From the filing
PDF page 241
Printed filing page 230 · Extracted page text
Read the page text
Table of Contents Every research project must ultimately advance our mission. This focus also guides what we choose not to pursue. For example, we do not currently focus on developing video and image generation models, despite their commercial appeal. That compute instead goes to coding, agentic skills, and safety research—a research decision that has also aided the capital efficiency of our models. Our research culture is highly collaborative and transparent. When multiple teams pursue parallel approaches, the norm is to consolidate behind the approach with the best empirical results rather than the loudest voice or most senior person in the room. This requires institutional trust we have built since our earliest days, and we believe it is a primary reason our research velocity has scaled significantly faster than headcount. How Research Becomes Product Anthropic was founded on the empirical conviction that Scaling Laws are the primary engine of AI progress. We operate with a sustained rigor in how we allocate research resources, including talent and compute, and measure whether each investment translates into scaled model performance. We believe that sustained and focused research can allow us to advance at an accelerated rate beyond what Scaling Laws alone predict through our compute multiplier advancements that have allowed us to push the frontier of model capability with better compute efficiency. Pre-training. Our pre-training process operates as a disciplined pipeline. Researchers propose potential improvements to model architecture, training methodology, or data composition, and each candidate is then validated at a smaller scale before we permit it to consume large amounts of compute. We set a high bar for inclusion in a frontier training run; we require evidence that demonstrates a given improvement holds as models increase in size. Frontier training runs are themselves planned capital deployments with defined inputs and strict inclusion criteria. Although not every run performs as our researchers expect, the process is designed to turn research investment into predictable capability gains with greater consistency over time. Post-training. Pre-training produces raw model capability. During post-training, that capability is refined into practical utility—the ability to write production-quality code, operate software tools, follow complex multi-step instructions, and act autonomously on a user’s behalf. This phase has evolved from a minor adjustment step into one of the largest areas of our research investment, reflecting evolution in model capability from conversational assistants to autonomous agents that perform tasks independently. Post-training centers on training models against complete, realistic tasks whose complexity reflects the demands of real professional environments. We have made substantial proprietary investments in developing this task portfolio, and its breadth and realism constitute one of our most durable scientific advantages, which we believe is difficult to observe from outside the organization and correspondingly difficult to replicate. Inference. Our model architecture and inference infrastructure are co-designed. AI researchers and inference infrastructure engineers share planning cycles, reporting lines, and performance targets. Model architecture decisions are made with inference economics as a priority, which is why successive versions of Claude models can deliver both higher capability and greater efficiency than their predecessors. Safety Research is Core to Our Business Any technology that is powerful enough to produce the enormous benefits that we believe AI will bring also poses risks. Responding to these risks requires a portfolio of safety research that spans how we train, evaluate, interpret, secure, and monitor our models. Anthropic’s Responsible Scaling Policy Our Responsible Scaling Policy describes how we approach the catastrophic risks posed by AI. The policy requires us to regularly publish detailed risk reports that reflect our perspectives on the overall level of risk posed by our models and how we are mitigating such risks. Our teams test each model extensively before release, assessing its capabilities in the domains most relevant to catastrophic risk and evaluating the extent to which the model can be misused. Our risk report describes an overall assessment of the results of these evaluations, the mitigations we have implemented to address these risks, and whether, and if so, why, we believe the identified risks are justified by corresponding benefits. As part of the current version of our Responsible Scaling Policy, we publish a Frontier Safety Roadmap, detailing our plans for measurable advances 230
Text is provided for reference. Check the original for tables, images, and exact formatting.