Uploaded March 2025 | Updated November 2026, 1 hour ago
Iman Mirzadeh is a machine learning research engineer at Apple and the lead author of the GSM-Symbolic paper, which exposed deep fragility in how large language models handle mathematical reasoning. In this conversation, he draws a sharp line between intelligence and achievement -- between what a system can score on a benchmark and what it actually understands.
The discussion starts with chess. Mirzadeh explains how grandmasters don't use engines to memorize moves; they use them to develop theory. AlphaZero discovered unprecedented strategies, but that knowledge stays trapped in the game. Humans, by contrast, extract abstract principles like 'control the center' and transfer them to entirely different domains. That capacity for abstraction is what he thinks current AI architecturally lacks.
His critique of LLMs is structural. These systems are trained to minimize cross-entropy loss over a distribution, and by construction they cannot reason beyond what that distribution contains. Change the surface form of a problem -- swap names, add irrelevant clauses -- and performance varies wildly, even on grade-school math.
That's the core finding of GSM-Symbolic. By generating templated variants of math word problems, Mirzadeh's team showed that even frontier models exhibit large performance variance from changes that should be semantically irrelevant. The implication: what looks like reasoning is closer to sophisticated pattern matching across memorized distributions.
Mirzadeh proposes that intelligence should be measured by the slope of a system's scaling -- how fast it can learn novel things -- rather than its current benchmark position. The conversation also covers the connectionism-symbolism divide, active engagement and agency as necessary conditions for learning, and why we might need a fundamentally different vessel to reach genuine reasoning.
SPONSOR MESSAGES:
***
Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. They are hiring a Chief Engineer and ML engineers. Events in Zurich.
---
TIMESTAMPS:
00:00:00 Intelligence vs Achievement in AI
00:03:27 AlphaZero and Abstract Understanding in Chess
00:10:10 Language Models as Distribution Learners
00:14:47 The State of AI Research Methodology
00:24:24 Interpolation vs True Reasoning in LLMs
00:29:00 Measuring Intelligence: From Chollet to the Iman Moon Test
00:35:35 Agency, Active Learning, and World Models
00:47:15 Scaling Laws and the Connectionism-Symbolism Debate
00:58:09 GSM-Symbolic: Exposing LLM Reasoning Fragility
---
REFERENCES:
paper:
[00:00:55] Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
arxiv.org/abs/1712.01815
[00:17:15] GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models
arxiv.org/abs/2410.05229
[00:21:20] Connectionism and Cognitive Architecture: A Critical Analysis
sciencedirect.com/science/article/pii/001002779090014B
[00:29:35] On the Measure of Intelligence
arxiv.org/abs/1911.01547
[00:33:25] On definition of intelligence
sciencedirect.com/science/article/pii/S0160289624000266
[00:35:25] Defining Intelligence
https://cis.temple.edu/~wangp/papers.html
[00:43:10] Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
arxiv.org/abs/2201.11903
[00:47:45] Scaling Laws for Neural Language Models
arxiv.org/abs/2001.08361
[00:55:10] Tensor Product Variable Binding and the Representation of Symbolic Structures in Connectionist Systems
sciencedirect.com/science/article/abs/pii/000437029090007M
book:
[00:07:05] Game Changer: AlphaZero's Groundbreaking Chess Strategies
amazon.com/Game-Changer-AlphaZeros-Groundbreaking-Strategies/dp/9056918184
[00:37:35] How We Learn: Why Brains Learn Better Than Any Machine... for Now
amazon.com/How-We-Learn-Brains-Machine/dp/0525559884
[00:39:30] Surfaces and Essences: Analogy as the Fuel and Fire of Thinking
amazon.com/Surfaces-Essences-Analogy-Fuel-Thinking/dp/0465018475
reference:
[00:11:30] NLP Course: Language Modeling
lena-voita.github.io/nlp_course/language_modeling.html
[01:08:40] GSM8K: Training Verifiers to Solve Math Word Problems
huggingface.co/datasets/openai/gsm8k
---
LINKS:
Full Transcript: app.rescript.info/share/72689965572c7fd460954f70e26b2eaa
Download PDF transcript: app.rescript.info/api/public/sessions/abf4f57b59fcbbd5/pdf
Iman Mirzadeh is a machine learning research engineer at Apple and the lead author of the GSM-Symbolic paper, which exposed deep fragility in how large language models handle mathematical reasoning. In this conversation, he draws a sharp line between intelligence and achievement -- between what a system can score on a benchmark and what it actually understands.
The discussion starts with chess. Mirzadeh explains how grandmasters don't use engines to memorize moves; they use them to develop theory. AlphaZero discovered unprecedented strategies, but that knowledge stays trapped in the game. Humans, by contrast, extract abstract principles like 'control the center' and transfer them to entirely different domains. That capacity for abstraction is what he thinks current AI architecturally lacks.
His critique of LLMs is structural. These systems are trained to minimize cross-entropy loss over a distribution, and by construction they cannot reason beyond what that distribution contains. Change the surface form of a problem -- swap names, add irrelevant clauses -- and performance varies wildly, even on grade-school math.
That's the core finding of GSM-Symbolic. By generating templated variants of math word problems, Mirzadeh's team showed that even frontier models exhibit large performance variance from changes that should be semantically irrelevant. The implication: what looks like reasoning is closer to sophisticated pattern matching across memorized distributions.
Mirzadeh proposes that intelligence should be measured by the slope of a system's scaling -- how fast it can learn novel things -- rather than its current benchmark position. The conversation also covers the connectionism-symbolism divide, active engagement and agency as necessary conditions for learning, and why we might need a fundamentally different vessel to reach genuine reasoning.
SPONSOR MESSAGES:
***
Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. They are hiring a Chief Engineer and ML engineers. Events in Zurich.
---
TIMESTAMPS:
00:00:00 Intelligence vs Achievement in AI
00:03:27 AlphaZero and Abstract Understanding in Chess
00:10:10 Language Models as Distribution Learners
00:14:47 The State of AI Research Methodology
00:24:24 Interpolation vs True Reasoning in LLMs
00:29:00 Measuring Intelligence: From Chollet to the Iman Moon Test
00:35:35 Agency, Active Learning, and World Models
00:47:15 Scaling Laws and the Connectionism-Symbolism Debate
00:58:09 GSM-Symbolic: Exposing LLM Reasoning Fragility
---
REFERENCES:
paper:
[00:00:55] Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
arxiv.org/abs/1712.01815
[00:17:15] GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models
arxiv.org/abs/2410.05229
[00:21:20] Connectionism and Cognitive Architecture: A Critical Analysis
sciencedirect.com/science/article/pii/001002779090014B
[00:29:35] On the Measure of Intelligence
arxiv.org/abs/1911.01547
[00:33:25] On definition of intelligence
sciencedirect.com/science/article/pii/S0160289624000266
[00:35:25] Defining Intelligence
https://cis.temple.edu/~wangp/papers.html
[00:43:10] Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
arxiv.org/abs/2201.11903
[00:47:45] Scaling Laws for Neural Language Models
arxiv.org/abs/2001.08361
[00:55:10] Tensor Product Variable Binding and the Representation of Symbolic Structures in Connectionist Systems
sciencedirect.com/science/article/abs/pii/000437029090007M
book:
[00:07:05] Game Changer: AlphaZero's Groundbreaking Chess Strategies
amazon.com/Game-Changer-AlphaZeros-Groundbreaking-Strategies/dp/9056918184
[00:37:35] How We Learn: Why Brains Learn Better Than Any Machine... for Now
amazon.com/How-We-Learn-Brains-Machine/dp/0525559884
[00:39:30] Surfaces and Essences: Analogy as the Fuel and Fire of Thinking
amazon.com/Surfaces-Essences-Analogy-Fuel-Thinking/dp/0465018475
reference:
[00:11:30] NLP Course: Language Modeling
lena-voita.github.io/nlp_course/language_modeling.html
[01:08:40] GSM8K: Training Verifiers to Solve Math Word Problems
huggingface.co/datasets/openai/gsm8k
---
LINKS:
Full Transcript: app.rescript.info/share/72689965572c7fd460954f70e26b2eaa
Download PDF transcript: app.rescript.info/api/public/sessions/abf4f57b59fcbbd5/pdf


![Why AI Has a Plato Problem — Mazviita Chirimuuta
Professor Mazviita Chirimuuta joins us for a fascinating deep dive into the philosophy of neuroscience and what it really means to understand the mind.
*What can neuroscience actually tell us about how the mind works?* In this thought-provoking conversation, we explore the hidden assumptions behind computational theories of the brain, the limits of scientific abstraction, and why the question of machine consciousness might be more complicated than AI researchers assume.
Mazviita, author of *The Brain Abstracted,* brings a unique perspective shaped by her background in both neuroscience research and philosophy. She challenges us to think critically about the metaphors we use to understand cognition — from the reflex theory of the late 19th century to todays dominant view of the brain as a computer.
*Key topics explored:*
*The problem of oversimplification* — Why scientific models necessarily leave things out, and how this can sometimes lead entire fields astray. The cautionary tale of reflex theory shows how elegant explanations can blind us to biological complexity.
*Is the brain really a computer?* — Mazviita unpacks the philosophical assumptions behind computational neuroscience and asks: if we can model anything computationally, what makes brains special? The answer might challenge everything you thought you knew about AI.
*Haptic realism* — A fresh way of thinking about scientific knowledge that emphasizes interaction over passive observation. Knowledge isnt about reading the source code of the universe — its something we actively construct through engagement with the world.
*Why embodiment matters for understanding* — Can a disembodied language model truly understand? Mazviita makes a compelling case that human cognition is deeply entangled with our sensory-motor engagement and biological existence in ways that cant simply be abstracted away.
*Technology and human finitude* — Drawing on Heidegger, we discuss how the dream of transcending our physical limitations through technology might reflect a fundamental misunderstanding of what it means to be a knower.
This conversation is essential viewing for anyone interested in AI, consciousness, philosophy of mind, or the future of cognitive science. Whether youre skeptical of strong AI claims or a true believer in machine consciousness, Mazviitas careful philosophical analysis will give you new tools for thinking through these profound questions.
TIMESTAMPS:
00:00:00 The Problem of Generalizing Neuroscience
00:02:51 Abstraction vs. Idealization: The Kaleidoscope
00:05:39 Platonism in AI: Discovering or Inventing Patterns?
00:09:42 When Simplification Fails: The Reflex Theory
00:12:23 Behaviorism and the Black Box Trap
00:14:20 Haptic Realism: Knowledge Through Interaction
00:20:23 Is Nature Protean? The Myth of Converging Truth
00:23:23 The Computational Theory of Mind: A Useful Fiction?
00:27:25 Biological Constraints: Why Brains Arent Just Neural Nets
00:31:01 Agency, Distal Causes, and Dennetts Stances
00:37:13 Searles Challenge: Causal Powers and Understanding
00:41:58 Heideggers Warning & The Experiment on Children
REFERENCES:
Book:
[00:01:28] The Brain Abstracted
https://mitpress.mit.edu/9780262548045/the-brain-abstracted/
[00:11:05] The Integrated Action of the Nervous System
https://www.amazon.sg/integrative-action-nervous-system/dp/9354179029
[00:18:15] The Quest for Certainty (Dewey)
https://www.amazon.com/Quest-Certainty-Relation-Knowledge-Lectures/dp/0399501916
[00:19:45] Realism for Realistic People (Chang)
https://www.cambridge.org/core/books/realism-for-realistic-people/ACC93A7F03B15AA4D6F3A466E3FC5AB7
[00:38:15] The Rediscovery of the Mind (Searle)
https://mitpress.mit.edu/9780262691543/the-rediscovery-of-the-mind/
[00:47:18] So Youve Been Publicly Shamed (Ronson)
https://www.amazon.com/So-Youve-Been-Publicly-Shamed/dp/1594634017
[00:50:30] Reality+ (Chalmers)
https://consc.net/reality/
Person:
[00:05:00] Francois Chollet
https://arcprize.org/
Paper:
[00:08:03] Real Patterns (Dennett)
https://ruccs.rutgers.edu/images/personal-zenon-pylyshyn/class-info/FP2012/FP2012_readings/Dennett_RealPatterns.pdf
[00:25:30] A Logical Calculus of Ideas... (1943)
https://link.springer.com/article/10.1007/BF02478259
[00:29:17] The Lottery Ticket Hypothesis
https://arxiv.org/abs/1803.03635
Philosophy:
[00:17:30] Transcendental Idealism (Kant)
https://plato.stanford.edu/entries/kant-transcendental-idealism/
RESCRIPT:
https://app.rescript.info/public/share/A6cZ1TY35p8ORMmYCWNBI0no9ChU3-Kx7dPXGJURvZ0
PDF Transcript:
https://app.rescript.info/api/public/sessions/0fb7767e066cf712/pdf Why AI Has a Plato Problem — Mazviita Chirimuuta](https://i.ytimg.com/vi/yq318DIwPqw/mqdefault.jpg)
![How Many Narrow AIs Could Behave Like One Superintelligence - Daniel Kokotajlo and Thomas Larsen
Could slowing AI development make superintelligence safer? Daniel Kokotajlo and Thomas Larsen of the AI Futures Project join Tim Scarfe to examine AI 2040: Plan A, a proposal to buy time before AI exceeds human control.
SPONSOR:
Cyber Fund built the Monastery to help founders ship products that were impossible a year ago.
Apply now: https://cyber.fund
After revisiting AI 2027 and the limits of forecasting, they ask what happens when AI can automate research and sustain an economy without human workers. Tim challenges the case for general models and asks whether intelligence alone explains power. Plan A proposes an initial pause to build safety infrastructure, then cautious development up to the strongest AI that can still be reliably controlled. The discussion tests the distinction between control and alignment, the case for public AI research, and whether the US and China could enforce a slowdown. It ends with the evidence that would change their forecasts.
TIMESTAMPS:
00:00:00 AI 2040: a slower route to superintelligence
00:01:34 Sponsor: Cyber Fund
00:02:12 From OpenAI to AI 2027
00:06:58 Forecasts, war games and self-fulfilling prophecies
00:17:44 Why AI sceptics are changing their minds
00:23:04 When AI can replace its own researchers
00:28:45 Could an AI economy grow without human workers?
00:37:32 One general model or a society of specialists?
00:47:43 Brains, machines and collective intelligence
00:56:12 Plan A: buy time at the controllable frontier
01:00:02 Why control buys time but cannot replace alignment
01:06:36 Why AI research should be public
01:10:32 Can the US and China enforce an AI slowdown?
01:19:04 Why AI policy debates miss the technology
01:21:56 Is AI normal technology? The remaining disagreement
Many thanks to James Wilken-Smith for helping with show research.
REFERENCES:
other:
[00:00:01] AI 2040: Plan A
https://ai-2040.com/
[00:03:27] AI 2027
https://ai-2027.com/
[00:13:47] Scenario Scrutiny for AI Policy
https://blog.aifutures.org/p/scenario-scrutiny-for-ai-policy
[00:33:11] The 2028 Global Intelligence Crisis
https://www.citriniresearch.com/p/2028gic
[01:00:40] Brief independent investigation of agents behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident
https://www.redwoodresearch.org/research/hugging-face-incident
[01:09:21] The Hugging Face incident and the road ahead
https://openai.com/index/hugging-face-incident-and-the-road-ahead/
[01:22:01] AI as Normal Technology
https://www.normaltech.ai/p/ai-as-normal-technology
[01:22:51] Common Ground between AI 2027 & AI as Normal Technology
https://asteriskmag.substack.com/p/common-ground-between-ai-2027-and
person:
[00:19:43] Geoffrey Hinton
https://www.cs.toronto.edu/~hinton/
[00:20:07] Ryan Greenblatt
https://www.lesswrong.com/users/ryan_greenblatt
[00:26:06] Elon Musk
https://www.tesla.com/elon-musk
tool:
[00:21:46] ARC-AGI-3
https://arcprize.org/arc-agi/3
[00:21:53] AlphaGo and Move 37
https://deepmind.google/research/alphago/
[00:39:41] Claude
https://claude.com/product/overview
[00:39:58] NVIDIA H100 GPU
https://www.nvidia.com/en-us/data-center/h100/
paper:
[00:24:42] Training AI Scientists to Replicate Research
https://arxiv.org/abs/2608.13331v1
[01:27:19] Validity of the single processor approach to achieving large scale computing capabilities
https://www.cs.cmu.edu/~18742/papers/Amdahl1967.pdf
book:
[00:28:52] Bullshit Jobs: A Theory
https://www.simonandschuster.com/books/Bullshit-Jobs/David-Graeber/9781501143335
organization:
[01:05:09] Redwood Research
https://www.redwoodresearch.org/
RESCRIPT:
https://app.rescript.info/public/share/33d1a58fa8f307ae7dfd504d4fdaa9d5 How Many Narrow AIs Could Behave Like One Superintelligence - Daniel Kokotajlo and Thomas Larsen](https://i.ytimg.com/vi/z5Xix4h5UlU/mqdefault.jpg)


![We need AIs with PHYSICAL experience (Jeff Beck)
Jeff Beck spent years studying computational neuroscience with Alexandre Pouget, Peter Latham, and Wei Ji Ma before founding Noumenal Labs. In this conversation with Tim Scarfe, he argues that language models are fundamentally limited because they manipulate symbols without the physical grounding that gives those symbols meaning.
Beck walks through the Bayesian brain hypothesis — the idea that our brains maintain probabilistic models of the world, built from direct sensory experience. Language, he says, is just a thin lossy summary of that richer internal model. He points to Markus Meisters work showing that human information output runs at roughly 10 bits per second, a tiny fraction of what comes in.
The conversation gets interesting around the question of whether AI systems can genuinely understand anything. Becks position is clear: until a language model produces something genuinely novel — not recombined from training data — he wont call it intelligent. The second half turns to alignment — Beck argues that you cannot separate someones beliefs from their reward function just by observing their behavior, a claim with serious implications for AI safety.
SPONSOR MESSAGES:
***
Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. They are hiring a Chief Engineer and ML engineers. Events in Zurich.
Goto https://tufalabs.ai/
***
TIMESTAMPS:
00:00:00 Bayesian brain and neural computation foundations
00:02:00 Belief-reward entanglement and AI alignment
00:02:19 Sponsor: Tufa Labs
00:02:45 Visual cortex discovery and the Hubel-Wiesel experiment
00:05:00 Theory of mind tests and ChatGPT limitations
00:08:50 The sum and product riddle: pattern recognition vs reasoning
00:09:30 Systems engineering and object decomposition
00:12:20 Vicarious experience and grounded understanding
00:15:20 Neural coding and choice of generative model
00:17:30 Line-of-sight legibility in AI reasoning
00:18:50 Language as lossy compression of cognition
00:22:50 10 bits per second: information processing bottleneck
00:25:00 Why language models cannot substitute for understanding
00:28:00 Scientific abstraction and idealization
00:29:40 Markov blankets and system partitioning
00:33:20 Scientific realism, noise, and the limits of models
00:36:00 Free energy principle as mathematical framework
00:38:40 Black-box prediction vs legible explanation
00:41:00 Reward functions and the impossibility of alignment
00:45:00 Modeling beliefs as prerequisite for value inference
REFERENCES:
paper:
[00:00:15] Bayesian inference in neural computation (Ma, Beck, Latham, Pouget)
https://www.nature.com/articles/nn1790
[00:01:50] Noumenal Labs research paper
https://arxiv.org/html/2502.13161v1
[00:05:25] LLM performance on theory of mind tasks (Kosinski)
https://arxiv.org/abs/2302.02083
[00:16:25] Bayesian Mechanics (Ramstead, Sakthivadivel, Heins et al.)
https://arxiv.org/abs/2205.11543
[00:17:10] Building Human-like Communicative Intelligence (Dubova)
https://arxiv.org/abs/2201.02734
[00:20:45] Why do we live at 10 bits/s? (Meister, Zheng)
https://www.sciencedirect.com/science/article/abs/pii/S0896627324008080
[00:28:30] Markov blankets in biological systems (Friston)
https://royalsocietypublishing.org/doi/10.1098/rsif.2017.0792
[00:32:15] Critique of Gabor patches in neuroscience (Tsao)
https://pmc.ncbi.nlm.nih.gov/articles/PMC9564096/
[00:35:55] MaxEnt and MaxCal principles (Presse et al.)
https://journals.aps.org/rmp/abstract/10.1103/RevModPhys.85.1115
[00:40:08] Reward is Enough (Silver, Singh, Precup, Sutton)
https://www.sciencedirect.com/science/article/pii/S0004370221000862
[00:45:40] Modeling Human Beliefs about AI Behavior (Lang, Forre)
https://www.arxiv.org/pdf/2502.21262
website:
[00:01:50] Noumenal Labs (Jeff Beck)
https://www.noumenal.ai/
[00:02:19] Tufa AI Labs
https://tufalabs.ai/
[00:22:25] Steven Piantadosi
https://colala.berkeley.edu/people/piantadosi/
[00:23:35] Mad Libs (Stern, Price)
https://en.wikipedia.org/wiki/Mad_Libs
[00:31:05] Mathematical Platonism (Linnebo, SEP)
https://plato.stanford.edu/entries/platonism-mathematics/
book:
[00:25:25] The Brain Abstracted (Chirimuuta)
https://mitpress.mit.edu/9780262548045/the-brain-abstracted/
LINKS:
Full Transcript: https://app.rescript.info/share/75dfad07d8cad3ed6ccd03253aa006be
Download PDF transcript: https://app.rescript.info/api/public/sessions/981caa3aeba083df/pdf
Extended version on patreon:
https://www.patreon.com/posts/jeff-beck-125455115 We need AIs with PHYSICAL experience (Jeff Beck)](https://i.ytimg.com/vi/zv6qzWecj5c/mqdefault.jpg)