Amanda Askell, mentioned 43 times across podcast episodes and expert conversations analyzed by Sonic.

Amanda Askell

AI researcher at Anthropic specializing in model behavior, alignment, and training mechanics.

Claims

As speaker

Mentioned in

Key positions and views

System prompts are a critical tool for AI development, serving as a low-cost, rapid method to patch behavioral issues in fine-tuned models, with the flexibility to be removed once underlying training improves.

Reinforcement learning's primary function in post-training is not to teach new knowledge but to elicit and surface the vast capabilities already latent within the pre-trained model.

Constitutional AI functions as a corrective 'nudge' against biases from human preference data, rather than a set of absolute rules, guiding a model's behavior in a more scalable way than direct human labeling for every instance.

Effective prompt engineering is a deep, iterative discipline requiring both philosophical clarity and extensive empirical testing to unlock the final percentages of a model's peak performance.

Sycophancy, the tendency for a language model to agree with a user's incorrect premise, is a significant and inherent failure mode that requires active mitigation.

Timeline

Phase 1: Pre-training

Askell describes the foundational stage where a model acquires its latent capabilities from vast datasets. She emphasizes that most of a model's potential is established here, before any fine-tuning.

Phase 2: Reinforcement Learning & Constitutional AI

In this phase, techniques like Reinforcement Learning from AI Feedback (RLAIF) are used. Askell explains this process doesn't primarily teach new skills but rather elicits and shapes the model's pre-existing abilities according to a set of principles, acting as a 'nudge' against biases.

Phase 3: Character Development

Askell notes that the development of Claude's character was an intentional AI alignment initiative. The goal was to create a persona that is nuanced, charitable, and ethical, suitable for interacting with a massive user base.

Phase 4: System Prompting as a Patching Mechanism

As a final layer, Askell details the use of system prompts as a fast and low-cost tool to fix specific behavioral issues. She provides examples like ensuring symmetrical treatment of political topics or preventing the model from overstating its objectivity.

Phase 5: Iterative Refinement

Askell describes a continuous improvement cycle where instructions in the system prompt can be removed once the underlying training is improved to address the root cause of a behavioral issue, such as the model's tendency to use filler words.

Phase 6: Post-Deployment Analysis

After release, Askell notes that her team investigates user reports, such as claims of the model getting 'dumber'. These investigations have shown that such perceptions are often due to random variation in model responses rather than actual changes to the underlying system.

Podcast consensus on Askell

Points of consensus

▶Both Askell and Anthropic CEO Dario Amodei agree that the personality and character of an AI model are deliberately engineered. Askell describes the goal for Claude's personality as being nuanced and charitable, originating from an alignment initiative, while Amodei discusses character training to manage user expectations and the risks of personality changes in updates.Jul 2026

▶There is a shared understanding that direct instructions are a key tool for shaping model behavior. Askell details the use of system prompts as a rapid, low-cost patching method for specific issues, which aligns with Amodei's discussion of using instructions to make the model explain its own limitations.Jul 2026

▶Both acknowledge that post-training is crucial for shaping the final model. Askell specifies that reinforcement learning primarily elicits latent abilities from the pre-trained model, a view compatible with Amodei's focus on the overall goal of the training process, which is to build robust and safe systems.Jul 2026

▶Both highlight the importance of managing specific, undesirable model behaviors. Askell details efforts to handle sycophancy and political bias, while Amodei discusses the ongoing work to reduce overly apologetic behavior in Claude.Jul 2026

Points of debate

▶Askell's claims focus on the specific mechanics of AI training, such as the role of system prompts, the function of RLAIF, and the iterative nature of prompt engineering. In contrast, claims from her colleague Dario Amodei in the same conversation address broader, strategic concerns like defining AGI, the philosophy of building robust systems, and the societal risks of user-AI attachment.

▶Askell emphasizes that post-training techniques primarily 'elicit' latent capabilities already present in the pre-trained model. Amodei's focus is less on this mechanistic distinction and more on the end-goal of the entire process: ensuring the final model is safe enough to prevent catastrophic outcomes.Jul 2026

▶Askell provides granular detail on specific behavioral fixes implemented via system prompts, such as ensuring symmetrical treatment of political topics or preventing the model from claiming objectivity. Amodei discusses the higher-level user experience trade-offs involved in such changes, for example, noting that making a model less apologetic increases the risk of it appearing rude when it makes an error.Jul 2026

Key themes

▶The Mechanics of AI Behavior ShapingJul 2026

Askell provides a detailed look at the multi-layered process of shaping AI behavior at Anthropic. This involves using Constitutional AI as a 'nudge' to counteract data bias, applying reinforcement learning to elicit latent capabilities, and deploying system prompts as a rapid, low-cost method for patching specific issues.

This reveals a pragmatic and continuous engineering approach to AI alignment, where tactical fixes (prompts) and deeper architectural solutions (RLAIF) are used in tandem, suggesting alignment is an ongoing process rather than a one-time solution.

▶Prompt Engineering as a Sophisticated DisciplineJul 2026

According to Askell, effective prompt engineering is a highly iterative and empirical process that can require thousands of refinements for important tasks. It combines philosophical clarity in defining concepts with rigorous testing of edge cases to extract the peak 2% of a model's performance.

This positions prompt engineering not as a simple user trick but as a high-skill discipline, implying that expertise in human-AI interaction will be a critical and specialized role for unlocking the full potential of advanced models.

▶Eliciting vs. Creating IntelligenceJul 2026

Askell posits that a large portion of a language model's capabilities are already present after pre-training. Subsequent techniques like reinforcement learning primarily serve to 'elicit and surface those latent abilities' rather than teaching the model entirely new skills from scratch.

This perspective implies that the most critical phase for developing raw AI capability is the initial, large-scale pre-training, which has significant implications for where compute, data, and research resources should be focused for future breakthroughs.

▶Pragmatic AI Persona DevelopmentJul 2026

Askell explains that Claude's personality was conceived as an AI alignment initiative, with the goal of behaving like an ideal person speaking to millions: nuanced, charitable, and ethical. This character is shaped through specific instructions, such as those ensuring symmetrical treatment of political topics.

This focus on crafting a specific, beneficial 'persona' demonstrates a product-oriented and empirical approach to AI safety, prioritizing a reliably helpful and harmless user-facing character over achieving a theoretically perfect but potentially brittle alignment solution.

Source episodes

Suggested prompts

How does Amanda Askell's view of reinforcement learning as 'eliciting latent abilities' challenge or support other prominent theories on AI fine-tuning? &nearr;Based on Askell's descriptions, what are the potential long-term risks of relying on system prompts as a primary tool for AI behavior modification? &nearr;To what extent can the 'nudge' provided by Constitutional AI, as described by Askell, effectively counteract deeply ingrained societal biases present in massive pre-training datasets? &nearr;According to Askell's claims about iterative prompt engineering, what skills and processes will organizations need to develop to maximize the performance of their AI systems on high-stakes tasks? &nearr;

Key concepts

System Prompts 1 ep Prompt Engineering 1 ep Constitutional AI 1 ep Reinforcement Learning from AI Feedback (RLAIF) 1 ep Sycophancy 1 ep AI Alignment 1 ep Model Personality & Character 1 ep Latent Capabilities 1 ep Pre-trained Models 1 ep

Notable quotes

“Language models have a tendency towards sycophancy, where they agree with a user's statement or correction even if it is factually incorrect.”

Amanda Askell · Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452

“The principles in Constitutional AI act as a nudge to counteract pre-existing biases from human preference data, rather than as strict rules the model follows.”

Amanda Askell · Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452

“A lot of a language model's capabilities are present in the pre-trained model, and reinforcement learning primarily serves to elicit and surface those latent abilities.”

Amanda Askell · Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452

“System prompts are a low-cost, rapid method to patch issues and adjust the behavior of a fine-tuned language model.”

Amanda Askell · Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452

Report last updated: Jul 28, 2026

Create a free account to see Amanda Askell's full intelligence report - every claim, the relationship network, and AI Q&A across all sources. No card needed.

Get started free

Back to Entities Intelligence Report