Discovery Lectures

Tuesday - Discovery Lecture Recap

“Generative AI: Teaching AI to Think”

Professor Yi Zhang

On Tuesday, Professor Yi Zhang introduced us to generative artificial intelligence and the research her lab is conducting to make AI systems more reliable and capable of reasoning. She began by explaining that artificial intelligence is a broad field, composed of different technologies (including a chess game from the 1960s!) that perform intelligent tasks, while machine learning allows computers to learn patterns from data. She then talked about generative AI, and explained that large language models like ChatGPT work by repeatedly predicting the most likely next word based on the words that came before it (similar to an advanced form of autocomplete). The secret ingredient is attention—the AI needs to know which of the earlier words matter the most, and adjusts each word’s weights depending on the context. 

Although modern AI systems are incredibly powerful, Professor Zhang discussed several important limitations they still face. She explained that AI cannot always determine whether the information it retrieves is trustworthy, how it often produces similar responses due to homogenization, and that it may struggle with unfamiliar real-world tasks, such as robotics. She also introduced the concept of hallucinations, where an AI confidently generates information that is completely false. 

To address these challenges, Professor Zhang presented several projects from the UCSC Generative AI Center. Her lab’s main question is, “What if AI could think, and double-check itself, before answering?” She’s currently working on giving AI an “inner voice,” much like what we humans have as we think and reason through our lives. Rather than immediately generating a response (and being at risk of hallucinations), her team is teaching AI to first ask the right questions, use tools such as web search or image generation to get more information, evaluate the quality of the information it finds, recognize when it has gathered enough evidence, and clearly explain its reasoning. One of her projects, IM-RAG, divides this process among several specialized AI agents: a Questioner decides what information is missing, a Retriever searches for relevant sources, a Refiner filters out unreliable information, and an Answerer determines whether there is enough evidence to respond, or if additional searching is needed. The system improves through reinforcement learning, where correct final answers are rewarded and the model uses trial and error to learn better research strategies.

Beyond just language models, Professor Zhang showed how reasoning can be applied to images and robotics. Her lab developed systems that encourage AI to ask itself questions about the images it sees before making predictions, guide blind or low-vision users toward areas to gather more information, and even generate images (like diagrams, flowcharts) based on a user’s description to better understand what they are searching for. She also introduced us to projects designed to reduce the “AI echo chamber” by encouraging models to explore multiple perspectives instead of producing identical responses. Finally, she showed us clips of robots that can assemble furniture by retrieving the correct instruction manual, reasoning through each step, and completing the task, without being retrained beforehand!

By the end of the lecture, Professor Zhang showed us that the future of artificial intelligence is not just building larger and faster models, but creating systems that can think more like a human, reason carefully, and learn to apply its skills to various scenarios.

Written by Clarina Manuel

Wednesday - Discovery Lecture Recap

“Spatialized Audio in Games”

Professor Tamara Duplantis

On Wednesday, Professor Tamara Duplantis introduced us to spatialized audio and how sound can be used to create more immersive experiences in video games and other forms of media. She explained that spatialized audio aims to realistically recreate sounds as they exist in three-dimensional space, allowing listeners to perceive not only what they are hearing, but also the direction and distance from which each sound originates. Beyond just improving the realistic qualities of the media, she also emphasized that the placement of sound is also an important artistic and storytelling tool.

Professor Duplantis then walked us through several different techniques used to spatialize audio. One approach, known as ambisonics, reconstructs an entire sound field by capturing the amplitude, direction, and location of sounds around the listener. She also discussed stereophonic and multichannel audio, which use psychoacoustics to create the illusion of three-dimensional sound through multiple speakers. To demonstrate this concept, we listened to the Bohemian Rhapsody and paid attention to how different vocal and instrumental parts seemed to move between the left and right speakers at the front of the lecture hall, creating the feeling that the music surrounded us instead of coming from a single direction.

Then, Professor Duplantis focused on binaural audio, a technique that models how humans naturally hear sound. She introduced the concept of the Head-Related Transfer Function (HRTF), which accounts for how our head and ears naturally filter sounds arriving from different directions. By recreating these effects digitally, binaural recordings can produce remarkably realistic 3D audio, especially when listening through headphones. She also introduced us to technologies such as the Dolby Atmos, which combines multiple surround-sound speakers and overhead channels to create an even more immersive listening experience. 

At the end of the lecture, Professor Duplantis shared one of her recent projects, which was presented at the Game Developers Conference (GDC) in San Francisco! Her team explored how different melodies and spatial audio techniques could guide a player's attention and shape the way they experience a game's story. We all gained a greater appreciation for how thoughtful sound design can transform games, music, and other interactive experiences into something so much more immersive and engaging.

Written by Clarina Manuel

Thursday - Discovery Lecture Recap

“Bioluminesence, Biodiversity, and Biotechnology”

Steven Haddock of the Monterey Bay Aquarium Research Institute (MBARI)

For the ninth Discovery Lecture, COSMOS decided to showcase a non-UCSC speaker. Scientist and researcher Steven Haddock works for the Monterey Bay Aquarium (MBARI) and studies deep sea organisms. He mainly works with organisms with gelatinous structure and bioluminescence, and their effects on ecosystems and ocean ecology alike. 

“Two-thirds of the earth is ‘covered’ by the ocean,” Haddock began. “But that is statistically false.” While this fact is an often parroted one—through children's books and fast media—its contents are not actually true, Haddock explained in his lecture. Ninety-nine percent of the earth is water in an equal-volume projection, is the actual correct representation of the ocean. 

Often misunderstood, the ocean is grandiose and a barely studied place, especially the deep sea. To illustrate common misunderstandings, Haddock showed images found off of the internet that attempted to demonstrate the depth of the ocean. Often portrayed in a comic-like form, places like the mariana trench are boiled down to their relation of how far deep sea vessels or aquatic forms of life can be found in its depths. While often having some semblance of truth, these portrayals are often overstated or exaggerated, or even plain wrong. Haddock desired to correct these wrongs, and show the wonders of the deep sea. 

To encompass his own research, Haddock focused mainly on speaking on four main topics: bioluminescence, fluorescence, Ctenophores, and Siphonophores, all very misunderstood terms. Haddock’s research primarily centers around cenotaphores, which are a type of siphonophore. A Siphonophore is a multitude of gelatinous sea dwelling creatures intertwined together that use tubes for their biological needs. For example, the Hydrozoan. They are partly a stationary creator, almost plant-like. They have tubes for nutritional needs, and for reproductive needs. But, as a part of the life cycle, they produce jellyfish, which are essentially fertilized eggs, to drop what can be called ‘seeds’ on the sea food, which ultimately ‘bloom’ into new Hydrozoans.

And while Ctenophores are similar in nature, feeding and reproducing through tube-like structures, they are different in the sense that they are one organism. While a Siphonophore is multiple organisms operating as one, a Ctenophore is simply an organism. For example, one may think of certain jellyfish as Ctenophora. Both are fundamentally different by nature, but share many common characteristics of deep sea animals, such as bioluminescence and fluorescence. 

Frequently used interchangeably, bioluminescence and fluorescence are not the same thing, Haddock explained passionately. Every living organism has fluorescence, even humans. Fluorescence is plainly just the process in which an organism absorbs light or radiation and emits it back with a typically different wavelength, or color. Bioluminescence, on the other hand, is a fundamentally different process.  

Bioluminescence is when an organism produces light completely individually from any other energy source. In an internal reaction, the organism produces enough energy to excite an electron, which then jumps up to an energy level– then back down. This jumping from energy levels produces a photon of light, which is then excreted. It is a purposeful function, often used for attracting prey, or protecting itself.

Haddock concluded with some suggestions, which sounded suspiciously like warnings. He cautioned against blindly believing internet sources, especially AI, and warned against non-peer reviewed sources. “Always be suspicious,” Haddock added. After, he opened up questions, and showed off a Ctenophore that he himself brought. As the lecture ended, students crowded around to take photos of the jellyfish looking creature. The ninth COSMOS lecture was a victory! 

Written by Emily Halmer-Lee

Friday - Discovery Lecture Recap

“Memristors: Why Do They Remember?”

Professor Nobuhiko P. Kobayashi

In the last Discovery Lecture of the week, Professor Nobuhiko (“Nobby”) Kobayashi showed us how tiny materials can have many impacts on modern technology. He began by explaining thin films, which are extremely thin layers of material ranging from just a few nanometers to a few micrometers thick. These films are used in countless technologies, like computer chips to telescope mirrors, and are typically put onto a solid substrate using techniques such as physical vapor deposition (PVD) and chemical vapor deposition (CVD).

Then, Professor Nobby introduced SALAD, a fabrication process that combines PVD and CVD, two manufacturing techniques that are fundamentally difficult to use together. He showed how this process can improve mirrors used in both ground-based and space telescopes. For example, silver mirrors are incredibly reflective for ground telescopes, but require protective coatings to prevent oxidation and sulfur damage. Aluminum mirrors, which are used in space telescopes such as the Hubble Space Telescope and the James Webb Space Telescope, also require protective layers while maintaining their ability to reflect ultraviolet, visible, and infrared light. By combining different deposition techniques, SALAD allows these protective coatings to be applied more effectively.

The second half of the lecture was about memristors, which are electronic devices that can both store and process information by “remembering” their previous electrical state even after power is removed. Professor Nobby explained that memristors are important components for neuromorphic computing, a field that designs computer hardware inspired by the way the human brain learns and processes information. Because the resistance of a memristor can change and retain multiple states, it can mimic the strength of connections between neurons, known as synaptic weights. His lab studies different materials that can be used to build memristors and develops computer models to better understand why these devices switch between resistance states. 

By understanding the underlying physics of these materials, Professor Nobby showed us how researchers hope to design faster, more energy-efficient computing systems that better resemble the human brain.

Written by Clarina Manuel