Reasoning in intelligence | Dr. Prathosh AP | #SangamTalks

Dr. Pratosh reveals a fundamental flaw at the heart of modern AI: large language models have been forced to mimic human reasoning, but this very imposition may be the ceiling that prevents them from making original discoveries. Discover why chain-of-thought training is a double-edged sword, why the Turing Test is already obsolete, and what the new Einstein Test tells us about the true limits of machine intelligence. Dr. Pratosh, an AI researcher and educator, exposes the deep tension between making LLMs safe and controllable versus freeing them to synthesise genuinely new knowledge, while drawing on Indian epistemological traditions — including the concept of Iccha (desire) and Nyaya logical graph theory — to ask questions Western AI discourse has yet to confront. Essential viewing for researchers, educators, technologists, and anyone seeking to understand not just how AI works, but whether it can ever truly think. Key Points • Unpack why imposing human chain-of-thought reasoning on LLMs limits their creative and discovery potential • Explore the gap between AI analysis and genuine knowledge synthesis — why LLMs pass the Turing Test but fail the Einstein Test • Understand how Indian epistemological concepts like Iccha and Nyaya logical graphs reframe the AI reasoning debate • Examine the interpretability crisis: why knowing how an AI answered is answerable but knowing why may be fundamentally impossible • Debate whether metacognition, agency, and desire are the true missing ingredients in artificial general intelligence About the speaker: Dr. Prathosh earned his Ph.D from IISc Bangalore in 2015, submitting his thesis three years after B.Tech. He worked at Xerox, Philips, and a US startup, generating 15 US patents in healthcare analytics, 10 granted, 6 commercialized. He served as Assistant Professor at IIT Delhi (2017-2021). Current research: guided deep-representational learning, cross-domain generalization, ML for scarce data, signal processing for healthcare. Co-founded Cogniable.Tech, winner of Govt. of India's AI startup challenge. Expertise: deep representation learning, generative models, ML for scarce data, digital healthcare. Subtopics: 0:00 Power statement 0:54 Redefining intelligence 2:00 This is a civilisational change 3:20 Indian epistemology meets AI 7:13 How LLMs were force-trained via reinforcement learning to mimic human reasoning graphs 8:51 The interpretability problem 11:46 Scaling laws vs reasoning limits 13:25 should we let LLMs develop their own representational reasoning? 17:51 Pedagogy analogy 23:07 Medical AI interpretability 26:03 Why meaning is stripped during LLM training 27:16 Closing synthesis Hash Tags: #ArtificialIntelligence #AIReasoning #LLM #IndianEpistemology #ChainOfThought #MachineLearning #AIPhilosophy #SangamTalks #DeepMind #EinsteinTest #Metacognition #AILimits #IndianKnowledgeSystems #ChatGPT #AICreativity Join this channel to get access to perks:    / @sangamtalks   Join our YouTube channel to get updated on new video releases: SangamTalks :    / @sangamtalks   SangamTalks Hindi :    / @sangamtalkshindi   SangamTalks Business :    / @sangambusiness   SangamTalks Punjabi :    / @sangamtalkspunjabi   SangamTalks Tamil :    / @sangamtalkstamiz   SangamTalks Malayalam :    / @sangamtalksmalayalam   SangamTalks Bangla :    / @sangamtalksbangla   SangamTalks Marathi :    / @sangamtalksmarathi   SangamTalks Kannada :    / @sangamtalkskannada   SangamTalks Bhojpuri :    / @sangamtalksbhojpuri   SangamTalks Religion :    / @sangamtalksreligion   SangamTalks History :    / @sangamtalkshistory   SangamTalks IKS :    / @sangamtalksiks   SangamTalks Media :    / @sangamtalksclips   SangamTalks Shorts :    / @sangamtalksshorts   For updates, you may follow us on: Facebook:   / sangamtalks   Instagram:   / sangamtalks   Telegram: https://t.me/sangamtalks Twitter /X :   / sangamtalks   WhatsApp: https://whatsapp.com/channel/0029Va95... LinkedIn:   / sangamtalks   Website: https://www.sangamtalks.org Donate: https://www.sangamtalks.org/donate