Idhant Gulati
3 followers · 1182 views
on the atlas · 38
- Shaping and Monitoring Models Through Their Internal Representations - Goodfire1 savers
- Do Not Rely on the Linear Representation Hypothesis1 savers
- Neural Networks Are Already Clean55 savers
- Moving Image Archive5 savers
- Reading a Pretraining Run — Chunyuan Deng2 savers
- How To Be Successful96 savers
- Clarifying VLM concerns and project directions - Claude1 savers
- Curius / Onboarding2704 savers
- Verbalizable Representations Form a Global Workspace in Language Models28 savers
- Canopy Labs3 savers
- Apps - order-evaluation | Modal1 savers
- Using Interpretability to Identify a Novel Class of Alzheimer's Biomarkers8 savers
- The Humane Interface1 savers
- Dario Amodei — The Adolescence of Technology8 savers
- Emergence of Human to Robot Transfer in Vision-Language-Action Models7 savers
- KFAC explained1 savers
- Modular Manifolds - Thinking Machines Lab16 savers
- World Models | Rohit Bandaru4 savers
- LoRA Without Regret - Thinking Machines Lab37 savers
- How Social Media Shortens Your Life - by Gurwinder2 savers
- How Does A Blind Model See The Earth? - by henry2 savers
- Everything I know about good system design37 savers
- Highly Opinionated Advice on How to Write ML Papers — LessWrong2 savers
- GLM-4.5: Reasoning, Coding, and Agentic Abililties1 savers
- Look Ma, No Bubbles! Designing a Low-Latency Megakernel for Llama-1B · Hazy Research9 savers
- Mastermind - Desktop Writing Pad · Baronfig1 savers
- Inner Alignment: Explain like I'm 12 Edition — LessWrong1 savers
- Books — LessWrong1 savers
- AGI safety from first principles: Introduction — LessWrong1 savers
- Context7 - Up-to-date documentation for LLMs and AI code editors5 savers
- Attribution Patching: Activation Patching At Industrial Scale — Neel Nanda7 savers
- Waypoint | Tzafon1 savers
- Cognition | Kevin-32B: Multi-Turn RL for Writing CUDA Kernels2 savers
- Monet: Mixture of Monosemantic Experts for Transformers Explained — LessWrong1 savers
- Qwen3: Think Deeper, Act Faster | Qwen2 savers
- Dario Amodei — The Urgency of Interpretability17 savers
- Circuit Tracing: Revealing Computational Graphs in Language Models20 savers
- Transformer Circuits Thread17 savers
highlights · 32
In an LLM, the language of sensory-motor transformations is simple: as both text inputs and outputs are represented using the same natural language tokens, representing their nexus using the same tokens may simply be the most parsimonious solution.
Verbalizable Representations Form a Global Workspace in Language ModelsAlternatively, it may be the case that LLMs privilege verbalizable representations not only because their outputs are verbal, but because their inputs are as well (or at least a large fraction of their inputs—some LLMs process image input as well). That is, LLMs largely receive inputs and produce outputs of the same kind. In contrast, humans take in sensory information, and output motor commands; sensory and motor information are represented in very different coordinates, and the brain must do computational work to translate between them. Indeed, some theories propose that such sensory-motor t…
Verbalizable Representations Form a Global Workspace in Language ModelsWhat you hold. This is your deepest evidence pile. Counterfactual animals, edited maps, illusions, and — importantly — a half-built mechanism: existing work already traced the prior-vs-pixel fight to specific layers and a handful of steerable attention heads.
Clarifying VLM concerns and project directions - Claude3. Attention-as-homonym. The N45 inattentional-blindness anecdote genuinely has no controlled study behind it — that empirical territory is unclaimed, which makes this the most executable direction. But it's a vocabulary case study with a lower ceiling, and C19 (attention weight ≠ importance) is a built-in complication you'd have to handle rather than discover.
Clarifying VLM concerns and project directions - ClaudeOn the other axis, processing also unfolds across the sequence dimension via the attention mechanism, which can recall and transform information from any prior sequence position.
Verbalizable Representations Form a Global Workspace in Language ModelsOn one axis, processing takes place across layers, through a fixed number of computational steps.
Verbalizable Representations Form a Global Workspace in Language ModelsTo first order, this relationship is linear, and is described by the Jacobian matrix .
Verbalizable Representations Form a Global Workspace in Language ModelsThe basic idea is to characterize an intermediate activation vector by its first-order causal effect on the model's outputs, over a broad distribution of potential contexts.
Verbalizable Representations Form a Global Workspace in Language ModelsWe wanted to see if simply scaling up robotic foundation models could allow for emergent human to robot transfer without any explicit transfer learning mechanism.
Emergence of Human to Robot Transfer in Vision-Language-Action Modelsoptimal update is given by subtracting out the radial component from the gradient, normalizing and multiplying by the learning rate
Modular Manifolds - Thinking Machines Labtake steps in the tangent space
Modular Manifolds - Thinking Machines Labwe could use a non-manifold optimizer and just project the weights back to the manifold after each step
Modular Manifolds - Thinking Machines Labco-design optimization algorithms with these manifold constraints.
Modular Manifolds - Thinking Machines Labby keeping the tensors constrained to submanifolds at each layer
Modular Manifolds - Thinking Machines LabThis uses causal attention. This architecture scales linearly rather than quadratically with the number of frames, which makes it efficient at processing video.
World Models | Rohit Bandarusing a suboptimal learning rate, we swept the LR for each experimental condition. We used constant learning rate schedule (no warmup or cooldown).
LoRA Without Regret - Thinking Machines Labover three orders of magnitude, with rank between 1 and 512, and compared these to full fine-tuning.
LoRA Without Regret - Thinking Machines LabLoRA is less tolerant of large batch sizes than full fine-tuning — it pays a larger penalty in loss as batch size increases beyond some point. This penalty is not mitigated by increasing the LoRA rank; it is a property of the product-of-matrices parametrization, which has different training dynamics than optimizing the original weight matrix.
LoRA Without Regret - Thinking Machines Labcan LoRA match the performance of full fine-tuning, and if so, under which conditions?
LoRA Without Regret - Thinking Machines LabLoRA replaces each weight matrix W from the original model with a modified version 𝑊 ′ = 𝑊 + 𝛾 𝐵 𝐴 W ′ =W+γBA, where B and A are matrices that together have far fewer parameters than W, and 𝛾 γ is a constant scaling factor.
LoRA Without Regret - Thinking Machines Labgood design is self-effacing: bad design is often more impressive than good
Everything I know about good system designOur initial approach constructs the trajectories as follows. Starting with the initial prompt, we append the chain of thought, kernel, and evaluation information after each refinement step. We then assign a single reward to the entire trajectory—defined as the maximum score achieved by any kernel—and use this sequence for training.
Cognition | Kevin-32B: Multi-Turn RL for Writing CUDA Kernelswe extract feedback from a generated kernel and have the model refine it. If the kernel fails to compile, we pass the model the error trace and ask it to fix it. If it’s correct, we measure the runtime and ask the model to improve it further.
Cognition | Kevin-32B: Multi-Turn RL for Writing CUDA KernelsRunning a forward pass of this replacement model is identical to running the original model, with two modifications:
Circuit Tracing: Revealing Computational Graphs in Language Modelsreceive input from the model’s residual stream at their associated layer, but are “cross-layer” in the sense that they can provide output to all subsequent layers
Circuit Tracing: Revealing Computational Graphs in Language Modelssparsely active features
Circuit Tracing: Revealing Computational Graphs in Language Modelsgoal of the model is to reconstruct the outputs of the MLPs of the underlying model
Circuit Tracing: Revealing Computational Graphs in Language Modelssame number of layers as the underlying model
Circuit Tracing: Revealing Computational Graphs in Language Modelsneurons (“features”) divided into layers
Circuit Tracing: Revealing Computational Graphs in Language ModelsHowever, we successfully apply them to understand the circuits underlying small number addition.
Circuit Tracing: Revealing Computational Graphs in Language ModelsWe find that across prompts, perturbation experiments are generally qualitatively consistent with our attribution graphs, though there are some deviations.
Circuit Tracing: Revealing Computational Graphs in Language Modelsdirect feature-feature interactions.
Circuit Tracing: Revealing Computational Graphs in Language Models