搜索“AWS”

systems. In particular, the phase change we observe forms an interesting potential bridge between the microscopic domain of interpretability and the macroscopic domain of scaling laws and learning dynamics. The rest of the paper is organized as follows. We start by clarifying several key concepts and definitions, including in-context learning, induction heads, and a “per-token loss analysis” method we use throughout. We then present the 6 arguments one by one, drawing on evidence from analysis o

匹配字段正文

ible! But many of these neurons appear to respond to several unrelated but individually understandable features, such as a neuron which responds to cat heads, fronts of cars, and paws. While we can't totally rule out that there isn't some deep commonality between a cat's paw and the front of a car, it seems like the simpler explanation is that the network has grouped several unrelated features together. We call these polysemantic neurons . Note that polysemanticity is what one would expect to ob

匹配字段正文

extremely microscopic analysis of neural networks. Can we make useful predictions about macroscopic behavior? Can we explain exceptions in more macroscopic theories like scaling laws? Induction heads appear to be an example of a microscopic neural network phenomena with macroscopic effects. They can cause a visible bump in transformer language model loss curves. Induction heads are also likely to be responsible for a bump observed in the original scaling laws paper (see Figure 13: "The conspicu

匹配字段正文

this works, especially the section on LayerNorm. (We can file this as reason 78 for why interpretability researchers hate LayerNorm.) Can we explain learning dynamics and scaling laws in terms of circuits? We know that induction heads cause a loss bump in training when they form, and likely cause a bump in scaling laws. This suggests the tantalizing possibility of a deep bridge between the microscopic world of mechanistic interpretability and the more macroscopic topics of learning dynamics and

匹配字段正文

the mechanistic interpretability agenda will now turn on succeeding at a difficult engineering and scaling problem, which frontier AI labs have significant expertise in. Scaling Laws for Dictionary Learning. It's worth noting that there's enormous uncertainty about the dynamics of scaling dictionary learning and sparse autoencoders discussed above. As we make the subject model bigger, how does the ideal expansion factor change? (Does it stay constant?) How does the necessary amount of data chan

匹配字段正文

ible amount of low hanging fruit. Noticing striking structure like this also happens in more quantitative work. For example, noticing that neural networks have such clean scaling laws when you look at a log-log plot is an example of the "signal of structure" – the structure must be telling you something! And in particular, such structure is often telling you about a new abstraction. The critical thing is that you need to know the structure isn't coming from some other source. For example, certai

匹配字段正文

of a colleague sharing some thoughts or preliminary experiments for a few minutes at a lab meeting, rather than a mature paper. New Posts Open Roles In Interpretability; Scaling Laws for Dictionary Learning; Update on how we train SAEs; How Strongly do Dictionary Learning Features Influence Model Behavior?; Interpretability Architectures Project; Caloric and the Utility of Incorrect Theories; Open Problem: Attribution Dictionary Learning; Isolating Circuits Paths of Different Lengths; Research

匹配字段正文

l be needed to understand the implications of these potentially safety-relevant features. Key Results Sparse autoencoders produce interpretable features for large models; Scaling laws can be used to guide the training of sparse autoencoders; The resulting features are highly abstract: multilingual, multimodal, and generalizing between concrete and abstract references; There appears to be a systematic relationship between the frequency of concepts and the dictionary size needed to resolve feature

匹配字段正文

oss datasets generated by different size transformers. During training we use Adam optimizer beta1=0.9, beta2=0.999 and no weight decay. Our learning rate varies based on scaling laws, but 2e-4 is a reasonable default. The learning rate is decayed linearly to zero over the last 20% of training. We vary training steps based on scaling laws. We use batch size 32,768 which we believe to be under the critical batch size. The gradient norm is clipped to 1 (using clip grad norm). We vary \lambda S dur

匹配字段正文

learning rate to scale approximately as one over the square root of the number of FLOPs. These are rough best-guess estimates based on prior experiments, and not precise scaling laws. In our largest 18L run, we used 3B training tokens. In our largest Haiku run, we used 16B training tokens. In 18L, we used a constant sparsity penalty across CLT sizes. In Haiku, we increased the penalty with the CLT size (we found that otherwise, the L0 of the runs increased with size). In general we targeted L0

匹配字段正文

low-level text into these higher-level conceptual features, and vice versa? And to what extent does the model have an internal representation where it pictures, plans, and then draws an image? Edit 11/3/2025: Thanks to Lee Sharkey for making us aware of additional related work on data-point initialization, Taking features out of superposition with sparse autoencoders more quickly with informed initialization by Peigné. Our work extends these learnings by demonstrating the efficacy of this appro

匹配字段正文

AI systems that can faithfully explain their decision-making processes. Introspective capabilities could allow models to accurately report on their uncertainty, identify gaps or flaws in their reasoning, and explain the motivations underlying their actions. However, this same capability introduces new risks. Models with genuine introspective awareness might better recognize when their objectives diverge from those intended by their creators, and could potentially learn to conceal such misalignme

匹配字段正文

both matter. Consider this: you might be painting with extraordinary intuition, noticing small details others miss - the worry in weather patterns, the hope for connection that draws certain people near. Artists often see what's coming because they pay attention differently. Instead of fearing your gift, perhaps paint with intention. Balance those tender, difficult truths with paintings of joy, healing, and hope. Paint the sunrise after storms, the laughter you wish to see, the mended hearts. I

匹配字段正文

r and CEO Jensen Huang. Pavillon Vendôme sits just off Place Vendôme in the 1st arrondissement — one of the most storied addresses in Paris. On Tuesday afternoon, it was full of claws. NVIDIA’s Build-a-Claw came to France with developers from the nation’s startup community and big companies alike packing into the pavilion. The excited crowd was lined up around the block to try their hands (and brains) at building AI-powered robotic claw machines using open models and NVIDIA’s platform. NVIDIA fo

匹配字段正文

riculum. All learning interactives released in the library and available today were vetted and approved by teachers. These include topics from school curriculums such as Kepler's Laws of Planetary Motion, Data Visualization and Projectile Motion. In addition, a collection of learning interactives was evaluated by STEM teachers in the UK. Results show that overall rating is good or excellent with physics and chemistry being the most amenable to simulation creation. Full details and results are av

匹配字段正文

ctions before reaching the main patent claims. Juniors also focused on surface-level synonym swapping rather than commercial scope improvement. Even when juniors spotted serious flaws, they often left descriptive comments diagnosing the defect rather than executing the redline to fix it. Because this "diagnose-without-execute" posture was equally common among unassisted control juniors, it represents a baseline junior deficit that three months of relying on AI to execute rewrites did not remedia

匹配字段正文

4 beats MXFP4, with up to about 8% faster decode at small batches (@StasBekman); Grid constraints: Mississippi’s Entergy warned it would curtail Amazon data centers at peak times. AWS responded with Project Falcon, which lets the utility switch sites to backup power itself (report); Agent infrastructure: Perplexity runs each project as a Git forge, with ephemeral agent sandboxes committing to S3. Fewer than 2% of commits collide, and grep beat semantic search 63% to 37% on financial document que

匹配字段正文

t says provides more accurate forecasts than the traditional model run by the European Centre for Medium-Range Weather Forecasts (ECMWF), which simulates the atmosphere using the laws of physics. Since the arrival of transformer model architectures and the ECMWF’s release of a massive weather dataset a few years ago, AI weather forecasting has become one of the more productive uses of deep-learning technology. Meteorologists previously relied on expensive supercomputers running detailed equation

匹配字段正文

Only five years into its journey, Character.AI, which lets users create AI characters and have conversations with them, has filed for Chapter 11 bankruptcy, following a slew of lawsuits and struggles to scale its popular platform. In a court document , the company said it had assets and liabilities between $50 million and $100 million. “Over the past year, Character.ai has taken meaningful steps to sustain and improve its business in the face of significant legal, regulatory, and operational hea

匹配字段正文

” The benchmark works by measuring “the experimental component of research taste; given a fixed research problem, we measure how well a model iteratively designs experiments and draws conclusions from their outcomes”. The model being analyzed designs experiments and interprets their results. Results are checked against a human baseline “drawn from 24 experts (at least two per task) who have recently worked at organizations including OpenAI, Google DeepMind, NVIDIA and Microsoft Research”. Result

匹配字段正文