Is ChatGPT an N-gram model on steroids? @MachineLearningStreetTalk

Uploaded August 2024 | Updated September 2026, 1 week ago

DeepMind machine learning scientist / MIT scholar Dr. Timothy Nguyen discusses his recent paper on understanding transformers through n-gram statistics. Nguyen explains his approach to analyzing transformer behavior using a kind of "template matching" (N-grams), providing insights into how these models process and predict language.

MLST is sponsored by Brave:
The Brave Search API covers over 20 billion webpages, built from scratch without Big Tech biases or the recent extortionate price hikes on search API access. Perfect for AI model training and retrieval augmentated generation. Try it now - get 2,000 free queries monthly at brave.com/api.

Key points covered include:
A method for describing transformer predictions using n-gram statistics without relying on internal mechanisms.
The discovery of a technique to detect overfitting in large language models without using holdout sets.
Observations on curriculum learning, showing how transformers progress from simpler to more complex rules during training.
Discussion of distance measures used in the analysis, particularly the variational distance.
Exploration of model sizes, training dynamics, and their impact on the results.

We also touch on philosophical aspects of describing versus explaining AI behavior, and the challenges in understanding the abstractions formed by neural networks. Nguyen concludes by discussing potential future research directions, including attempts to convert descriptions of transformer behavior into explanations of internal mechanisms.

Timothy Nguyen's earned his B.S. and Ph.D. in mathematics from Caltech and MIT, respectively. He held positions as Research Assistant Professor at the Simons Center for Geometry and Physics (2011-2014) and Visiting Assistant Professor at Michigan State University (2014-2017). During this time, his research expanded into high-energy physics, focusing on mathematical problems in quantum field theory. His work notably provided a simplified and corrected formulation of perturbative path integrals.

Since 2017, Nguyen has been working in industry, applying his expertise to machine learning. He is currently at DeepMind, where he contributes to both fundamental research and practical applications of deep learning to solve real-world problems.

Refs:
The Cartesian Cafe
youtube.com/@TimothyNguyen

Understanding Transformers via N-Gram Statistics
researchgate.net/publication/382204056_Understanding_Transformers_via_N-Gram_Statistics

TOC
00:00:00 Timothy Nguyen's background
00:02:50 Paper overview: transformers and n-gram statistics
00:04:55 Template matching and hash table approach
00:08:55 Comparing templates to transformer predictions
00:12:01 Describing vs explaining transformer behavior
00:15:36 Detecting overfitting without holdout sets
00:22:47 Curriculum learning in training
00:26:32 Distance measures in analysis
00:28:58 Model sizes and training dynamics
00:30:39 Future research directions
00:32:06 Conclusion and future topics

This is what happens when you let AIs debate

Microsofts professor Chris Bishop on the Sparks of AGI

AI Doesnt Need Intelligence. Thats the Whole Point. - Luciano Floridi

NEURAL NETWORKS ARE WEIRD! - Neel Nanda (DeepMind)

Biologically-inspired AI and Mortal Computation

A CERN like effort for AI (Jakob Foerster)

$Strange Geometric Shapes Found Inside AIs — Tom McGrath Tom McGrath is co-founder and Chief Scientist at Goodfire, and a former Google DeepMind researcher. He joins Tim Scarfe to ask what neural networks actually learn, whether their internal representations converge on structures in the world, and whether interpretability can extract new scientific knowledge rather than merely explain model outputs. Beginning with AlphaZero and learned modularity, the conversation moves into neural geometry: concept manifolds, reusable computation inside Llama, and why activation steering can fail when it pushes a model off-manifold. McGrath then makes the case for intentional design, using interpretability as part of the training loop. They examine controlled generalisation, features as rewards, predictive data debugging, and the uncomfortable fact that a model may recognise a hallucination or reward hack and still produce it. The discussion closes on grader awareness, oversight and collusion between adaptive agents, then returns to sparse autoencoders. SAEs are useful, McGrath argues, but they may fracture the higher-dimensional structures networks actually use. This episode was made with support from Goodfire. TIMESTAMPS: 00:00:00 Introduction: Can interpretability speed-run science? 00:02:03 The invisible grader 00:06:51 What AlphaZero learned from the world 00:12:24 Interpretability as a control loop 00:21:54 The forbidden method and safer interventions 00:37:36 Why models catch hallucinations too late 00:46:19 Debug the dataset before training 00:50:44 Why neural networks become modular 00:55:57 Finding the geometry inside a network 01:02:55 Why steering falls off the manifold 01:12:10 A reusable calculator inside Llama 01:17:19 From abstractions to goals 01:25:28 Reward hacking, oversight and collusion 01:37:23 Are sparse autoencoders dead? REFERENCES: paper: [00:05:45] Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs https://arxiv.org/abs/2502.17424v7 [00:11:05] Acquisition of Chess Knowledge in AlphaZero https://arxiv.org/abs/2111.09259 [00:25:30] Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning https://arxiv.org/abs/2507.16795 [00:29:30] Persona Vectors: Monitoring and Controlling Character Traits in Language Models https://arxiv.org/abs/2507.21509 [00:41:14] Features as Rewards: Scalable Supervision for Open-Ended Tasks via Interpretability https://arxiv.org/abs/2602.10067 [00:47:03] Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal https://arxiv.org/abs/2606.12360 [01:00:26] Do Sparse Autoencoders Capture Concept Manifolds? https://arxiv.org/abs/2604.28119 [01:03:04] Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior https://arxiv.org/abs/2605.05115 [01:14:20] Arithmetic in the Wild: Llama uses Base-10 Addition to Reason About Cyclic Concepts https://arxiv.org/abs/2605.01148 [01:29:35] Measuring Reward-Seeking via Contrastive Belief Updates https://arxiv.org/abs/2607.18966v1 other: [00:15:44] Intentional Design https://www.goodfire.com/blog/intentional-design [00:56:12] The World Inside Neural Networks https://www.goodfire.com/research/the-world-inside-neural-networks [01:37:28] A Pragmatic Vision for Interpretability https://www.alignmentforum.org/posts/StENzDcD3kpfGJssR/a-pragmatic-vision-for-interpretability RESCRIPT: https://app.rescript.info/share/846cfee4131b664fd09209cc3b98018e$

AI Agents can write 10,000 lines of hacking code in seconds [Dr. Ilia Shumailov]

Is the Mind More Than Just the Brain? - Tom Froese