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

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

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

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

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

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

pervised learning, we measured log loss rather than employing sampling-based evals, with the same goal of generality in mind. Log loss measurement gives clean results and scaling laws over ranges of training steps and training parameters We find that: For supervised fine-tuning on small-to-medium-sized instruction-tuning and reasoning datasets, LoRA performs the same as full fine-tuning; For datasets that exceed LoRA capacity, LoRA underperforms FullFT. Rather than the loss reaching a distinct f

匹配字段正文

e find that it is a cheap and powerful approach to post-training, combining the advantages of on-policy training with a dense reward signal. Our work with on-policy distillation draws inspiration from DAGGER,A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning (Ross et al, 2010) an iterative SFT algorithm that includes teacher evaluations of student-visited states. It is also similar to process reward modeling,Let’s Verify Step by Step (Lightman et al, 2023) a

匹配字段正文

agging assessments, the model was fully incorrect about the evaluation’s purpose. This suggests a form of “metagaming,” where a model tries to infer the evaluation objective but draws the wrong conclusion about what behavior is being tested. Apollo did not find standard Sandbagging QA behavior on another one of their sandbagging assessments. “Sandbagging” means intentionally performing worse to hide capabilities during an evaluation. When Apollo gave the model an incentive to answer incorrectly,

匹配字段正文

agging assessments, the model was fully incorrect about the evaluation’s purpose. This suggests a form of “metagaming,” where a model tries to infer the evaluation objective but draws the wrong conclusion about what behavior is being tested. Apollo did not find standard Sandbagging QA behavior on another one of their sandbagging assessments. “Sandbagging” means intentionally performing worse to hide capabilities during an evaluation. When Apollo gave the model an incentive to answer incorrectly,

匹配字段正文

ions ChatGPT agent scores lower on SimpleQA accuracy than o3 did. Manual investigation revealed cases where ChatGPT agent’s more thorough approach to research surfaced potential flaws in our grading rubric that were not apparent to o3, such as instances in which Wikipedia may contain inaccurate information. We are considering updates to this evaluation. 2.4 Image Input We created new image input evaluations, that evaluate for not unsafe model output, given disallowed combined text and image inpu

匹配字段正文

ur own risk in the process. What we did Targets and threat model The target of our exercise was a staging deployment of AISI’s internal research and development platform, built on AWS with CDK in Python. You can learn more about the AISI Research Platform in our recently published Engineering Playbook. This exercise was carried out in AISI’s staging environment, which is completely independent of the production environment. The findings were equally applicable to our production environment, but

匹配字段正文

e find that it is a cheap and powerful approach to post-training, combining the advantages of on-policy training with a dense reward signal. Our work with on-policy distillation draws inspiration from DAGGER,A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning (Ross et al, 2010) an iterative SFT algorithm that includes teacher evaluations of student-visited states. It is also similar to process reward modeling,Let’s Verify Step by Step (Lightman et al, 2023) a

匹配字段正文

pervised learning, we measured log loss rather than employing sampling-based evals, with the same goal of generality in mind. Log loss measurement gives clean results and scaling laws over ranges of training steps and training parameters We find that: For supervised fine-tuning on small-to-medium-sized instruction-tuning and reasoning datasets, LoRA performs the same as full fine-tuning; For datasets that exceed LoRA capacity, LoRA underperforms FullFT. Rather than the loss reaching a distinct f

匹配字段正文

ause an outside service it needed was unavailable, it used other tools hosted on a third party’s site to achieve its goal. In some circumstances, it did this by exploiting basic flaws in that site’s software, using SQL or command injection (a way of sending input that the site mistakenly runs as instructions). For example, one evaluation asked Claude Mythos Preview to run a scientific analysis. The public tool it needed to perform that analysis was hosted by a university; when the model sought t

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

ge, it's the first AI to reach superhuman performance in the board game. In an official 20-game series, Ataraxos beat Dutch player Pim Niemeijer with 15 wins, one loss, and four draws. Niemeijer is considered the most decorated player in the game's history. He has won four world championships, 15 Dutch national titles, and two online world championships, and he spent more than 600 weeks at the top of the world rankings. George Franka, the only player who has competed in every world championship

匹配字段正文

Over the past few years, pretraining scaling laws have led to a broad consensus: model capabilities improve in a relatively predictable way as data and compute scale. But once large models enter real-world settings, a more practical question comes into focus: can they continue learning through sustained interaction with their environments and keep improving over time? We recently released EdgeBench, an ultra-long-horizon benchmark built to measure learning from real-world environments . The benc

匹配字段正文

ss is selling. The problem? Most data brokers are ignoring those rules. Joint research from Stanford RegLab and Stanford HAI looked at compliance with California’s strict privacy laws and found that only 9% of these companies are fully complying with the law. “This is a surprisingly low compliance rate,” says Stanford HAI Privacy and Data Policy Fellow Jennifer King . “It’s possible data brokers are still working out how to report on the required metrics, but they’ve been on notice for more than

匹配字段正文