Thomas Thompson’s Post

Prompt engineering meets Bloom’s taxonomy: Jackson’s STEL article highlights how AI-powered prompts can deepen critical thinking and engagement in higher education. How are you integrating advanced AI and pedagogical frameworks in your teaching or learning design? #EdTech #PromptEngineering #LearningDesign. https://lnkd.in/g4wch98F

I’m thrilled to share my latest AI project — a Retrieval-Augmented Generation (RAG)-based AI Teaching Assistant, developed as part of my journey into advanced Artificial Intelligence and Natural Language Processing. 💡 Project Overview: The system is designed to function as an intelligent teaching assistant that answers academic or conceptual queries with precision and contextual understanding. It combines retrieval mechanisms (for fetching relevant information) with generative AI (for natural language responses), ensuring accuracy and relevance in every answer. 🧠 Core Concept: Retrieval-Augmented Generation (RAG) bridges the gap between static language models and dynamic, knowledge-grounded systems. By integrating a retriever to fetch external data and a generator (LLM) to produce responses, the assistant provides context-aware, reliable information — beyond the limitations of a standalone model. ⚙️ System Architecture Highlights: • Video to Audio Conversion: Extracts the audio track from lecture or learning videos. • Audio to Text Transcription: Converts speech into text data using speech recognition. • Document Ingestion: The transcribed text is converted into vector embeddings for efficient storage. • Context Retrieval: Performs similarity search to fetch the most relevant chunks of information. • Response Generation: The LLM generates coherent, factually grounded answers using retrieved context. • Evaluation Layer: Ensures quality through relevance and coherence scoring. 🎯 Objective: To build an AI assistant capable of enhancing self-learning by delivering context-rich, accurate, and explainable answers to user queries — simulating an intelligent academic mentor. 📊 Key Outcomes: •Improved contextual accuracy in responses •Enhanced explainability of AI-generated content •Modular structure enabling future scalability and deployment This project has been a remarkable learning journey in understanding how retrieval-based methods and large language models can complement each other to create more reliable and intelligent AI systems. I’m eager to keep exploring and refining this concept further. #ArtificialIntelligence#NLP#RAG #AIProjects #Research #DataScience #LLM #Innovation

Fascinating work that I was lucky to collaborate on! We created a fine-tuned LLM mentor that significantly improves on pedagogical capabilities over SOTA models! I'm also especially excited about Janos's novel multi-turn evaluation paradigm for benchmarking AI tutors. Read the paper here: https://lnkd.in/gvCXdxHJ

If you ask me, real teaching is one of the hardest skills in the world. Great teachers build genuine rapport, establish key background and context, structure a path of inquiry that sparks students' curiosity, encourage dialogue and expression of ideas, and don't give answers away but scaffold the process for the student. Great teachers do this dynamically and empathically, and constantly pivot and adapt based on how students respond. Students are losing out on this essential guidance when they rely on LLMs for assistance today, which we know (https://lnkd.in/gHrsd559) seriously harms their critical thinking skills. This should worry all of us -- cognitive labor is an imperative for real learning. The struggle is the point. That's why I am so excited about this new research by our founders Janos Perczel and Jin Yun Chow (along with Stanford's Dorottya Demszky), which demonstrates how we can train models on what real, good pedagogy looks like using authentic learning data. The results are pretty exciting -- check it out! So proud to work with these leaders.

Thrilled to share preliminary findings from new research study TeachLM! It's been a dream collaborating with my cofounder Janos Perczel and Stanford researcher Dorottya (Dora) Demszky on this exciting work. With the world's ever increasing dependence on AI for all facets of learning and work, there's really no better investment than increasing the pedagogical capabilities of LLMs. To learn how to teach, LLMs must first learn how to learn. That's the core thesis behind the novel approach of this study. Would love to hear your thoughts and feedback!

Me: Chat GPT How should schools and teachers balance use of chat GPT? Chat GPT: "Excellent and timely question, Sunil especially as schools grapple with whether Chat GPT should be treated like a calculator, a textbook, or a cheat code. The real goal is to make AI a thinking partner, not a thinking substitute. Here’s how schools can balance it thoughtfully: 1. Set Clear Philosophy; Not Just Policy Before creating rules, schools must define why AI is being used. Is it to enhance thinking or to speed up tasks? Is it a creative companion or a fact checker? Schools should adopt an “AI Literacy Charter” a shared code that celebrates curiosity, honesty, and responsibility in AI use. 2. Teach AI Literacy Early Students should learn: How Chat GPT works (patterns, probabilities, biases) What it can and cannot do (e.g., it predicts, not thinks) How to fact-check AI outputs How to cite AI-generated material ethically Make “Prompt Engineering” part of digital literacy; it’s today’s version of learning how to search or code. 3. Blend Human Thinking with AI Thinking Encourage assignments that require both: Human judgment + AI assistance. Example: Students use Chat GPT to brainstorm, but must justify which ideas they kept or rejected. Reflection journals: “How did AI shape your thinking today?” The motto: Use Chat GPT to start your thinking, not to stop it. 4. Redesign Assessments Traditional rote-based tests invite misuse. Instead: Focus on open-ended projects, design challenges, simulations. Use oral defenses or in-class reflections to check understanding. Introduce “AI-assisted assignments” where use is expected but reasoning must be transparent. 5. Empower Teachers Teachers need space to experiment too. Offer: AI co-teaching sessions e.g., lesson plans or quiz creation with Chat GPT. Peer circles where educators share what works and what backfires. Guidelines on data ethics and student privacy when using AI tools. 6. Create a Culture of Ethical Curiosity Make it cool to ask: “Did Chat GPT help you think better or just faster?” Encourage peer accountability, classroom debates (“AI vs Human Mind”), and school exhibitions on ‘Responsible Use of AI in Learning.’ In Short Do More Of, Do Less Of Using AI to brainstorm, visualize, summarize, prototype, Copying homework or essays, Reflecting on how AI changed your answer, Submitting AI output as-is, Combining human insights + AI creativity Relying on AI as authority Teaching thinking about thinking, Punishing tech use blindly etc." Me: Thank you Chat GPT But our schools never learn!

“The thing I find fascinating is how sophisticated the answers are,” Tomescu said. “For 60% of students to say they are concerned that AI tools encourage copying rather than doing original work, that’s a very deep understanding of what your schoolwork is meant to help you do, and what the pitfalls and benefits are associated with this technology. “Young people who are using this technology actually have a pretty sophisticated, quite mature understanding of what the technology does in relation to their schoolwork, which is fascinating because we don’t give young people enough credit when it comes to using technology in an educational space, unaided, in this way.” OUP’s findings follow empirical studies on the use of AI in education. One published this year by the Massachusetts Institute of Technology (MIT) measured brain electrical activity during essay writing among students using large language models (LLM) such as ChatGPT, and concluded: “These results raise concerns about the long-term educational implications of LLM reliance and underscore the need for deeper inquiry into AI’s role in learning.” Nearly half of the 2,000 students surveyed by OUP said they were worried their classmates were “secretly using AI” for schoolwork without their teachers being able to spot it. Many reported that they wanted more help from teachers for the appropriate use of AI and in judging whether its output was reliable. OUP said it is launching a new AI education hub aimed at supporting teachers. “Some of these findings will be very interesting for teachers, especially around how much students are expecting guidance from teachers. We sometimes think there is a technological generational divide, and yet they are still looking at their teachers for guidance in how to use this technology productively, and I find that very positive,” Tomescu said. Daniel Williams, an assistant headteacher and AI lead at Bishop Vesey’s grammar school in Birmingham, said: “The findings closely reflect what I see in school. Many pupils recognise AI’s value for creativity, revision, and problem-solving but often use it as a shortcut rather than a learning tool.” https://lnkd.in/e6rPq7Xy

AI Literacy Beyond Deleuze's Dogmatic Image of Thought AI literacy often begins with the question of accuracy. The goal, it seems, is to teach students to test the reliability of LLM outputs, to separate true from false, fact from hallucination. The exercise appears responsible, but it risks reducing thinking to fact-checking. Accuracy becomes a proxy for mastery without addressing how language itself generates meaning. When students are trained to check without creating, they inherit a model of cognition in which knowledge functions as an inventory of verified statements rather than an ongoing process of sense-making. The philosopher Gilles Deleuze offered a different image of thought. For him, philosophy, science, and art all begin with problems, not as gaps in knowledge, but as conditions that give rise to thinking. Thought begins, he argued, when something resists classification, when the mind meets what cannot be easily absorbed into what it already knows. The task of thought is to respond to this resistance, not by reducing it to definitions, but by creating with it. Consider this quote from a basic AI output generated in response to the prompt “What is photosynthesis?” “Light energy becomes chemical energy.” Within conventional AI literacy, this sentence becomes an opportunity to verify accuracy. But that kind of fact-checking approach reflects what Deleuze called the dogmatic image of thought: the idea that thinking begins only once truth has already been decided, and that “critical thinking” means checking for correctness. Photosynthesis, here, becomes a fixed process awaiting recognition. The role of the learner is to confirm: true/false; check A,B,C, or D. Problematizing, in the Deleuzean sense, involves something else. It treats the sentence as a site of potential, a field from which new sense might emerge: “Light energy becomes chemical energy.” It appears to describe a smooth transformation, one form (light) becomes another (energy). But it also gestures toward a deeper encounter: light meeting leaf, matter feeling its way into pattern, energy folding through time until something unanticipated takes shape. The process does more than just transfer energy; it generates difference. With each modulation, new conditions for meaning take form. Posed as a problem in the Deleuzean sense, the line now shows us something new: how energy becomes expressive, how, in the encounter between light and matter, new meanings or relationships emerge. This happens when description stops merely representing and begins to transform perception. To teach AI literacy through this lens is to shift from checking to creating, from confirming what is said to producing the conditions for new thought. Every AI output can become an opening, a surface of potential where knowledge begins again and language experiments with what it might become. Checking accuracy may close this process, but problematization opens it up to the new. #AlLiteracy #Deleuze

“I have been able to understand maths better and it helps me to solve difficult questions.” “I now think faster than I used to.” These are quotes from pupils in a new report on AI in education, featured in https://lnkd.in/eBy2fubD At school, kids are mostly told that using AI is cheating and not proper learning. At work, we’re told we won’t be replaced by AI, but by someone who can use AI. The same technology that’s seen as a threat to learning is being celebrated as a skill in the workplace. Bit contradictory innit

How can educators deliver consistent, rubric-aligned feedback in real time - without compromising quality? At Queen Mary University of London, the EduMark AI project is answering that question. Led by Dr. Deepshikha and supported by a cross-disciplinary team, EduMark AI is an exciting proof-of-concept pilot project which integrates large language models into the assessment process to: ⏱️ Reduce marking time 📊 Provide personalised, rubric-based feedback 🎓 Enhance student experience and assessment literacy In her new blog post, Dr Deepshikha reports back on the findings from the EduMark pilot https://lnkd.in/e8sYDV-k