Podcast
Jay Shah Podcast
Hosted by Jay Shah · EN
Interviews with scientists and engineers working in Machine Learning and AI, about their journey, insights, and discussion on latest research topics.
11episodes
Episodes
Newest firstAll episodes
The Hidden Flaws in AI Safety & Evaluation Benchmarks | Prof. Jackie Chi Kit Cheung
Feb 1801:26:24Tap to summarizeDr. Jackie Cheung is an Associate Professor at McGill University where he co-directs the Reasoning and Learning Lab. He is also an Associate Scientific Director at Mila-Quebec Artificial Intelligence Institute. He and his team are developing computational models to improve the reliability, pragmatics, and evaluation of large language models to ensure they are contextually appropriate and factually grounded.Jackie was worked as a consultant researcher with Microsoft Research and before his current appointments, he earned his PhD and MSc in Computer Science from the University of Toronto, focusing on computational linguistics, and his BSc from the University of British Columbia.00:00:00 Highlight & Introduction00:02:04 Entrypoint in AI & NLP00:04:47 Academia vs. Industry: Career choices00:09:48 Language Revitalization using AI00:12:24 Addressing Biases & Data sovereignty in language revitalization 00:15:49 Evaluating LLMs as Judges00:17:14 Validity and reliability in LLM evaluation 00:25:11 Evidence-centered benchmark design (ECBD) framework00:30:38 Gaps in LLM benchmarks and meaning of "general purpose" AI00:35:24 General purpose intelligence vs reasoning00:40:16 Safety as an undefined bundle in LLMs00:51:45 Stochastic chameleons: how LLMs generalize and hallucinate 01:03:02 Potential & Biases of agentic frameworks for research01:05:52 Evaluating LLMs for summarization01:11:43 Scaling large language models01:16:33 Advice to beginners entering AI in 202601:20:33 Pitfalls to avoid in AI research & development More about Jackie & his research: https://www.cs.mcgill.ca/~jcheung/About the Host:Jay is a Machine Learning Engineer III at PathAI working on improving AI for medical diagnosis and prognosis. Linkedin: https://www.linkedin.com/in/shahjay22/Twitter: https://twitter.com/jaygshah22Homepage: https://jaygshah.github.io/ for any queries.Stay tuned for upcoming webinars!***Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.***
Transcribe →The Future of AI in Pathology: Transforming Diagnosis & Drug Development | Andrew Beck, PathAI
Feb 201:20:12Tap to summarizeAndrew Beck, MD, PhD is the Co-founder and CEO of PathAI, where he and his team are developing AI tools to improve the precision of pathology and the efficacy of drug development for diagnosis of cancer and also many other complex diseases.Before founding PathAI, Andrew was an Associate Professor at Harvard Medical School, where his research focused on the application of machine learning to cancer pathology. He earned his MD from Brown University and his PhD in Biomedical Informatics from Stanford University, where he pioneered some of the first computational models used to predict patient outcomes in oncology.Time stamps of the conversation:00:00:00 Highlights00:01:28 Introduction00:02:18 Entrypoint in AI00:07:02 Background in Medicine and Bioinformatics 00:10:00 Leap from academia to entrepreneurship00:16:20 Translating AI developments to Pathology00:21:15 Specialist vs Generalist AI models in medicine00:24:15 What sets PathAI apart?00:26:32 AI adoption medicine00:34:25 Usage of AI tools in clinical workflows, example MASH00:40:10 AI in Dermatopathology00:42:15 AI for biomarker discovery00:47:05 Will AI models replace pathologists?00:52:28 Avoiding over-reliance on AI00:57:40 Is AI living unto the hype?01:01:00 Challenges in clinical trials 01:05:12 AI reaching patients directly01:09:50 Working at intersection of AI & Healthcare01:15:30 Pitfalls to learn fromMore about PathAI: https://www.pathai.com/and Andy: https://www.pathai.com/about-us/andy-beckAbout the Host:Jay is a Machine Learning Engineer III at PathAI working on improving AI for medical diagnosis and prognosis. Linkedin: https://www.linkedin.com/in/shahjay22/Twitter: https://twitter.com/jaygshah22Homepage: https://jaygshah.github.io/ for any queries.Stay tuned for upcoming webinars!***Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.***
Transcribe →Beyond Accuracy: Evaluating the learned representations of Generative AI models | Aida Nematzadeh
Oct 2300:53:17Tap to summarizeDr. Aida Nematzadeh is a Senior Staff Research Scientist at Google DeepMind where her research focused on multimodal AI models. She works on developing evaluation methods and analyze model’s learning abilities to detect failure modes and guide improvements. Before joining DeepMind, she was a postdoctoral researcher at UC Berkeley and completed her PhD and Masters in Computer Science from the University of Toronto. During her graduate studies she studied how children learn semantic information through computational (cognitive) modeling. Time stamps of the conversation00:00 Highlights01:20 Introduction02:08 Entry point in AI03:04 Background in Cognitive Science & Computer Science 04:55 Research at Google DeepMind05:47 Importance of language-vision in AI10:36 Impact of architecture vs. data on performance 13:06 Transformer architecture 14:30 Evaluating AI models19:02 Can LLMs understand numerical concepts 24:40 Theory-of-mind in AI27:58 Do LLMs learn theory of mind?29:25 LLMs as judge35:56 Publish vs. perish culture in AI research40:00 Working at Google DeepMind42:50 Doing a Ph.D. vs not in AI (at least in 2025)48:20 Looking back on research careerMore about Aida: http://www.aidanematzadeh.me/About the Host:Jay is a Machine Learning Engineer at PathAI working on improving AI for medical diagnosis and prognosis. Linkedin: shahjay22 Twitter: jaygshah22 Homepage: https://jaygshah.github.io/ for any queries.Stay tuned for upcoming webinars!**Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.**
Transcribe →Why Open-Source AI Is the Future and needs its 'Linux Moment'? | Manos Koukoumidis
Apr 15, 202501:19:38Tap to summarizeManos is the CEO of Oumi, a platform focused on open sourcing the entire lifecycle of foundation and large models. Prior to that he was at Google leading efforts on developing large language models within Cloud services. He also has experience working at Facebook on AR/VR projects and at Microsoft’s cloud division developing machine learning based services. Manos received his PhD in computer engineering from Princeton University and has extensive hands-on experience building and deploying models at large scale. Time stamps of the conversation00:00:00 Highlights00:01:20 Introduction00:02:08 From Google to Oumi00:08:58 Why big tech models cannot beat ChatGPT00:12:00 Future of open-source AI00:18:00 Performance gap between open-source and closed AI models00:23:58 Parts of the AI stack that must remain open for innovation00:27:45 Risks of open-sourcing AI00:34:38 Current limitations of Large Language Models00:39:15 Deepseek moment 00:44:38 Maintaining AI leadership - USA vs. China00:48:16 Oumi 00:55:38 Open-sourcing a model with AGI tomorrow, or wait for safeguards?00:58:12 Milestones in open-source AI01:02:50 Nurturing a developers community01:06:12 Ongoing research projects01:09:50 Tips for AI enthusiasts 01:13:00 Competition in AI nowadays More about Manos: https://www.linkedin.com/in/koukoumidis/And Oumi: https://github.com/oumi-ai/oumiAbout the Host:Jay is a PhD student at Arizona State University working on improving AI for medical diagnosis and prognosis. Linkedin: https://www.linkedin.com/in/shahjay22/Twitter: https://twitter.com/jaygshah22Homepage: https://jaygshah.github.io/ for any queries.Stay tuned for upcoming webinars!***Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.***
Transcribe →Differential Privacy, Creativity & future of AI research in the LLM era | Niloofar Mireshghallah
Feb 4, 202501:29:23Tap to summarizeNiloofar is a Postdoctoral researcher at University of Washington with research interests in building privacy preserving AI systems and studying the societal implications of machine learning models. She received her PhD in Computer Science from UC San Diego in 2023 and has received multiple awards and honors for research contributions. Time stamps of the conversation 00:00:00 Highlights 00:01:35 Introduction 00:02:56 Entry point in AI 00:06:50 Differential privacy in AI systems 00:11:08 Privacy leaks in large language models 00:15:30 Dangers of training AI on public data on internet 00:23:28 How auto-regressive training makes things worse 00:30:46 Impact of Synthetic data for fine-tuning 00:37:38 Most critical stage in AI pipeline to combat data leaks 00:44:20 Contextual Integrity 00:47:10 Are LLMs creative? 00:55:24 Under vs. Overpromises of LLMs 01:01:40 Publish vs. perish culture in AI research recently 01:07:50 Role of academia in LLM research 01:11:35 Choosing academia vs. industry 01:17:34 Mental Health and overarching More about Niloofar: https://homes.cs.washington.edu/~niloofar/ And references to some of the papers discussed: https://arxiv.org/pdf/2310.17884 https://arxiv.org/pdf/2410.17566 https://arxiv.org/abs/2202.05520 About the Host: Jay is a PhD student at Arizona State University working on improving AI for medical diagnosis and prognosis. Linkedin: https://www.linkedin.com/in/shahjay22/ Twitter: https://twitter.com/jaygshah22 Homepage: http://jayshah.me/ for any queries. Stay tuned for upcoming webinars! ***Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.***
Transcribe →Reasoning in LLMs, role of academia and keeping up with AI research | Dr. Vivek Gupta
Dec 24, 202401:48:32Tap to summarizeVivek is an Assistant Professor at Arizona State university. Prior to that, he was at the University of Pennsylvania as a postdoctoral researcher and completed his PhD in CS from the University of Utah. His PhD research focused on inference and reasoning for semi structured data and his current research spans reasoning in large language models (LLMs), multimodal learning, and instilling models with common sense for question answering. He has also received multiple awards and fellowships for his research works over the years. Conversation time stamps: 00:01:40 Introduction 00:02:52 Background in AI research 00:05:00 Finding your niche 00:12:42 Traditional AI models vs. LLMs in semi-structured data 00:18:00 Why is reasoning hard in LLMs? 00:27:10 Will scaling AI models hit a plateau? 00:31:02 Has ChatGPT pushed boundaries of AI research 00:38:28 Role of Academia in AI research in the era of LLMs 00:56:35 Keeping up with research: filtering noise vs. signal 01:09:14 Getting started in AI in 2024? 01:20:25 Maintaining mental health in research (especially AI) 01:34:18 Building good habits 01:37:22 Do you need a PhD to contribute to AI? 01:45:42 Wrap up More about Vivek: https://vgupta123.github.io/ ASU lab website: https://coral-lab-asu.github.io/ And Vivek's blog on research struggles: https://vgupta123.github.io/docs/phd_struggles.pdf About the Host:Jay is a PhD student at Arizona State University working on improving AI for medical diagnosis and prognosis. Linkedin: https://www.linkedin.com/in/shahjay22/ Twitter: https://twitter.com/jaygshah22 Homepage: http://jayshah.me/ for any queries. Stay tuned for upcoming webinars! ***Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.***
Transcribe →Time series Forecasting using GPT models | Max Mergenthaler Canseco
Sep 19, 202401:10:21Tap to summarizeMax is the CEO and co-founder of Nixtla, where he is developing highly accurate forecasting models using time series data and deep learning techniques, which developers can use to build their own pipelines. Max is a self-taught programmer and researcher with a lot of prior experience building things from scratch. 00:00:50 Introduction 00:01:26 Entry point in AI 00:04:25 Origins of Nixtla 00:07:30 Idea to product 00:11:21 Behavioral economics & psychology to time series prediction 00:16:00 Landscape of time series prediction 00:26:10 Foundation models in time series 00:29:15 Building TimeGPT 00:31:36 Numbers and GPT models 00:34:35 Generalization to real-world datasets 00:38:10 Math reasoning with LLMs 00:40:48 Neural Hierarchical Interpolation for Time Series Forecasting 00:47:15 TimeGPT applications 00:52:20 Pros and Cons of open-source in AI 00:57:20 Insights from building AI products 01:02:15 Tips to researchers & hype vs Reality of AI More about Max: https://www.linkedin.com/in/mergenthaler/ and Nixtla: https://www.nixtla.io/ Check out TimeGPT: https://github.com/Nixtla/nixtla About the Host: Jay is a PhD student at Arizona State University working on improving AI for medical diagnosis and prognosis. Linkedin: https://www.linkedin.com/in/shahjay22/ Twitter: https://twitter.com/jaygshah22 Homepage: https://www.public.asu.edu/~jgshah1/ for any queries. Stay tuned for upcoming webinars! ***Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.***
Transcribe →Generative AI and the Art of Product Engineering | Golnaz Abdollahian
Sep 5, 202400:35:22Tap to summarizeGolnaz Abdollahian is currently the senior director of big idea innovation at Dolby Laboratories. She has a lot of experience developing and shaping technological products around augmented and virtual reality, smart homes, and generative AI. Before joining Dolby, she had experience working at Microsoft, Apple, and Sony. She also holds PhD in electrical engineering from Purdue University. Time stamps of the conversation 00:00 Highlights 01:08 Introduction 01:52 Entry point in AI 03:00 Leading Big Idea Innovation at Dolby 06:55 Generative AI, Entertainment and Dolby 08:45 How do content creators feel about AI? 10:30 From a Researcher to a Product person 14:27 Traditional Tech products versus AI products 17:52 From concept to product 20:35 Lesson in Product design from - Apple, Microsoft, Song & Dolby 25:34 Interpreting trends in AI 29:25 Good versus Bad Product 31:25 Advice to people interested in productization More about Golnaz: https://www.linkedin.com/in/golnaz-abdollahian-93938a5/ About the Host: Jay is a PhD student at Arizona State University working on improving AI for medical diagnosis and prognosis. Linkedin: https://www.linkedin.com/in/shahjay22/ Twitter: https://twitter.com/jaygshah22 Homepage: https://www.public.asu.edu/~jgshah1/ for any queries. Stay tuned for upcoming webinars! ***Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.***
Transcribe →Future of Software Development with LLMs, Advice on Building Tech startups & more | Pritika Mehta
Aug 14, 202400:37:31Tap to summarizePritika is the co-founder of Butternut AI, a platform that allows the creation of professional websites without hiring web developers. Before butternut, Pritika had entrepreneurship experience building some other products, which later got acquired. Time stamps of the conversation 00:00 Highlights 01:15 Introduction 01:50 Entry point in AI 03:04 Motivation behind Butternut AI 05:00 Can software engineering be automated? 06:36 Large Language Models in Software Development 08:00 AI as a replacement vs assistant 10:32 Automating website development 13:40 Limitations of current LLMs 18:12 Landscape of startups using LLMs 19:50 Going from an idea to a product 27:48 Background in AI for building AI-based startup 30:00 Entrepreneurship 34:32 Startup Culture in USA vs. India More about Butternut AI: https://butternut.ai/ Pritika's Twitter: https://x.com/pritika_mehta And LinkedIn: https://www.linkedin.com/in/pritikam/ About the Host: Jay is a PhD student at Arizona State University working on improving AI for medical diagnosis and prognosis. Linkedin: https://www.linkedin.com/in/shahjay22/ Twitter: https://twitter.com/jaygshah22 Homepage: https://www.public.asu.edu/~jgshah1/ for any queries. Stay tuned for upcoming webinars! ***Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.***
Transcribe →Instruction Tuning, Prompt Engineering and Self Improving Large Language Models | Dr. Swaroop Mishra
Jul 9, 202401:31:39Tap to summarizeSwaroop is a research scientist at Google-Deepmind, working on improving Gemini. His research expertise includes instruction tuning and different prompt engineering techniques to improve reasoning and generalization performance in large language models (LLMs) and tackle induced biases in training. Before joining DeepMind, Swaroop graduated from Arizona State University, where his research focused on developing methods that allow models to learn new tasks from instructions. Swaroop has also interned at Microsoft, Allen AI, and Google, and his research on instruction tuning has been influential in the recent developments of LLMs. Time stamps of the conversation: 00:00:50 Introduction 00:01:40 Entry point in AI 00:03:08 Motivation behind Instruction tuning in LLMs 00:08:40 Generalizing to unseen tasks 00:14:05 Prompt engineering vs. Instruction Tuning 00:18:42 Does prompt engineering induce bias? 00:21:25 Future of prompt engineering 00:27:48 Quality checks on Instruction tuning dataset 00:34:27 Future applications of LLMs 00:42:20 Trip planning using LLM 00:47:30 Scaling AI models vs making them efficient 00:52:05 Reasoning abilities of LLMs in mathematics 00:57:16 LLM-based approaches vs. traditional AI 01:00:46 Benefits of doing research internships in industry 01:06:15 Should I work on LLM-related research? 01:09:45 Narrowing down your research interest 01:13:05 Skills needed to be a researcher in industry 01:22:38 On publish or perish culture in AI research More about Swaroop: https://swarooprm.github.io/ And his research works: https://scholar.google.com/citations?user=-7LK2SwAAAAJ&hl=en Twitter: https://x.com/Swarooprm7 About the Host: Jay is a PhD student at Arizona State University working on improving AI for medical diagnosis and prognosis. Linkedin: https://www.linkedin.com/in/shahjay22/ Twitter: https://twitter.com/jaygshah22 Homepage: https://www.public.asu.edu/~jgshah1/ for any queries. Stay tuned for upcoming webinars! ***Disclaimer: The information in this video represents the views and opinions of the speaker and does not necessarily represent the views or opinions of any institution. It does not constitute an endorsement by any Institution or its affiliates of such video content.***
Transcribe →