AI Safety: Constitutional AI vs Human Feedback

We review four clips from the Dwarkesh Patel Podcast with Satya Nadella, Microsoft's CEO. I highly recommend Dwarkesh’s show—technical & nerdy, but excellent.Satya talks about scaffolding—the software wrapped around AI models to make them actually work.So we speak with someone building that scaffolding: Neil McKechnie runs two AI-first startups as a CTO. He discusses how he orchestrates up to twelve different language models—GPT-5, Claude, Gemini, Llama, Mistral, Cohere, Perplexity. We discuss what it actually takes to build production systems with LLMs today—and what that reveals about the agent future we’re being pitched.Dwarkesh's Podcast: https://www.youtube.com/@DwarkeshPatelTo stay in touch, sign up for our newsletter at https://superprompt.substack.com

Description:AI agents from OpenAI, Google, and Anthropic promise to act on your behalf—booking flights, handling tasks, making decisions. What kind of agency do these systems actually have? And whose interests are they serving?Enterprise AI agents are already deployed in customer support, code generation, and task automation. Consumer agents—ChatGPT Agent Mode, personal task assistants—face a wider gap between marketing promises and actual capabilities.The alignment problem: agents need access to your calendar, email, and personal preferences to help you effectively. But the agent that knows you well enough to serve you is also positioned to steer you. When you delegate decisions to an agent, who decides what success looks like?To stay in touch, sign up for our newsletter at https://superprompt.substack.com

With great power comes great responsibility. How do leading AI companies implement safety and ethics as language models scale? OpenAI uses Model Spec combined with RLHF (Reinforcement Learning from Human Feedback). Anthropic uses Constitutional AI. The technical approaches to maximizing usefulness while minimizing harm. Solo episode on AI alignment.REFERENCEOpenAI Model Spechttps://cdn.openai.com/spec/model-spec-2024-05-08.html#overviewAnthropic Constitutional AIhttps://www.anthropic.com/news/claudes-constitutionTo stay in touch, sign up for our newsletter at https://superprompt.substack.com

Notable open source large language models from Meta, French AI company Mistral (valued at $2B), Microsoft, and Apple. Not all open source models are equally open—the restrictions and licensing constraints you need to know before deploying one. Solo episode.To stay in touch, sign up for our newsletter at https://superprompt.substack.com

Why enterprises and entrepreneurs choose open source LLMs like Meta's Llama—cost-effectiveness, control, privacy, and security. The safety and ethics debate: which poses greater risk to humanity, open source or proprietary AI models? Both? Neither? Solo episode on open source LLMs.To stay in touch, sign up for our newsletter at https://superprompt.substack.com

Beyond ChatGPT and Gemini: Anthropic's Claude and the $4 billion Amazon investment. How AI industry benchmarks work, including LMSYS Arena Elo and MMLU (Measuring Massive Multitask Language Understanding). How benchmarks are constructed, what they measure, and how to use them to evaluate LLMs. Solo episode.Anthropic's Claude https://claude.ai [Note: I am not sponsored by Anthropic]LMSYS Leaderboardhttps://chat.lmsys.org/?leaderboardTo stay in touch, sign up for our newsletter at https://superprompt.substack.com

Recent updates from Google and OpenAI feature multimodal capabilities—AI that processes multiple input types simultaneously. Why multimodal models outperform single-modality systems, demonstrated through a hypothetical chatCAT that helps owners understand their cats. Solo episode on multimodal architecture.To stay in touch, sign up for our newsletter at https://superprompt.substack.com

Google's Gemini family of multimodal AI models compared to OpenAI equivalents. What Nano, Pro, and Ultra each do, how they compare to GPT-3.5, GPT-4, and ChatGPT, and what "multimodal" means in practice. Solo episode on Google's LLM strategy. https://gemini.google.com. (Note: I am not sponsored by Google.) To stay in touch, sign up for our newsletter at https://superprompt.substack.com

Creating a travel planning AI using OpenAI's GPT Builder with no code. Seven takeaways from building a custom GPT, what the process reveals about prompt engineering, and the constraints of the no-code approach. Includes demonstration of the Holiday travel GPT. Solo episode. You'll need to be Open AI Plus subscriber to access this GPT. (Note: I am not sponsored by Open AI.) https://chatgpt.com/g/g-FURrBQAh8-holidayTo stay in touch, sign up for our newsletter at https://superprompt.substack.com

Deploying large language models for enterprise applications. Rackspace CTO Jeff DeVerter discusses implementing Google PaLM for sales, enabling Azure and AWS customers, and why your LLM choice should probably match your cloud provider. Private cloud deployment for data confidentiality and thinking in data pipelines from the start. Featuring Jeff DeVerter, Rackspace Technology.To stay in touch, sign up for our newsletter at https://superprompt.substack.com