I, scientist with Balazs Kegl

00:00:00 Intro. Gaels's journey from physics to AI through coding, then health and social science applications.00:06:17 Looking for impact. Are we using our energy to solve the best problems? How to estimate future impact?00:12:18 How did the interacting with a wide variety of sciences changes you as an AI researcher? Out-of-the box. Empirical research.00:13:19 Benchmarks. How they incorporate value and drive AI research. AI went from a mathematical to an empirical science. Fei-Fei Li and ImageNet.00:19:07 The Autism Challenge: predict the condition from brain imaging. How to avoid fooling ourselves?00:25:24 How did the medical community react? The clash between what is true and what is valuable.00:27:15 How do you measure your scientific impact?00:31:09 Scientific/technological and societal progress.00:33:01 Recommender AI and the 2007 Netflix challenge.00:35:14 How to deal with social media addiction.00:42:06 Scikit-learn. The Toyota of AI. Origin story.A well-designed tool for scientist is also useful for business. 00:47:03 Open source organizational structure. Ecosystem building.00:55:57 Deep learning and scikit-learn.00:59:37 Sociology ad psychology of scikit-learn.01:09:57 How to bring science home. AI has become an ice breaker.01:13:13 Gael's question: what excites me these days. Being seen; clarifying my thoughts through dialogs and writing; agency, RL, and putting AI in hardware we connect with; moving my body.I, scientist blog: https://balazskegl.substack.comTwitter: https://twitter.com/balazskeglArtwork: DALL-EMusic: Bea Palya https://www.youtube.com/channel/UCBDp3qcFZdU1yoWIRpMSaZw Hosted on Acast. See acast.com/privacy for more information.

00:00:00 Intro00:04:20 The LHCb experiment. Fundamental particle physics. Why isn't there as much antimatter as matter? Timescales.00:17:18 Shortcomings of the Standard Model. Dark matter. The LHC.00:22:24 The LHCb collaboration. Organization of a scientific experiment beyond the Dunbar number. Carrier development in academia and physics.00:27:03 The skills and day job of an LHC physicist. The messy organization of a big experiment. Technical vs physics work.00:32:33 The (lack of) management levers and incentives. 70+ institutes. Where does meaning come from?00:38:26 Vava's journey from war-torn Yugoslavia through Vienna and Oxford to CERN in Geneva and permanent position in Paris.00:41:23 Why physics? Curiosity and introversion. The helpers on a hero's journey.00:45:03 The real-time aspects: how we take the data. 30 million+ proton-proton collisions, a few to 30+ Terabits per second. The real-time trigger reduces the rate by 3-4 orders of magnitude.00:48:39 Working in a small group. Career without planning in the early days vs students today.00:51:48 Early adoption of AI and GPUs in the real-time trigger. Separate signal (interesting events) from background (known particles) in a million-dimensional space. Reconstruction cuts it down to 10-20 features where we apply Boosted Decision Trees. Training data and simulation. Neural nets? Sometimes, in complex feature spaces, for example in the calorimeter.01:01:16 Simulation to real data: systematic uncertainties. How to prioritize what to care about? The soft process and social structure of scrutinizing results. The effect of the aggregated knowledge of the collaboration.01:09:29 The delicacies of the scientific method: the look-elsewhere effect and unknown unknowns. The soft side of the Popperian ideal.01:16:46 Who decides what to go after in physics? LHCb: 20 years x 100 PhD thesis is a lot of investment. The role of the critical mass.01:20:27 The International Linear Collider and the sociology of the next big experiment.01:24:10 ATLAS = Cathedral. The deep metaphor: multi-generational experiments. The sacrifice of early-career scientists.01:29:56 Vava's dream for the rest of his career. Survival guilt.01:36:01 Science at home. 01:39:43 Why the podcast? - Vava's question to me. Spirituality and science. Anger and separation anxiety. This little corner of the internet. Truth and importance: the daily dilemma at the Paris-Saclay Center for Data Science.I, scientist blog: https://balazskegl.substack.comTwitter: https://twitter.com/balazskeglArtwork: DALL-EMusic: Bea Palya https://www.youtube.com/channel/UCBDp3qcFZdU1yoWIRpMSaZw Hosted on Acast. See acast.com/privacy for more information.

00:00:00 Intro. Bogdan's journey through the French system towards PhD in AI. Inspiration by early DeepMind papers, research on LSTM and other recurrent architectures.00:05:29 Oxford postdoc between ML and Neuroscience, theory of mind. Turn towards safety. Influence of Nick Bostrom's 2014 book Superintelligence https://en.wikipedia.org/wiki/Superintelligence:_Paths,_Dangers,_Strategies.00:07:29 AI acceleration. AlphaGo, Startcraft, BART, GPT-2. Language. Learning people's preferences.00:10:24 The curious path of an independent AI safety researcher. 80000 hours https://80000hours.org/, the Alignment Forum https://www.alignmentforum.org/, LessWrong https://www.lesswrong.com/, effective altruism. 00:18:15 GPT: in-context learning, math.00:23:47 GPT: planning and thinking. Planning in the real world (reinforcement learning) vs planning in a math proof, planning as problem solving.00:27:29 GPT: chain of thought. "Let's think about this step by step."00:31:47 GPT: lying? HAL from 2001 Space Odyssey. Does GPT have the will to do something? Simulators, Bayesian inference, simulacra, autoregressivity. The surprising coherence of GPT-4. Playing personas.00:43:38 GPT: reinforcement learning with human feedback. RLHF is like an anti-psychotic drug? Or cognitive behavioral therapy?00:45:38 GPT: Vervaeke's dialog Turing test https://balazskegl.substack.com/p/gpt-4-in-conversation-with-itself.00:52:36 AI Safety. The issue of timescale. The OpenAI initiative https://openai.com/blog/our-approach-to-ai-safety. Aligning by debating.00:57:46 Direct alignment research; Bogdan's pessimism.The 2-step approach: automate alignment research. Who will align the aligner AI?01:04:11 Alignment by giving agency to AI. Embodiment. Let them confabulate but confront reality.01:12:09 Max Tegmark's waterfall metaphor. Munk debate on AI https://www.youtube.com/watch?v=144uOfr4SYA, Yoshua Bengio's interview https://www.youtube.com/watch?v=0RknkWgd6Ck.01:22:21 Open source AI. George Hotz interview https://www.youtube.com/watch?v=dNrTrx42DGQ. Bogdan's counterargument: engineering a pandemic. Some tools make a few people very powerful.01:28:15 Adversarial examples.01:31:32 Bogdan's dreams and fears, where are we heading? Hosted on Acast. See acast.com/privacy for more information.

00:00:00 Intro.00:01:34 Jonas' story. Math, physics towards computer science, AI, and robotics.00:05:09 Embodiment in intelligence.00:07:25 LLMs in robots. Should we put LLMs in robots? Can we teach them as kids if they can speak? Will they develop their personalities depending on their unique experience? Should we add episodic memory to them?00:19:02 Relevance realization. How to filter important information from the immense incoming flow of signals? The top-down aspect of perception. Cultural learning and binding.00:23:30 Limits of GI. Is there an inherent limit to how intelligent a being can be? What if too much intelligence makes the map take over the agent, leading to something like schizophrenia?00:29:20 Continual learning. The brain is a little scientist. The scientific method: where do the hypotheses come from? Where does the value of a proposition come from? How do we decide what proposition to prove or what experiment to run? Why did I work on the Higgs boson?00:37:37 Dog intelligence. Do dogs want to "go beyond" what is "visible", or is it a purely human drive?00:40:00 Collective alignment. Higher level collective consciousness and its relationship to human and AI alignment.00:47:21 AGI. How far are we from AGI?00:50:38 Robots. Bodies are the bottleneck of robotics research.00:57:09 Jonas' dream: connecting the dots, merging the cognitive modules and experiment in the real world, towards a dog intelligence in 5-10 years.01:01:13 High-functioning zombies. Should we be afraid of the them: an agent smart enough to plan, but not smart enough to see the harm that some of the planned actions may cause?I, scientist blog: https://balazskegl.substack.comTwitter: https://twitter.com/balazskeglArtwork: DALL-EMusic: Bea Palya https://www.youtube.com/channel/UCBDp3qcFZdU1yoWIRpMSaZw Hosted on Acast. See acast.com/privacy for more information.

1:35 Screwdriver hands: how Giuseppe became a scientist.6:07 Reinforcement learning (RL) explained to Giuseppe's grandma.9:01 Model-based reinforcement learning: how to fry an egg.10:45 The three components of model-based reinforcement learning: the actor, the model, and the planner.16:29 Planning = thinking: ham & eggs and Google map19:01 RL is responsible for collecting its own data22:05 Vervaeke's first two Ps: propositional (GPT) and procedural (RL)24:05 Fear from AI: the paperclip scenario, evil, indifference, and foolishness28:19 Should we add agency into AI? Can we socialize AI as we do with kids?32:22 Do we need to embody AI to align it? Can GPT bike? Is textual knowledge everything?36:28 Why to aspire to the dream of creating AI? 1) Why not :)? 2) Curiosity. 3) To understand ourselves better.41:52 How AI changed our lives. Recommendation engines, addiction, jujitsu, conspiracy theories, the attention economy.49:12 Open source large language models. AI as a mirror. Individuation of LLMs.51:16 Putting AI into things: they will individuate.52:56 Vervaeke's 2nd person Turing test. Let GPTs talk to each other.54:37 Could AI manipulate us?56:11 Closing.I, scientist blog: https://balazskegl.substack.comTwitter: https://twitter.com/balazskeglArtwork: DALL-EMusic: Bea Palya https://www.youtube.com/channel/UCBDp3qcFZdU1yoWIRpMSaZw Hosted on Acast. See acast.com/privacy for more information.