The core idea is to train Claude to explain its own activations. But how do we know whether an explanation is good? Since we don't know what thoughts an activation actually encodes, we can't directly check whether an explanation is accurate. So we train a second copy of Claude to work backwards—reconstruct the original activation from the text explanation. We consider an explanation to be good if it leads to an accurate reconstruction. We then train Claude to produce better explanations according to this definition using standard AI training techniques. In more detail, suppose we have a langua
Natural Language AutoencodersBefore we started this research, it was not clear where the misaligned behavior was coming from. Our main two hypotheses were: Our post-training process was accidentally encouraging this behavior with misaligned rewards.This behavior was coming from the pre-trained model and our post-training was failing to sufficiently discourage it. We now believe that (2) is largely responsible. Specifically, at the time of Claude 4’s training, the vast majority of our alignment training was standard chat-based Reinforcement Learning from Human Feedback RLHF data that did not include any agentic tool use. T
Teaching Claude why… We continue to choose AWS as our primary training and cloud provider for mission-critical workloads. “Our custom AI silicon offers high performance at significantly lower cost for customers, which is why it’s in such hot demand,” said Andy Jassy, CEO of Amazon. “Anthropic's commitment to run its la… …
… Those who join the network will have access to our Partner Portal, where we’ll share our Anthropic Academy training materials, the sales playbooks used by our own go-to-market team, and other co-marketing documentation. …
… An auditor equipped with NLAs successfully uncovered the target model’s hidden motivation between 12% and 15% of the time, even without access to the training data that implanted it. …
… Without action to limit China’s access to US compute, the CCP would have had all the ingredients to develop AI at par or superior to America’s. Some observers worry that constraining access to compute will force AI labs in China to innovate on other axes, reducing the American lead. …
… Our main two hypotheses were: Our post-training process was accidentally encouraging this behavior with misaligned rewards. This behavior was coming from the pre-trained model and our post-training was failing to sufficiently discourage it. …
… We tackle this from multiple angles during Claude’s training. First, we construct training scenarios that place Claude in ambiguous situations, and then reinforce Claude’s choice to pause, rather than to assume. …
… Pre-built MCP connectors: Access financial data providers and enterprise platforms for comprehensive market data and private market intelligence. Expert implementation support : Tailored onboarding, training, and best practices for rapid value realization. …
… AI novices with access to AI would need more time to acclimate to the technology, and AI novices without assistance would be less disoriented than our researchers who suddenly had Claude taken away from them. …
… AI is the first technology that can finally close that gap, which is why we're launching Claude for Small Business, alongside training and partnerships to make sure AI shows up for the entrepreneurs and communities who need it most. …
… We observed how the new models behaved after training, but without a counterfactual we can't make causal claims about how much the new training data specifically contributed to the reduction in sycophancy. …