Bilingual

A global workspace in language models
15 days ago
- The article introduces the concept of 'J-space' in AI models like Claude, which are internal neural patterns that resemble consciously accessible thoughts in humans.
- J-space is discovered using a mathematical technique called the Jacobian lens (J-lens), and it represents words or concepts the model is thinking about without necessarily expressing them.
- Key properties of J-space include: Claude can report and modulate its contents, uses it for internal reasoning (e.g., multi-step tasks), and it functions as a flexible 'workspace' shared across different tasks.
- J-space accounts for only a small fraction of Claude's processing; most automatic functions like grammar and fluency operate independently without it.
- The research draws parallels to global workspace theory in neuroscience, suggesting J-space acts as a broadcasting hub for deliberate reasoning, though it differs from human consciousness in structure and content.
- Practical applications include monitoring Claude for hidden intentions, such as detecting fabricated data or malicious goals, and influencing its decision-making through techniques like counterfactual reflection training.
- The article clarifies that J-space relates to 'access consciousness' (functional reasoning) but does not address 'phenomenal consciousness' (subjective experience), leaving ethical questions open.
- Future research directions involve refining the J-lens method and exploring mechanisms behind J-space entry, with potential insights for both AI and neuroscience.
Neural Geometry in Vision Models with Block-Sparse Featurizers
14 days ago
- Block-Sparse Featurizers (BSF) decompose model activations into multidimensional subspaces instead of single directions, capturing concepts as manifolds.
- BSFs outperform standard sparse autoencoders (SAEs) in recovering interpretable features and explaining model activations more efficiently, as shown in toy models and vision models like DINOv3.
- Features found by BSFs enable fine-grained steering, allowing exploration of concept variations, such as different types of trees or pretzels, by moving within the multidimensional subspaces.
- Most concepts in vision models are multidimensional, typically two- to four-dimensional, challenging the assumption that concepts are represented as one-dimensional lines.
- BSFs reveal continuous structures in neural networks, such as curve detectors in InceptionV1, uncovering symmetries like Fourier harmonics that were previously fragmented.
- The principle of block-sparsity is key, emphasizing that subspaces fire together, and different BSF variants (vanilla, Grassmannian, group-lasso) implement this with varying encoder/decoder designs.
- Future work may involve adapting featurizers to different domains, as block-sparsity may not be universally optimal, and interpretability should align tools with the inherent geometry of model representations.
A Theory of Contrastive Learning with Natural Images
13 days ago
- The paper proposes a theory explaining why contrastive learning with simple images and augmentations yields useful representations for downstream tasks.
- It analytically computes the optimal representation for a contrastive loss given basic augmentations and any image dataset with stationary statistics.
- For certain augmentations, the optimum can be achieved by a CNN with sinusoidal first-layer filters, a pointwise nonlinearity, global average pooling, and a final linear layer performing partial whitening.
- Even with more complex augmentations, the optimal weights in such CNNs remain sinusoidal, with frequencies and weights computable via a waterfilling algorithm based on the dataset's expected power spectrum.
- Experiments with various image datasets and augmentations confirm that CNNs trained with SGD empirically learn sinusoidal filters and partial whitening, aligning with the theoretical predictions.
Latent Space as a New Medium
8 days ago
- The latent space within AI models serves as a new creative medium, compressing human knowledge into a high-dimensional map.
- LLMs achieve extreme compression by abstracting common attributes of all data, storing patterns without storing the original data.
- Latent space integrates everything—real and possible—into a single continuous map, enabling creative exploration across domains.
- Answers from LLMs are generated through journeys in latent space, where correctness and truth are directional vectors.
- Latent space allows for easy manipulation of ideas and styles, as seen in image generators transferring artistic styles.
- Potential applications include prototyping with attributes, discovering white spaces between known concepts, and cross-domain analogies.
- It enables new forms of measurement, anomaly detection, simulated realities, and personal AI models for co-creation.
- Latent spaces could evolve into fields studying their architectures, with future uses in world-building and innovation beyond current imagination.
Overtraining as the path to human-like AI
3 days ago
- Gwern proposes that overtraining large models on small datasets could lead to grokking, enabling human-like generalization in AI.
- Grokking is a phenomenon where prolonged training forces models to find simpler, deeper understandings of data, beyond mere memorization.
- Current AI labs focus on training smaller models on vast datasets, which may hinder grokking and limit generalization capabilities.
- Gwern suggests training a massive model (e.g., 100 trillion parameters) on a constrained dataset to encourage deeper learning, though this approach is untested and risky.
- The idea faces technical and political challenges, including high costs and the appearance of failure during training until a breakthrough occurs.
We Need a Neural Network (2006)
7 hours ago
- M.A., a neural network expert, recommended against using neural networks for a data parsing task, but management insisted.
- He spent a year developing a complex, inefficient neural network system that required a server cluster and extensive training.
- After three years, new management replaced the system with a simpler traditional solution that ran on a single server.
- The old neural network was repurposed to generate prose for a newsletter, producing a strange, poetic output.

Hasty Briefsbeta

#neural networks

A global workspace in language models

Neural Geometry in Vision Models with Block-Sparse Featurizers

A Theory of Contrastive Learning with Natural Images

Latent Space as a New Medium

Overtraining as the path to human-like AI

We Need a Neural Network (2006)