Patreon: https://www.patreon.com/mlst Discord: https://discord.gg/ESrGqhf5CB Prof. Melanie Mitchell argues that the concept of "understanding" in AI is ill-defined and multidimensional - we can't simply say an AI system does or doesn't understand. She advocates for rigorously testing AI systems' capabilities using proper experimental methods from cognitive science. Popular benchmarks for intelligence often rely on the assumption that if a human can perform a task, an AI that performs the task must have human-like general intelligence. But benchmarks should evolve as capabilities improve. Large language models show surprising skill on many human tasks but lack common sense and fail at simple things young children can do. Their knowledge comes from statistical relationships in text, not grounded concepts about the world. We don't know if their internal representations actually align with human-like concepts. More granular testing focused on generalization is needed. There are open questions around whether large models' abilities constitute a fundamentally different non-human form of intelligence based on vast statistical correlations across text. Mitchell argues intelligence is situated, domain-specific and grounded in physical experience and evolution. The brain computes but in a specialized way honed by evolution for controlling the body. Extracting "pure" intelligence may not work. Other key points: - Need more focus on proper experimental method in AI research. Developmental psychology offers examples for rigorous testing of cognition. - Reporting instance-level failures rather than just aggregate accuracy can provide insights. - Scaling laws and complex systems science are an interesting area of complexity theory, with applications to understanding cities. - Concepts like "understanding" and "intelligence" in AI force refinement of fuzzy definitions. - Human intelligence may be more collective and social than we realize. AI forces us to rethink concepts we apply anthropomorphically. The overall emphasis is on rigorously building the science of machine cognition through proper experimentation and benchmarking as we assess emerging capabilities. TOC: [00:00:00] Introduction and Munk AI Risk Debate Highlights [05:00:00] Douglas Hofstadter on AI Risk [00:06:56] The Complexity of Defining Intelligence [00:11:20] Examining Understanding in AI Models [00:16:48] Melanie's Insights on AI Understanding Debate [00:22:23] Unveiling the Concept Arc [00:27:57] AI Goals: A Human vs Machine Perspective [00:31:10] Addressing the Extrapolation Challenge in AI [00:36:05] Brain Computation: The Human-AI Parallel [00:38:20] The Arc Challenge: Implications and Insights [00:43:20] The Need for Detailed AI Performance Reporting [00:44:31] Exploring Scaling in Complexity Theory Eratta: Note Tim said around 39 mins that a recent Stanford/DM paper modelling ARC “on GPT-4 got around 60%”. This is not correct and he misremembered. It was actually davinci3, and around 10%, which is still extremely good for a blank slate approach with an LLM and no ARC specific knowledge. Folks on our forum couldn’t reproduce the result. See paper linked below. Books (MUST READ): Artificial Intelligence: A Guide for Thinking Humans (Melanie Mitchell) https://www.amazon.co.uk/Artificial-Intelligence-Guide-Thinking-Humans/dp/B07YBHNM1C/?&_encoding=UTF8&tag=mlst00-21&linkCode=ur2&linkId=44ccac78973f47e59d745e94967c0f30&camp=1634&creative=6738 Complexity: A Guided Tour (Melanie Mitchell) https://www.amazon.co.uk/Audible-Complexity-A-Guided-Tour?&_encoding=UTF8&tag=mlst00-21&linkCode=ur2&linkId=3f8bd505d86865c50c02dd7f10b27c05&camp=1634&creative=6738 Show notes (transcript, full references etc) https://atlantic-papyrus-d68.notion.site/Melanie-Mitchell-2-0-15e212560e8e445d8b0131712bad3000?pvs=25 YT version: https://youtu.be/29gkDpR2orc

We explore connections between FEP and enactivism, including tensions raised in a paper critiquing FEP from an enactivist perspective. Dr. Maxwell Ramstead provides background on enactivism emerging from autopoiesis, with a focus on embodied cognition and rejecting information processing/computational views of mind. Chris shares his journey from robotics into FEP, starting as a skeptic but becoming convinced it's the right framework. He notes there are both "high road" and "low road" versions, ranging from embodied to more radically anti-representational stances. He doesn't see a definitive fork between dynamical systems and information theory as the source of conflict. Rather, the notion of operational closure in enactivism seems to be the main sticking point. The group explores definitional issues around structure/organization, boundaries, and operational closure. Maxwell argues the generative model in FEP captures organizational dependencies akin to operational closure. The Markov blanket formalism models structural interfaces. We discuss the concept of goals in cognitive systems - Chris advocates an intentional stance perspective - using notions of goals/intentions if they help explain system dynamics. Goals emerge from beliefs about dynamical trajectories. Prof Friston provides an elegant explanation of how goal-directed behavior naturally falls out of the FEP mathematics in a particular "goldilocks" regime of system scale/dynamics. The conversation explores the idea that many systems simply act "as if" they have goals or models, without necessarily possessing explicit representations. This helps resolve tensions between enactivist and computational perspectives. Throughout the dialogue, Maxwell presses philosophical points about the FEP abolishing what he perceives as false dichotomies in cognitive science such as internalism/externalism. He is critical of enactivists' commitment to bright line divides between subject areas. Prof. Karl Friston - Inventor of the free energy principle https://scholar.google.com/citations?user=q_4u0aoAAAAJ Prof. Chris Buckley - Professor of Neural Computation at Sussex University https://scholar.google.co.uk/citations?user=nWuZ0XcAAAAJ&hl=en Dr. Maxwell Ramstead - Director of Research at VERSES https://scholar.google.ca/citations?user=ILpGOMkAAAAJ&hl=fr We address critique in this paper: Laying down a forking path: Tensions between enaction and the free energy principle (Ezequiel A. Di Paolo, Evan Thompson, Randall D. Beere) https://philosophymindscience.org/index.php/phimisci/article/download/9187/8975 Other refs: Multiscale integration: beyond internalism and externalism (Maxwell J D Ramstead) https://pubmed.ncbi.nlm.nih.gov/33627890/ MLST panel: Dr. Tim Scarfe and Dr. Keith Duggar TOC (auto generated): 0:00 - Introduction 0:41 - Defining enactivism and its variants 6:58 - The source of the conflict between dynamical systems and information theory 8:56 - Operational closure in enactivism 10:03 - Goals and intentions 12:35 - The link between dynamical systems and information theory 15:02 - Path integrals and non-equilibrium dynamics 18:38 - Operational closure defined 21:52 - Structure vs. organization in enactivism 24:24 - Markov blankets as interfaces 28:48 - Operational closure in FEP 30:28 - Structure and organization again 31:08 - Dynamics vs. information theory 33:55 - Goals and intentions emerge in the FEP mathematics 36:58 - The Good Regulator Theorem 49:30 - enactivism and its relation to ecological psychology 52:00 - Goals, intentions and beliefs 55:21 - Boundaries and meaning 58:55 - Enactivism's rejection of information theory 1:02:08 - Beliefs vs goals 1:05:06 - Ecological psychology and FEP 1:08:41 - The Good Regulator Theorem 1:18:38 - How goal-directed behavior emerges 1:23:13 - Ontological vs metaphysical boundaries 1:25:20 - Boundaries as maps 1:31:08 - Connections to the maximum entropy principle 1:33:45 - Relations to quantum and relational physics

Please check out Numerai - our sponsor @ http://numer.ai/mlst Patreon: https://www.patreon.com/mlst Discord: https://discord.gg/ESrGqhf5CB The Second Law: Resolving the Mystery of the Second Law of Thermodynamics Buy Stephen's book here - https://tinyurl.com/2jj2t9wa The Language Game: How Improvisation Created Language and Changed the World by Morten H. Christiansen and Nick Chater Buy here: https://tinyurl.com/35bvs8be Stephen Wolfram starts by discussing the second law of thermodynamics - the idea that entropy, or disorder, tends to increase over time. He talks about how this law seems intuitively true, but has been difficult to prove. Wolfram outlines his decades-long quest to fully understand the second law, including failed early attempts to simulate particles mixing as a 12-year-old. He explains how irreversibility arises from the computational irreducibility of underlying physical processes coupled with our limited ability as observers to do the computations needed to "decrypt" the microscopic details. The conversation then shifts to discussing language and how concepts allow us to communicate shared ideas between minds positioned in different parts of "rule space." Wolfram talks about the successes and limitations of using large language models to generate Wolfram Language code from natural language prompts. He sees it as a useful tool for getting started programming, but one still needs human refinement. The final part of the conversation focuses on AI safety and governance. Wolfram notes uncontrolled actuation is where things can go wrong with AI systems. He discusses whether AI agents could have intrinsic experiences and goals, how we might build trust networks between AIs, and that managing a system of many AIs may be easier than a single AI. Wolfram emphasizes the need for more philosophical depth in thinking about AI aims, and draws connections between potential solutions and his work on computational irreducibility and physics. Show notes: https://docs.google.com/document/d/1hXNHtvv8KDR7PxCfMh9xOiDFhU3SVDW8ijyxeTq9LHo/edit?usp=sharing Pod version: TBA https://twitter.com/stephen_wolfram TOC: 00:00:00 - Introduction 00:02:34 - Second law book 00:14:01 - Reversibility / entropy / observers / equivalence 00:34:22 - Concepts/language in the ruliad 00:49:04 - Comparison to free energy principle 00:53:58 - ChatGPT / Wolfram / Language 01:00:17 - AI risk Panel: Dr. Tim Scarfe @ecsquendor / Dr. Keith Duggar @DoctorDuggar

Please check out Numerai - our sponsor @ http://numer.ai/mlst Patreon: https://www.patreon.com/mlst Discord: https://discord.gg/ESrGqhf5CB Professor Jürgen Schmidhuber, the father of artificial intelligence, joins us today. Schmidhuber discussed the history of machine learning, the current state of AI, and his career researching recursive self-improvement, artificial general intelligence and its risks. Schmidhuber pointed out the importance of studying the history of machine learning to properly assign credit for key breakthroughs. He discussed some of the earliest machine learning algorithms. He also highlighted the foundational work of Leibniz, who discovered the chain rule that enables training of deep neural networks, and the ancient Antikythera mechanism, the first known gear-based computer. Schmidhuber discussed limits to recursive self-improvement and artificial general intelligence, including physical constraints like the speed of light and what can be computed. He noted we have no evidence the human brain can do more than traditional computing. Schmidhuber sees humankind as a potential stepping stone to more advanced, spacefaring machine life which may have little interest in humanity. However, he believes commercial incentives point AGI development towards being beneficial and that open-source innovation can help to achieve "AI for all" symbolised by his company's motto "AI∀". Schmidhuber discussed approaches he believes will lead to more general AI, including meta-learning, reinforcement learning, building predictive world models, and curiosity-driven learning. His "fast weight programming" approach from the 1990s involved one network altering another network's connections. This was actually the first Transformer variant, now called an unnormalised linear Transformer. He also described the first GANs in 1990, to implement artificial curiosity. Schmidhuber reflected on his career researching AI. He said his fondest memories were gaining insights that seemed to solve longstanding problems, though new challenges always arose: "then for a brief moment it looks like the greatest thing since sliced bread and and then you get excited ... but then suddenly you realize, oh, it's still not finished. Something important is missing.” Since 1985 he has worked on systems that can recursively improve themselves, constrained only by the limits of physics and computability. He believes continual progress, shaped by both competition and collaboration, will lead to increasingly advanced AI. On AI Risk: Schmidhuber: "To me it's indeed weird. Now there are all these letters coming out warning of the dangers of AI. And I think some of the guys who are writing these letters, they are just seeking attention because they know that AI dystopia are attracting more attention than documentaries about the benefits of AI in healthcare." Schmidhuber believes we should be more concerned with existing threats like nuclear weapons than speculative risks from advanced AI. He said: "As far as I can judge, all of this cannot be stopped but it can be channeled in a very natural way that is good for humankind...there is a tremendous bias towards good AI, meaning AI that is good for humans...I am much more worried about 60 year old technology that can wipe out civilization within two hours, without any AI.” [this is truncated, read show notes] YT: https://youtu.be/q27XMPm5wg8 Show notes: https://docs.google.com/document/d/13-vIetOvhceZq5XZnELRbaazpQbxLbf5Yi7M25CixEE/edit?usp=sharing Note: Interview was recorded 15th June 2023. https://twitter.com/SchmidhuberAI Panel: Dr. Tim Scarfe @ecsquendor / Dr. Keith Duggar @DoctorDuggar Pod version: TBA TOC: [00:00:00] Intro / Numerai [00:00:51] Show Kick Off [00:02:24] Credit Assignment in ML [00:12:51] XRisk [00:20:45] First Transformer variant of 1991 [00:47:20] Which Current Approaches are Good [00:52:42] Autonomy / Curiosity [00:58:42] GANs of 1990 [01:11:29] OpenAI, Moats, Legislation

Patreon: https://www.patreon.com/mlst Discord: https://discord.gg/ESrGqhf5CB George Hotz and Connor Leahy discuss the crucial challenge of developing beneficial AI that is aligned with human values. Hotz believes truly aligned AI is impossible, while Leahy argues it's a solvable technical challenge.Hotz contends that AI will inevitably pursue power, but distributing AI widely would prevent any single AI from dominating. He advocates open-sourcing AI developments to democratize access. Leahy counters that alignment is necessary to ensure AIs respect human values. Without solving alignment, general AI could ignore or harm humans.They discuss whether AI's tendency to seek power stems from optimization pressure or human-instilled goals. Leahy argues goal-seeking behavior naturally emerges while Hotz believes it reflects human values. Though agreeing on AI's potential dangers, they differ on solutions. Hotz favors accelerating AI progress and distributing capabilities while Leahy wants safeguards put in place.While acknowledging risks like AI-enabled weapons, they debate whether broad access or restrictions better manage threats. Leahy suggests limiting dangerous knowledge, but Hotz insists openness checks government overreach. They concur that coordination and balance of power are key to navigating the AI revolution. Both eagerly anticipate seeing whose ideas prevail as AI progresses. Transcript and notes: https://docs.google.com/document/d/1smkmBY7YqcrhejdbqJOoZHq-59LZVwu-DNdM57IgFcU/edit?usp=sharing Note: this is not a normal episode i.e. the hosts are not part of the debate (and for the record don't agree with Connor or George). TOC: [00:00:00] Introduction to George Hotz and Connor Leahy [00:03:10] George Hotz's Opening Statement: Intelligence and Power [00:08:50] Connor Leahy's Opening Statement: Technical Problem of Alignment and Coordination [00:15:18] George Hotz's Response: Nature of Cooperation and Individual Sovereignty [00:17:32] Discussion on individual sovereignty and defense [00:18:45] Debate on living conditions in America versus Somalia [00:21:57] Talk on the nature of freedom and the aesthetics of life [00:24:02] Discussion on the implications of coordination and conflict in politics [00:33:41] Views on the speed of AI development / hard takeoff [00:35:17] Discussion on potential dangers of AI [00:36:44] Discussion on the effectiveness of current AI [00:40:59] Exploration of potential risks in technology [00:45:01] Discussion on memetic mutation risk [00:52:36] AI alignment and exploitability [00:53:13] Superintelligent AIs and the assumption of good intentions [00:54:52] Humanity’s inconsistency and AI alignment [00:57:57] Stability of the world and the impact of superintelligent AIs [01:02:30] Personal utopia and the limitations of AI alignment [01:05:10] Proposed regulation on limiting the total number of flops [01:06:20] Having access to a powerful AI system [01:18:00] Power dynamics and coordination issues with AI [01:25:44] Humans vs AI in Optimization [01:27:05] The Impact of AI's Power Seeking Behavior [01:29:32] A Debate on the Future of AI