 tick on average. The approach breaks down one link at a time to reveal why a rule actually works.

UNC Chapel Hill Launches New AI Minor for All Students

UNC Chapel Hill introduced a 15-credit-hour AI minor open to students of all majors, not just computer science. The curriculum covers using AI for writing, research and data analysis along with ethics and safety. Rei Sanchez-Arias, the Associate Dean who helped develop the program, said it teaches students to use AI responsibly. Students from political science, nursing, business and other fields are already enrolled.

AI Is the Great Equalizer Changing What Expertise Is Worth

An opinion piece argues that AI is democratizing knowledge that once required years of training to acquire. The author describes using an AI tool to plan a city tour in under 12 hours and seeing a company compress weeks of work into hours. The real equalizer is not expertise but the software itself. Judgment and the ability to apply knowledge in specific situations are now what truly set people apart.

University of Nevada hosts AI cybersecurity conference

The University of Nevada, Reno will host a cybersecurity conference on March 22, 2023, focused on the risks and rewards of artificial intelligence. Cybersecurity expert Heather Case-Hall will deliver the keynote address titled 'AI didn't break in. We let it in.' The event will include panel discussions, workshops, and networking opportunities. The conference is open to the public and registration details are available on the university's website.

OpenEvidence and Johns Hopkins partner on AI medical training

OpenEvidence and Johns Hopkins Medicine announced a collaboration on October 13 to develop AI tools for physician training and clinical care. The partnership focuses on redesigning how doctors are trained in an AI-enabled environment and building systems that connect medical research with patient histories. The goal is to support care decisions while maintaining core skills like clinical reasoning and diagnosis. Specific capabilities and commercial arrangements remain under development.

Building accountable AI in government requires strong governance

Federal agencies are testing AI but face challenges in embedding it into core systems like financial operations and grants management. Strong governance is essential because the bigger risk is AI failing quietly rather than loudly. Agencies like EXIM and the State Department manage AI as part of their IT portfolios with zero-trust and privacy controls. Leaders say organizational alignment and leadership courage are tougher challenges than technology selection.

Agency AI adoption tests infrastructure costs and oversight

Government and industry leaders say federal agencies need adaptable platforms and human oversight as AI takes on more decision-making roles. The National Institute of Allergy and Infectious Diseases is building a shared platform to speed AI integration across enterprise systems. HP and GAO officials addressed AI compute costs and accountability for autonomous systems. New risks emerge once systems make consequential decisions, and agencies must confirm AI systems pursue intended goals.

Open specs could speed AI data center hardware builds

Open hardware specifications could cut AI data center integration time from a year to just months. Hardware interoperability is becoming as important as raw performance as accelerators, memory and storage come from many suppliers. OCP's server project showed that different suppliers reading specifications differently caused server fleets to drift apart. Standardized specs could reduce reintegration time from nine to twelve months down to three to six months.

AI misuse in schools stems from lack of student curiosity

Harvard students argue that AI misuse in education is not caused by AI itself but by a failure to cultivate curiosity. About 40 percent of students in one survey admitted to using AI for coursework in ways that violated class policies. Two-thirds of those students said they understood the rules. The authors say AI simply lowers the cost of cheating. They believe schools should focus on inspiring students to love learning rather than only trying to prevent AI misuse.

Sources

• Can AI Understand Space, Time and Pain? FAU Researcher Says It May Be a Word Game
• FAU Researcher Asks: Can AI Understand Space, Time and Pain?
• When and How Order-Flow Interventions Work After Dealers Adapt
• UNC Chapel Hill new AI minor
• AI is the great equalizer
• University of Nevada cybersecurity conference to explore risks, rewards of AI
• OpenEvidence, Johns Hopkins Partner on AI Medical Training
• What it takes to build accountable AI in government
• Agency AI Adoption Tests Infrastructure, Costs, Oversight
• Hardware interoperability: Open specs speed AI data center builds
• AI Isn’t the Problem. We Are.