# Impact/Challenges of GAI in STEM Education/Research ## Harvard Day ### Natural Science Approach to Understanding and Interpreting Generative AI Dr. Hidenori Tanaka * Natural Science of Artificial Intellgience * Conventional Paradigm in COmputer Science * a computer precisely executes algorithms based on programs written by humans * Paradigm of Deep Learning: Engineering with Emergent Abilities * capabilities of generative AI emerges with the scaling of data, model, and compute * challenges: hallucinations. biases/fairness, unintended plagiarism * *empirical characterization and theoretical modeling of emergent phenomena, akin to physics* * Intertwining of "technology" and "scientific understanding" (industrial revolutions) drives creation of new physics * Therefore what physics will emerge during this next industrial revolution? * And what does this say about emergent abilities and deep leanring with AI * compositional concepts illustrated through stable diffusion images * "concept distance" on a graph quanitfies the "imagination" necessary to generate images from the training set. * AI can compose and sequentially learn to generate from the least to the most imaginative object * MW Q: how is "imagination" here being defined? * theory, based on the graphs: imagination and creativity are being defined according to the "concept distance" of the AI creation away from the training data. * capabilities that require composition of atomic abilities (skills) show emergent curves * Conclusions * INSERT IMAGE TRANSCRIPTION * Questions * someone asked about a paper out of europe that demonstrated that these were not "emergent" abilities, but far more discrete when observed/given data in a more controlled way. ### Enhancing Academic Research with Locally-Run LLMs Dr Nikhil Mukund * Question that drove this research: * what if you had an "AI Co-Pilot," but for research? * word of caution: while thgey function as excellent idea generators, they are essentially text-predictors and must be used with cuation * Now, why a local LLM instead of what is currently available? * Outlining the issues with chatGPT * mostly privacy, cost, and limitations in domain knowledge * Argues that open models are catching up with chatGPT (especoially chatGPT 3.5) * specificially references Mixtral 8X7B LLM * the architectural differences in these models, allow locally run models to at least compete with larger, private ones. * Can/Should you finetune your locally run LLM? * you can, espeiclaly ig you have highly specific use cases that require knowledge of specifc data * BUT: * it requires specialized hardware and is costly * Mukund presents an alternative to fine-tuning: "retrieval-augmented-generation" * essentially, embedding, which I believe is what Evangelos is doing with his tutor bot, but with a LLM thaqt is running locally ### Panel Discussion * In these talks, we have focused on the use of LLMs for higher education. How about younger students? * Dr. Kavita Bala mentioned Khan Academy-- which has rolled out a suite of characters and tutor bots * however, she has said it is difficult to predict, especially how far behind we are on analyzing social networks * A question on cost * how much these [complex searches and AI use-cases](https://app.undermind.ai/home/) (espeically for STEM work) cost? Not just the cost to researchers, but to the company and then for energy consumption * one of the researchers responded that while it is more expensive than using google, it is far less expesnive than if they had hired a human * everyone agreed that as this scales, the cost will go down * just like with every technology * however, good note on the obscene amount of energy it costs to train these models, costs we aren't seeiong passed on to the customers (us) but exist regardless * how much programming should a STEM undergrad have? * Dr Bala: too early to say, this will change and evolve in ways we cannot see * back to cost: how can we afford ot build these models to compete? How can we ensure equity? * Dr. Bala is advocating that governments and universities should we leading th eway in creating open-source, well-trained models with good data. We must lead the way to keep this equitable-- or else business and capitalism will win out. * privacy and security with senstivie data: * what's better-- contracts with these companies that ensure privacy or building a system we can run locally? * Dr. Bala: currently, both. while costs will go lower over time, we also need to be honest about th edata we're putting in. In some domains, these compnaies have no interest. Others are/need to be highly protected. * philosophical question from an education/pedagogy standpoint: how do these mahcines change us? especially around hallucinations * Joshua, from [undermind](https://app.undermind.ai/home/), is not worried at all. Humans are, if anything, more likely to provide false information, and we've been dealing with that for all of history. * you always need to have due dilligence * Dr. Bala nuances that point,arguin that these machines are far more convincing than humans, and will defend their false positions at a higher level than humans. * regardless, we will get very addicted to having these ai "companions" is all parts of our life, especially as functionalities are collapsed into single bots with tailored-for-us personalities * and this, like most technologies, may have net "good" but there will be extreme examples on either end. * a huge lack of trust in anything in younger generations --- lunch hour --- ### what should today's undergraduate know about GAI? Dr. Logan McCarty * Showed a demo of live-polling using OpenAI * So, what do undergrads need? * some considerations: * AI-created content will eventually be undetectable, unavoidable, and, often, unreliable * need essential new skills: * traceable validation of informatyion and reasoning * skill hierarchies will be upended: * coding, datascience, legal advice, etc. * competence will be devauled * discover and exploit unique/unusual skills. ### Automated question modification to challenge AI models Dr Soroush Vosoughi * Introduction * he has been working on machine learning and LLM for years * exciting to have all the attention on his field, but there is also significant pushback on full integration of these technologies * he works with the Minds, Machines, and Society group at Dartmouth, which has a goal of mitgating GAIs anti-social tendencies and to increase transparency and trustworthiness. * This is an approach to assist with student use of LLMs to assist them with courework * gave data on how well chatGPT scored on various tests (AP, LSAT, etc.) * amazing at most subjects, except for high level math and English * reference to the Cornell paper on [Generative Artificial Intelligence for Education and Pedagogy](https://teaching.cornell.edu/generative-artificial-intelligence/cu-committee-report-generative-artificial-intelligence-education) * He asks where people fall on the spectrum from prohibiting chatGPT, allowing with attribution, or to encourage. * Fortifying Assignments Aginst Generative AI * Senstiivty of LLMs to Prompts * they are highly sensitive to the textual specifics of the ionput, often yielding disparate responses to prompts that are semantically similar of textually distinct. * this has given rise to prompt enginerring * but the theoretical basis for such prompt dependency remains largely unexplored. * Knowledge discontinuities in LLMs * we need to define and investigate the conditions necessitating prompt engineering * better prompt = better answer, therefore there exists a knowlkedge inequity among users * there also exist "knowledge discontinuities"-- essentially gaps in the knowledge base of LLMs * *so, theoretically, you could use prompt engineering to nudge your assignment into a knowledge discontinuity without changing the semantic content of the question itself*. * write your question in a way that mirrors the style of a different educational domain * they have created an automation that does this automatically, based on their research * MW question: but for how long will this last? this feels like intellectual pacman-- you are constantly attempting to "outrun" the knowledge of a LLM by mapping out these knowledge discontinuities and exploiting them. ### AI and STEM Education PhD Jack Maier * Introduction * "When will I need to use this?" * so why should they learn STEM? * when we have apps and tools to solve virtually any question outside of the highest levels * He is a phd student, so how does he use it? * explaining concepts * writing code * highlighting AI use in experimental work * AI finding crystal structures robots carrying out 24/7 lab work * Two major questions: * Where do humans fit into the scientific process? * What is the use of STEM education in the age of AI? * So, is there anything we can ultimately do better than AI and robotics? * what is our place in the universe? * we are better at being human * MW: what does that mean? organic bodies with souls? lol * Therefore: * education should go from "producing specialized workers" to "making the best human" * What does it mean to be human? * holistic education * "let no one ignorant of geometry enter here" -Plato * math and logic are foundational, before getting to work in the world of rhetoric and the humanities ### Comparing AI-Supported Instruction vs. Active Learning in Introductory Physics Dr. Gregory Kestin * AI tutors * Sal Khan as a propoonent, sayign that AI tutoring bots will revolutionize education * though critics argue that AI could have degenerative effects * They wanted to use an AI tutor for their class, with a focus on educationla principles from traditional teaching * They did not do embeddings or fine-tuning, but they did a lot of system prompts. * caps lock, repetition of direction, specific directions * They tested "active learning" vs AI tutor-supported learning * they split the class in half * TDLR via many graphs; the AI-supported students learned twice as much and felt more engaged and motivated * What is it? * It's a UI with OpenAI API calls, with a system prompt ### A and Mathematics: Applications in Low Income, K-12 Contexts Damion Mannings * This talk is a response to the emergence of tutor bot usage * with a specific look at low-income communities, as one of the best indicators of future success * adds a note for rural communities * MW is pleased * and a specific focus on mathematics as a case study * Why math? * math anxiety is high * low math can indicate a lack of prepadness for tests, which correlates to later success ### ChaGPT in CS50 * same presentation as usual * duck debugger * scaled to students in class and to those who are taking it online * the "heart" system to control use * which they say has a learning objective, but I think it's to limit the cost of the API calls --- ## MIT DAY ### Keynote: Generative AI: What's it good for, and what's it good at? * What's it good for, and what it's good at? * outline * does generative AI present an existential threat or is it a remarkable opportunity? * yes. * GAI changes the nature of interaction with computers * lowers barrier of entry for sophisticated computing * presents challenges regarding verification and validation * fast moving and scalable * GAI disrupts our current model of higher education * 1188 Explorer Assistant * a custom GPT, with guardrails and constraints * a basic system prompt * What GAI good for in STEM? * Teaching * Admin * Research * INSERT IMAGE TRANSCRIPT * Stubs examples: * editing student work * creating a cover for a book * crafting homework assignments for students * live student input * creating histograms and python out of complex data sets * Work in progress: * facilitated browsing of 20 years of Rubin Observatory project documentation * GAI integration into commissioning and operations of the observatory * facilitated administrative screening of data use * GENED course on GAI * Is there a prospet fo GAI making independent intellectual breakthroughs? * currently mostly an efficiency tool * MW NOTE: though I would argue examples like [this](https://arstechnica.com/ai/2023/11/googles-deepmind-finds-2-2m-crystal-structures-in-materials-science-win/) demonstrate some breakthroughs, though not novel. * but yes, eventually (especially with reinforcement learning models) * Will the intinsic lack of GAI reproducibility impact the scientific method? * yes, until we fully understand the "black box" qualities ### A2rchi [slides here](https://drive.google.com/file/d/1l3RsMdECoe5vklpyh__z4kM3qw0UsBy1/view?usp=drivesdk) * For academic usecases, how to modfy LLMs for our purposes * potential solutions * our own LLM * finetuning * Or RAG (and embeddings) * cheapest, easiest form, one step above a system prompts * RAG outperfroms finetuning, and is far cheaper * RAG tools * Not directly application: * DSPy * LlamaIndex * LangChain * HuggingFace * Too niche: * Moveworks * cohere * Julius * So they made their own! * [A2rchi](https://github.com/mit-submit/A2rchi) * end-to-end RAG, compatible with an API * all open source * command line interface * built in monitoring * admin portal * many interfaces * automated help desk service, etc ### GenPhys and Beyond: Unveiling the Synergy of Physics and Diffusion Models in Generative Processes * Physics is generative * so, how can we apply physics to diffusion models (camera vision)? * is there a universal converter to build a bridge beetween physics and generative models? * PDEs as a converter, espeically between equations and illustrations * especially in dispersion and diffusion * PDEs are also smooth and linera, but if you go beyond to other equations, such as reaction-diffusion and Naiver Stokes equations ### Supernova Science in the AI Era * Realtime ANalysis with Large Language Models * a golden age of (triaged) discovery * very few supernovae were observed with lesser technology * one observatory (the Rubin) will detect 10 million supernovae candidates per night * The Young Suopernova Experiment (YSE) * purchasing time on the hawaiian observatory * overlapping the captured data with other telescopes in order to triage phenomena that could be supernova * a variable mix of data (some written, some birghtness, etc). * Question: can we use AI to help us triage all this data effieciently * the SpeakYSE (lol) * a LLM referencing this data, which you are able to quickly query * up to five different tables * however, there are still hallucinations * they have added a vectorized natural language database to help reduce hallucinations * They are attempting to create an enormous embedding via RAG, in many different modalities, to give the LLM expertise knowledge to improve the way in queries and analyzes the data * this allows the LLM to contextualize the data it is retrieving (and helps the researchers not miss anything, even if it outside of the "norm") * multimodal embeddings (so this plus A2rchi?) * gaglian2@mit.edu