Uploaded November 2024 | Updated November 2026, 1 hour ago
Jonas Hübotter from ETH presents SIFT (Select Informative data for Fine-Tuning), a breakthrough algorithm that dramatically improves language model performance through test-time adaptation. Using intelligent data selection, SIFT achieves state-of-the-art results with a 3.8B parameter model - 30x smaller than previous approaches. The system combines a parametric controller with non-parametric memory to optimize training example selection, showing impressive results across mathematics, coding, and legal domains. This novel approach points toward more efficient and adaptable AI systems that can continuously improve through interaction.
This was the first physical meetup of Tufa AI Labs, are you an ML researcher interested in joining or presenting at one of these sessions? Please get in touch with Benjamin Crouzier benjamin@tufa.ai - tufalabs.ai
SLIDES:
dropbox.com/scl/fi/sys3iasc63lgj8lm5t0ld/JONAS_SLIDES.pdf?rlkey=ak6ir61a2pyhrfuwyvgrdvq66&st=9cloopv9&dl=0
SPONSOR MESSAGE:
CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments.
centml.ai/pricing
Jonas Hübotter
Doctoral Researcher at ETH Zurich working on Active Fine-Tuning and Local Learning.
jonhue.github.io
Test-Time Training on Nearest Neighbors for Large Language Models
arxiv.org/abs/2305.18466 (IMPORTANT BACKGROUND READING)
TOC:
1. SIFT Algorithm Core Concepts
[00:00:00] 1.1 Introduction to Test-Time Adaptation and SIFT Algorithm
[00:02:45] 1.2 The Pile Benchmark and Parameter Efficiency
[00:07:00] 1.3 Local Learning Models and Vapnik's Principle
[00:12:33] 1.4 SIFT Performance and Domain-Specific Comparisons
2. Training and Data Selection Methods
[00:22:50] 2.1 Data Selection and Error Measurement Methods
[00:32:33] 2.2 Non-IID Training Experiments on MNIST
3. Scaling and Implementation and Audience QnA
[00:35:50] 3.1 Scaling Experiments to Larger Datasets and Models
[00:42:30] 3.2 Model Scaling and Performance Across Architectures
[00:44:25] 3.3 Exploration-Exploitation Trade-offs in Fine-tuning
[00:47:54] 3.4 Two-Stage Local Learning Architecture and SIFT Implementation
SHOWNOTES (transcript, references, best quotes etc):
dropbox.com/scl/fi/os3ny3sy446u07yldz0zg/JONAS_PRESENTS.pdf?rlkey=kmu2pxfx8xbmiy283diof0kj1&st=xi5ouc9t&dl=0
REFS:
[0:00:25] Paper: 'Efficiently Learning at Test-Time: Active Fine-Tuning of LLMs' introducing SIFT algorithm for optimizing LLM performance through test-time fine-tuning (Jonas Hübotter, Sascha Bongni, Ido Hakimi, Andreas Krause)
arxiv.org/pdf/2410.08020
[0:02:45] The Pile: An 800GB Dataset of Diverse Text for Language Modeling - A comprehensive dataset comprising 22 diverse high-quality subsets for training large-scale language models (Leo Gao et al.)
arxiv.org/abs/2101.00027
[0:03:20] Language Models are Unsupervised Multitask Learners (GPT-2) -
cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
[0:11:05] Vladimir Vapnik's principle from Statistical Learning Theory: 'When solving a problem of interest, do not solve a more general problem as an intermediate step. Try to get the answer that you really need, but not a more general one.'
amazon.com/Statistical-Learning-Information-Science-Statistics/dp/0387987800
[0:22:05] Paper discussed at ICML 'The Linear Representation Hypothesis and the Geometry of Large Language Models' by Kiho Park et al.
arxiv.org/abs/2311.03658
[0:23:20] On choosing and bounding probability metrics - Paper discussing Total Variation (TV) distance and its applications in probability theory (ALISON L. GIBBS AND FRANCIS EDWARD SU)
arxiv.org/pdf/math/0209021
[0:33:25] MNIST dataset - Standard database of handwritten digits containing 60,000 training images and 10,000 test images of size 28x28 pixels (Yann LeCun, Corinna Cortes)
yann.lecun.com/exdb/mnist
[0:35:50] CIFAR-100 dataset - A dataset of 32x32 color images in 100 classes, with 600 images per class (Alex Krizhevsky)
https://www.cs.toronto.edu/~kriz/cifar.html
[0:36:00] ImageNet - Large-scale hierarchical image database with over 14 million images organized according to the WordNet hierarchy (Jia Deng et al)
ieeexplore.ieee.org/document/5206848
[0:42:55] Llama 2: Collection of foundation and fine-tuned chat models ranging from 7B to 70B parameters (Hugo Touvron et al.)
arxiv.org/abs/2307.09288
[0:43:35] Scaling Instruction-Finetuned Language Models - Paper introducing Flan-T5, showing performance improvements through instruction finetuning (Hyung Won Chung et al.)
arxiv.org/abs/2210.11416
[0:45:10] Active Few-Shot Fine-Tuning methodology paper discussing exploration-exploitation trade-offs in the context of fine-tuning neural networks. (Jonas Hübotter et al.)
arxiv.org/abs/2402.15898
Jonas Hübotter from ETH presents SIFT (Select Informative data for Fine-Tuning), a breakthrough algorithm that dramatically improves language model performance through test-time adaptation. Using intelligent data selection, SIFT achieves state-of-the-art results with a 3.8B parameter model - 30x smaller than previous approaches. The system combines a parametric controller with non-parametric memory to optimize training example selection, showing impressive results across mathematics, coding, and legal domains. This novel approach points toward more efficient and adaptable AI systems that can continuously improve through interaction.
This was the first physical meetup of Tufa AI Labs, are you an ML researcher interested in joining or presenting at one of these sessions? Please get in touch with Benjamin Crouzier benjamin@tufa.ai - tufalabs.ai
SLIDES:
dropbox.com/scl/fi/sys3iasc63lgj8lm5t0ld/JONAS_SLIDES.pdf?rlkey=ak6ir61a2pyhrfuwyvgrdvq66&st=9cloopv9&dl=0
SPONSOR MESSAGE:
CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments.
centml.ai/pricing
Jonas Hübotter
Doctoral Researcher at ETH Zurich working on Active Fine-Tuning and Local Learning.
jonhue.github.io
Test-Time Training on Nearest Neighbors for Large Language Models
arxiv.org/abs/2305.18466 (IMPORTANT BACKGROUND READING)
TOC:
1. SIFT Algorithm Core Concepts
[00:00:00] 1.1 Introduction to Test-Time Adaptation and SIFT Algorithm
[00:02:45] 1.2 The Pile Benchmark and Parameter Efficiency
[00:07:00] 1.3 Local Learning Models and Vapnik's Principle
[00:12:33] 1.4 SIFT Performance and Domain-Specific Comparisons
2. Training and Data Selection Methods
[00:22:50] 2.1 Data Selection and Error Measurement Methods
[00:32:33] 2.2 Non-IID Training Experiments on MNIST
3. Scaling and Implementation and Audience QnA
[00:35:50] 3.1 Scaling Experiments to Larger Datasets and Models
[00:42:30] 3.2 Model Scaling and Performance Across Architectures
[00:44:25] 3.3 Exploration-Exploitation Trade-offs in Fine-tuning
[00:47:54] 3.4 Two-Stage Local Learning Architecture and SIFT Implementation
SHOWNOTES (transcript, references, best quotes etc):
dropbox.com/scl/fi/os3ny3sy446u07yldz0zg/JONAS_PRESENTS.pdf?rlkey=kmu2pxfx8xbmiy283diof0kj1&st=xi5ouc9t&dl=0
REFS:
[0:00:25] Paper: 'Efficiently Learning at Test-Time: Active Fine-Tuning of LLMs' introducing SIFT algorithm for optimizing LLM performance through test-time fine-tuning (Jonas Hübotter, Sascha Bongni, Ido Hakimi, Andreas Krause)
arxiv.org/pdf/2410.08020
[0:02:45] The Pile: An 800GB Dataset of Diverse Text for Language Modeling - A comprehensive dataset comprising 22 diverse high-quality subsets for training large-scale language models (Leo Gao et al.)
arxiv.org/abs/2101.00027
[0:03:20] Language Models are Unsupervised Multitask Learners (GPT-2) -
cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
[0:11:05] Vladimir Vapnik's principle from Statistical Learning Theory: 'When solving a problem of interest, do not solve a more general problem as an intermediate step. Try to get the answer that you really need, but not a more general one.'
amazon.com/Statistical-Learning-Information-Science-Statistics/dp/0387987800
[0:22:05] Paper discussed at ICML 'The Linear Representation Hypothesis and the Geometry of Large Language Models' by Kiho Park et al.
arxiv.org/abs/2311.03658
[0:23:20] On choosing and bounding probability metrics - Paper discussing Total Variation (TV) distance and its applications in probability theory (ALISON L. GIBBS AND FRANCIS EDWARD SU)
arxiv.org/pdf/math/0209021
[0:33:25] MNIST dataset - Standard database of handwritten digits containing 60,000 training images and 10,000 test images of size 28x28 pixels (Yann LeCun, Corinna Cortes)
yann.lecun.com/exdb/mnist
[0:35:50] CIFAR-100 dataset - A dataset of 32x32 color images in 100 classes, with 600 images per class (Alex Krizhevsky)
https://www.cs.toronto.edu/~kriz/cifar.html
[0:36:00] ImageNet - Large-scale hierarchical image database with over 14 million images organized according to the WordNet hierarchy (Jia Deng et al)
ieeexplore.ieee.org/document/5206848
[0:42:55] Llama 2: Collection of foundation and fine-tuned chat models ranging from 7B to 70B parameters (Hugo Touvron et al.)
arxiv.org/abs/2307.09288
[0:43:35] Scaling Instruction-Finetuned Language Models - Paper introducing Flan-T5, showing performance improvements through instruction finetuning (Hyung Won Chung et al.)
arxiv.org/abs/2210.11416
[0:45:10] Active Few-Shot Fine-Tuning methodology paper discussing exploration-exploitation trade-offs in the context of fine-tuning neural networks. (Jonas Hübotter et al.)
arxiv.org/abs/2402.15898

![Why Scientists Cant Rebuild a Polaroid Camera [César Hidalgo]
César Hidalgo has spent years trying to answer a deceptively simple question: What is knowledge, and why is it so hard to move around?
We all have this intuition that knowledge is just... information. Write it down in a book, upload it to GitHub, train an AI on it—done. But César argues thats completely wrong. Knowledge isnt a thing you can copy and paste. Its more like a living organism that needs the right environment, the right people, and constant exercise to survive.
Guest: César Hidalgo, Director of the Center for Collective Learning
The Big Ideas
1. Knowledge Follows Laws (Like Physics)
Just as temperature and gravity follow predictable rules, so does knowledge. César outlines three laws:
- Time: How knowledge grows (fast at first, then it plateaus)
- Space: How knowledge spreads (its way harder than you think)
- Value: How we can measure a countrys knowledge potential
2. You Cant Download Expertise
The most memorable stories in this conversation prove that knowledge is embodied—it lives in people, teams, and organizations, not in manuals.
3. Why Big Companies Fail to Adapt
César explains architectural innovation—the idea that small changes (like shipping books directly to customers) can require a completely different organizational structure.
4. The Infinite Alphabet of Economies
Every skill, every industry, every capability is like a letter in an alphabet. Césars research shows you can actually predict which countries will grow by counting their letters.
If you think AI can just copy human knowledge, or that development is just about throwing money at poor countries, or that writing things down preserves them forever—this conversation will change your mind. Knowledge is fragile, specific, and collective. It decays fast if you dont use it.
The Infinite Alphabet [César A. Hidalgo]
https://www.penguin.co.uk/books/458054/the-infinite-alphabet-by-hidalgo-cesar-a/9780241655672
https://x.com/cesifoti
Rescript link.
https://app.rescript.info/public/share/eaBHbEo9xamwbwpxzcVVm4NQjMh7lsOQKeWwNxmw0JQ
TIMESTAMPS:
00:00:00 The Three Laws of Knowledge
00:02:28 Rival vs. Non-Rival: The Economics of Ideas
00:05:43 Why You Cant Just Download Knowledge
00:08:11 The Detective Novel Analogy
00:11:54 Collective Learning & Organizational Networks
00:16:27 Architectural Innovation: Amazon vs. Barnes & Noble
00:19:15 The First Law: Learning Curves
00:23:05 The Samuel Slater Story: Treason & Memory
00:28:31 Physics of Knowledge: Joules Cannon
00:32:33 Extensive vs. Intensive Properties
00:35:45 Knowledge Decay: Ise Temple & Polaroid
00:41:20 Absorptive Capacity: Sony & Donetsk
00:47:08 Disruptive Innovation & S-Curves
00:51:23 Team Size & The Cost of Innovation
00:57:13 Geography of Knowledge: Vespas Origin
01:04:34 Migration, Diversity & Planet China
01:12:02 Institutions vs. Knowledge: The China Story
01:21:27 Economic Complexity & The Infinite Alphabet
01:32:27 Do LLMs Have Knowledge?
REFERENCES:
Book:
[00:47:45] The Innovators Dilemma (Christensen)
https://www.amazon.com/Innovators-Dilemma-Revolutionary-Change-Business/dp/0062060244
[00:55:15] Why Greatness Cannot Be Planned
https://amazon.com/dp/3319155237
[01:35:00] Why Information Grows
https://amazon.com/dp/0465048994
Paper:
[00:03:15] Endogenous Technological Change (Romer, 1990)
https://web.stanford.edu/~klenow/Romer_1990.pdf
[00:03:30] A Model of Growth Through Creative Destruction (Aghion & Howitt, 1992)
https://dash.harvard.edu/server/api/core/bitstreams/7312037d-2b2d-6bd4-e053-0100007fdf3b/content
[00:14:55] Organizational Learning: From Experience to Knowledge (Argote & Miron-Spektor, 2011)
https://www.researchgate.net/publication/228754233_Organizational_Learning_From_Experience_to_Knowledge
[00:17:05] Architectural Innovation (Henderson & Clark, 1990)
https://www.researchgate.net/publication/200465578_Architectural_Innovation_The_Reconfiguration_of_Existing_Product_Technologies_and_the_Failure_of_Established_Firms
[00:19:45] The Learning Curve Equation (Thurstone, 1916)
https://dn790007.ca.archive.org/0/items/learningcurveequ00thurrich/learningcurveequ00thurrich.pdf
[00:21:30] Factors Affecting the Cost of Airplanes (Wright, 1936)
https://pdodds.w3.uvm.edu/research/papers/others/1936/wright1936a.pdf
[00:52:45] Are Ideas Getting Harder to Find? (Bloom et al.)
https://web.stanford.edu/~chadj/IdeaPF.pdf
[01:33:00] LLMs/ Emergence
https://arxiv.org/abs/2506.11135
Person:
[00:25:30] Samuel Slater
https://en.wikipedia.org/wiki/Samuel_Slater
[00:42:05] Masaru Ibuka (Sony)
https://www.sony.com/en/SonyInfo/CorporateInfo/History/SonyHistory/1-02.html
[01:01:45] Corradino DAscanio
https://link.springer.com/chapter/10.1007/978-3-319-09858-6_38#:~:text=6%20Conclusions,%2C%20comfort%2C%20and%20technical%20performance.
[01:16:00] Chen Chunxian
https://thebhc.org/sites/default/files/tzeng.pdf
Event/Place: Why Scientists Cant Rebuild a Polaroid Camera [César Hidalgo]](https://i.ytimg.com/vi/vzpFOJRteeI/mqdefault.jpg)
![François Chollet on OpenAI o-models and ARC
SPONSOR MESSAGES:
***
CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments.
https://centml.ai/pricing/
Francois Chollet joins Tim Scarfe to discuss the outcomes of the 2024 ARC-AGI Prize, his departure from Google to start a new research lab focused on program synthesis, and why he believes current frontier models including o1 still cannot genuinely adapt to novelty.
Chollet breaks down the two paradigms that dominated the competition: deep learning-guided program synthesis (induction) and test-time training with direct prediction (transduction). Both approaches reached roughly 55% accuracy, but the striking finding is that solutions using $10 of compute matched those using $10,000. Compute is a multiplier for ideas, not a replacement for them.
The conversation goes deep into Clement Bonnets latent program search approach, Kevin Elliss hybrid induction-transduction strategy, and the OmniArc framework that trains a single model across multiple ARC-related tasks. Chollet explains why he sees program graphs rather than token-by-token code generation as the more promising architecture for program synthesis.
On the philosophical side, Chollet distinguishes two forms of reasoning memorized pattern application versus genuine on-the-fly recombination of cognitive building blocks. He argues that consciousness might emerge as a self-consistency mechanism needed for iterative reasoning, and that the question can LLMs reason? is less interesting than can they adapt to novelty?
Chollet also reveals his plans for ARC-2, discusses the logarithmic relationship between compute and accuracy that his data shows, and argues that the future of programming is democratization: anyone should be able to describe what they want automated, without writing code.
REFERENCES:
person:
[00:00:00] Francois Chollet
https://scholar.google.com/citations?user=VfYhf2wAAAAJ
[00:36:40] Kevin Ellis - Combining Induction and Transduction
https://scholar.google.com/citations?user=5YGiV0YAAAAJ
[00:45:00] Clement Bonnet - Latent Program Networks
https://scholar.google.com/citations?user=UQ3IbeoAAAAJ
tool:
[00:00:53] Keras
https://keras.io/
[00:11:00] ARC-AGI Prize
https://arcprize.org/
[00:11:00] ARC-AGI Dataset
https://github.com/fchollet/ARC-AGI
paper:
[00:16:03] On the Measure of Intelligence
https://arxiv.org/abs/1911.01547
[01:16:40] o3 ARC Breakthrough
https://arcprize.org/blog/oai-o3-pub-breakthrough
LINKS:
Full Transcript: https://app.rescript.info/share/44fe8a0aab7235883da9cb4744848cfe
Download PDF transcript: https://app.rescript.info/api/public/sessions/7b446884aa257347/pdf François Chollet on OpenAI o-models and ARC](https://i.ytimg.com/vi/w9WE1aOPjHc/mqdefault.jpg)




![YUDKOWSKY + WOLFRAM ON AI RISK.
Eliezer Yudkowsky and Stephen Wolfram discuss artificial intelligence and its potential existen‑
tial risks. They traversed fundamental questions about AI safety, consciousness, computational irreducibility, and the nature of intelligence.
The discourse centered on Yudkowsky’s argument that advanced AI systems pose an existential threat to humanity, primarily due to the challenge of alignment and the potential for emergent goals that diverge from human values. Wolfram, while acknowledging potential risks, approached the topic from a his signature measured perspective, emphasizing the importance of understanding computational systems’ fundamental nature and questioning whether AI systems would necessarily develop the kind of goal‑directed behavior Yudkowsky fears.
SHOWNOTES (transcription, references, summary, best quotes etc):
https://www.dropbox.com/scl/fi/3st8dts2ba7yob161dchd/EliezerWolfram.pdf?rlkey=b6va5j8upgqwl9s2muc924vtt&st=vemwqx7a&dl=0
***
MLST IS SPONSORED BY TUFA AI LABS!
The current winners of the ARC challenge, MindsAI are part of Tufa AI Labs. They are hiring ML engineers. Are you interested?! Please goto https://tufalabs.ai/
***
https://en.wikipedia.org/wiki/Eliezer_Yudkowsky
https://en.wikipedia.org/wiki/Stephen_Wolfram
TOC:
1. Foundational AI Concepts and Risks
[00:00:00] 1.1 AI Optimization and System Capabilities Debate
[00:06:46] 1.2 Computational Irreducibility and Intelligence Limitations
[00:20:09] 1.3 Existential Risk and Species Succession
[00:23:28] 1.4 Consciousness and Value Preservation in AI Systems
2. Ethics and Philosophy in AI
[00:33:24] 2.1 Moral Value of Human Consciousness vs. Computation
[00:36:30] 2.2 Ethics and Moral Philosophy Debate
[00:39:58] 2.3 Existential Risks and Digital Immortality
[00:43:30] 2.4 Consciousness and Personal Identity in Brain Emulation
3. Truth and Logic in AI Systems
[00:54:39] 3.1 AI Persuasion Ethics and Truth
[01:01:48] 3.2 Mathematical Truth and Logic in AI Systems
[01:11:29] 3.3 Universal Truth vs Personal Interpretation in Ethics and Mathematics
[01:14:43] 3.4 Quantum Mechanics and Fundamental Reality Debate
4. AI Capabilities and Constraints
[01:21:21] 4.1 AI Perception and Physical Laws
[01:28:33] 4.2 AI Capabilities and Computational Constraints
[01:34:59] 4.3 AI Motivation and Anthropomorphization Debate
[01:38:09] 4.4 Prediction vs Agency in AI Systems
5. AI System Architecture and Behavior
[01:44:47] 5.1 Computational Irreducibility and Probabilistic Prediction
[01:48:10] 5.2 Teleological vs Mechanistic Explanations of AI Behavior
[02:09:41] 5.3 Machine Learning as Assembly of Computational Components
[02:29:52] 5.4 AI Safety and Predictability in Complex Systems
6. Goal Optimization and Alignment
[02:50:30] 6.1 Goal Specification and Optimization Challenges in AI Systems
[02:58:31] 6.2 Intelligence, Computation, and Goal-Directed Behavior
[03:02:18] 6.3 Optimization Goals and Human Existential Risk
[03:08:49] 6.4 Emergent Goals and AI Alignment Challenges
7. AI Evolution and Risk Assessment
[03:19:44] 7.1 Inner Optimization and Mesa-Optimization Theory
[03:34:00] 7.2 Dynamic AI Goals and Extinction Risk Debate
[03:56:05] 7.3 AI Risk and Biological System Analogies
[04:09:37] 7.4 Expert Risk Assessments and Optimism vs Reality
8. Future Implications and Economics
[04:13:01] 8.1 Economic and Proliferation Considerations YUDKOWSKY + WOLFRAM ON AI RISK.](https://i.ytimg.com/vi/xjH2B_sE_RQ/mqdefault.jpg)
![Can Rewriting an AI Agent Bend the Intelligence Curve? - Zhengyao Jiang
Weco let an AI coding agent rewrite the harness around another agent for eight days: its code, prompts and tools, while the underlying language model stayed fixed. Tim Scarfe asks Weco co-founder Zhengyao Jiang what the reported gains over two years of human engineering actually demonstrate.
The discussion examines AIDE 85s generated code, held-out evaluation and the difficulty of separating useful discoveries from reward hacking. Jiang explains Wecos four levels of recursive self-improvement and compares the experiment with AlphaEvolve and the Darwin Gödel Machine.
The limits matter as much as the gains. Jiang explains why the experiment did not establish that the system had become a better improver. The conversation closes with open-ended search, human-designed primitives and Parameter Golf: where does the next useful idea come from when the agent is searching inside a space that people designed?
TIMESTAMPS:
00:00:00 Eight days of self-improvement: what counts?
00:03:25 AIDE and the puzzle of useful spaghetti code
00:08:38 Four levels of recursive self-improvement
00:12:02 What AIDE 85 changed and how it was tested
00:20:04 AlphaEvolve, Darwin Gödel Machine and the RSI claim
00:26:21 Reward hacking and the limits of detection
00:33:09 Open-ended search, harness tuning and creativity
00:39:43 Parameter Golf and the limits of self-improvement
REFERENCES:
organization:
[00:00:30] Weco AI
https://www.weco.ai/
other:
[00:00:33] AIDE²: The First Evidence of Recursive Self-Improvement
https://www.weco.ai/blog/first-evidence-of-recursive-self-improvement
[00:14:11] Faulty reward functions in the wild
https://openai.com/index/faulty-reward-functions/
[00:29:59] The Hugging Face incident and the road ahead
https://openai.com/index/hugging-face-incident-and-the-road-ahead/
tool:
[00:03:29] AIDE
https://github.com/WecoAI/aideml
[00:04:29] MLE-bench
https://github.com/openai/mle-bench
[00:04:33] ALE-Bench
https://github.com/SakanaAI/ALE-Bench
[00:04:52] WeatherBench 2
https://github.com/google-research/weatherbench2
[00:08:18] ReAct
https://react-lm.github.io/
[00:39:43] Parameter Golf
https://github.com/openai/parameter-golf
paper:
[00:20:08] AlphaEvolve: A coding agent for scientific and algorithmic discovery
https://arxiv.org/abs/2506.13131v1
[00:21:35] Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
https://arxiv.org/abs/2505.22954v3
[00:23:45] Hyperagents
https://arxiv.org/abs/2603.19461v1
[00:27:01] SpecBench: Measuring Reward Hacking in Long-Horizon Coding Agents
https://arxiv.org/abs/2605.21384
book:
[00:33:14] Why Greatness Cannot Be Planned: The Myth of the Objective
https://link.springer.com/book/10.1007/978-3-319-15524-1
LINKS:
https://app.rescript.info/share/3a9dc6189cb539c6a05fcc4f75c101b3
PDF:
https://app.rescript.info/api/public/sessions/9eda60ede2b31c92/pdf Can Rewriting an AI Agent Bend the Intelligence Curve? - Zhengyao Jiang](https://i.ytimg.com/vi/yB6_iFGTq9k/mqdefault.jpg)
![Moving Beyond Surface Statistics (Apple researcher) [Iman Mirzadeh]
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 dont 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.
Thats the core finding of GSM-Symbolic. By generating templated variants of math word problems, Mirzadehs 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 systems 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
https://arxiv.org/abs/1712.01815
[00:17:15] GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models
https://arxiv.org/abs/2410.05229
[00:21:20] Connectionism and Cognitive Architecture: A Critical Analysis
https://www.sciencedirect.com/science/article/pii/001002779090014B
[00:29:35] On the Measure of Intelligence
https://arxiv.org/abs/1911.01547
[00:33:25] On definition of intelligence
https://www.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
https://arxiv.org/abs/2201.11903
[00:47:45] Scaling Laws for Neural Language Models
https://arxiv.org/abs/2001.08361
[00:55:10] Tensor Product Variable Binding and the Representation of Symbolic Structures in Connectionist Systems
https://www.sciencedirect.com/science/article/abs/pii/000437029090007M
book:
[00:07:05] Game Changer: AlphaZeros Groundbreaking Chess Strategies
https://www.amazon.com/Game-Changer-AlphaZeros-Groundbreaking-Strategies/dp/9056918184
[00:37:35] How We Learn: Why Brains Learn Better Than Any Machine... for Now
https://www.amazon.com/How-We-Learn-Brains-Machine/dp/0525559884
[00:39:30] Surfaces and Essences: Analogy as the Fuel and Fire of Thinking
https://www.amazon.com/Surfaces-Essences-Analogy-Fuel-Thinking/dp/0465018475
reference:
[00:11:30] NLP Course: Language Modeling
http://lena-voita.github.io/nlp_course/language_modeling.html
[01:08:40] GSM8K: Training Verifiers to Solve Math Word Problems
https://huggingface.co/datasets/openai/gsm8k
LINKS:
Full Transcript: https://app.rescript.info/share/72689965572c7fd460954f70e26b2eaa
Download PDF transcript: https://app.rescript.info/api/public/sessions/abf4f57b59fcbbd5/pdf Moving Beyond Surface Statistics (Apple researcher) [Iman Mirzadeh]](https://i.ytimg.com/vi/yQPduek-Q5s/mqdefault.jpg)
