The progress of AI is fundamentally and predictably driven by scaling laws; increasing compute, data, and model size reliably yields more capable systems.
Training AI models on publicly available internet data constitutes 'fair use' and is a legally reasonable basis for building large language models.
Proactive and transparent safety development is essential. This includes creating tiered risk frameworks like AI Safety Levels (ASL), publishing responsible scaling policies, and investing heavily in research like interpretability and scalable supervision.
The future of AI lies beyond simple Q&A chatbots and in more 'agentic' systems that can use tools, self-correct, and perform long-horizon tasks, ultimately automating complex knowledge work.
AI companies should be transparent about their risk assessments and safety testing, and government regulation should be explored, with states acting as 'laboratories' for different policy approaches.
Pre-Anthropic
Identified by Ilya Sutskever as a key contributor, alongside Alec Radford and Dario Amodei, to the work that ushered in the 'age of pre-training' in AI.
2019
Kaplan notes this as the period when precise scaling laws began to be observed over many orders of magnitude, providing conviction that AI capabilities would improve predictably with scale.
2022
Co-founds Anthropic, which at the time was a research company without a product. The initial motivation was to leverage computational resources unavailable in academia to scale AI, leading to the first development of Claude as a dialogue agent.
2023-2024
As a public face for Anthropic, Kaplan's discourse shifts to business strategy, product capabilities, and policy. He discusses partnerships with Amazon, articulates the company's 'fair use' legal stance, advocates for state-level AI regulation experiments, and details safety frameworks like the Responsible Scaling Policy and AI Safety Levels (ASL).
Near Future
Kaplan's focus is on the next generation of models (e.g., 'Claude 5') trained on new large-scale clusters like Trainium 2. He predicts the achievement of human-level AI within 2-3 years and emphasizes the critical need for 'scalable supervision' to manage the risks of increasingly powerful systems.
▶The Unwavering Faith in Scaling LawsJul 2026
Kaplan's worldview is fundamentally anchored in the existence of 'clean scaling laws,' the principle that AI performance predictably improves with more compute, data, and larger models. He views this as the primary driver of AI progress and believes that when scaling appears to fail, it is due to implementation errors rather than a failure of the underlying principle. This conviction extends to both pre-training and reinforcement learning.
For investors, Kaplan's conviction in scaling laws signals that Anthropic's core strategy is a capital-intensive race for computational superiority, making their access to cutting-edge hardware like Trainium 2 a critical competitive advantage.
▶Proactive and Evolving AI Safety
Kaplan details a multi-faceted and evolving approach to AI safety, moving from foundational techniques like Constitutional AI to more advanced concepts. This includes developing AI Safety Levels (ASL) to classify risks, using AI to red team other AIs, and pursuing 'scalable supervision' where AI systems monitor and steer other AIs. This reflects a belief that safety mechanisms must advance in lockstep with model capabilities.
Analysts should note that Kaplan's safety narrative is shifting from static, principle-based alignment to dynamic, AI-driven oversight, suggesting that future safety solutions will be as computationally complex as the models they are meant to control.
▶Strategic Vision Beyond the ChatbotJul 2026
Kaplan articulates a business strategy for Anthropic that moves beyond the simple question-and-answer chatbot paradigm, which he views as 'limiting'. The focus is on enabling more complex, 'agentic' capabilities and empowering customers to build novel applications. This is supported by a legal stance defending the use of public data for training as 'fair use' and a policy of not directly competing with customers.
This theme indicates Anthropic is positioning itself as a foundational infrastructure provider for a future AI-driven economy, prioritizing ecosystem enablement over building a single dominant consumer-facing application.
▶The Imminent Arrival of Transformative AI
Kaplan consistently projects a future of rapid, transformative AI progress, with timelines for human-level AI shrinking to just 2-3 years. He predicts AI will soon tackle long-horizon tasks that take humans days or weeks, automate knowledge work, and even surpass human experts in fields like theoretical physics. This rapid progress is fueled by 3x to 10x annual gains in efficiency and capability.
Kaplan's accelerated timeline for AGI serves as a call to action for developers to build for future capabilities, but it also places immense pressure on policymakers and safety researchers to address existential risks on a much shorter timescale than previously anticipated.