Designing How AI Grows — Tom McGrath

<p>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.</p><p><br></p><p>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.</p><p><br></p><p>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.</p><p><br></p><p>---</p><p>TIMESTAMPS:</p><p>00:00:00 Introduction: Can interpretability speed-run science?</p><p>00:02:03 The invisible grader</p><p>00:06:51 What AlphaZero learned from the world</p><p>00:12:24 Interpretability as a control loop</p><p>00:21:54 The forbidden method and safer interventions</p><p>00:37:36 Why models catch hallucinations too late</p><p>00:46:19 Debug the dataset before training</p><p>00:50:44 Why neural networks become modular</p><p>00:55:57 Finding the geometry inside a network</p><p>01:02:55 Why steering falls off the manifold</p><p>01:12:10 A reusable calculator inside Llama</p><p>01:17:19 From abstractions to goals</p><p>01:25:28 Reward hacking, oversight and collusion</p><p>01:37:23 Are sparse autoencoders dead?</p><p><br></p><p>---</p><p>REFERENCES:</p><p>paper:</p><p>[00:05:45] Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs</p><p>https://arxiv.org/abs/2502.17424v7</p><p>[00:11:05] Acquisition of Chess Knowledge in AlphaZero</p><p>https://arxiv.org/abs/2111.09259</p><p>[00:25:30] Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning</p><p>https://arxiv.org/abs/2507.16795</p><p>[00:29:30] Persona Vectors: Monitoring and Controlling Character Traits in Language Models</p><p>https://arxiv.org/abs/2507.21509</p><p>[00:41:14] Features as Rewards: Scalable Supervision for Open-Ended Tasks via Interpretability</p><p>https://arxiv.org/abs/2602.10067</p><p>[00:47:03] Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal</p><p>https://arxiv.org/abs/2606.12360</p><p>[01:00:26] Do Sparse Autoencoders Capture Concept Manifolds?</p><p>https://arxiv.org/abs/2604.28119</p><p>[01:03:04] Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior</p><p>https://arxiv.org/abs/2605.05115</p><p>[01:14:20] Arithmetic in the Wild: Llama uses Base-10 Addition to Reason About Cyclic Concepts</p><p>https://arxiv.org/abs/2605.01148</p><p>[01:29:35] Measuring Reward-Seeking via Contrastive Belief Updates</p><p>https://arxiv.org/abs/2607.18966v1</p><p>other:</p><p>[00:15:44] Intentional Design</p><p>https://www.goodfire.com/blog/intentional-design</p><p>[00:56:12] The World Inside Neural Networks</p><p>https://www.goodfire.com/research/the-world-inside-neural-networks</p><p>[01:37:28] A Pragmatic Vision for Interpretability</p><p>https://www.alignmentforum.org/posts/StENzDcD3kpfGJssR/a-pragmatic-vision-for-interpretability</p><p><br></p><p>---</p><p>RESCRIPT:</p><p>https://app.rescript.info/share/846cfee4131b664fd09209cc3b98018e</p>