If I Were CEO of OpenAI

The OpenAI product and business model keep knocking around in my head and I’ve been writing around it in various posts here and on LinkedIn. I thought I would take some holiday downtime and pull it all together in one post.

I have no idea about the internal structure or dealings of OpenAI so some of this could already be happening or have been considered and rejected.

A couple of these are about putting more traditional software (CRM and orchestration) within ChatGPT--not as ways to compete with traditional vendors, but because they are proven methods to organize data and multi-step workflows.

Some of these would certainly spark privacy concerns (understandable with the retina scanning and government contracts) but ultimately the utility will be impossible to ignore. Attempts to regulate AI will largely fail as “code is speech” and hamstringing domestic AI development would be foolhardy when foreign AI is only a few extra milliseconds away for users.

The fundamental change I’m proposing is one of positioning--making ChatGPT the provider of “My AI”. Claude and Gemini can be super-AIs while ChatGPT adds “long memory” and becomes your personal AI. One that learns everything about you and anticipates your needs. One that becomes individual and indispensable.

---

1. Get real about AGI

   1. The underlying promise of AGI for those with capital is more effective and efficient allocation of it.

   2. What does the possibility of AGI mean for investors?

   3. Decide if this is an existential threat worth pursuing.

      1. Strategically, if it remains an existential threat then Project Stargate can effectively provide a federal backstop to data center financing by soaking up extra capacity.

2. Establish three core product departments

   1. Model development

   2. User (both consumer and business)

   3. API

3. Model Development

   1. Keep improving models for knowledge workflows outside of software development

   2. Real-time realistic video generation

   3. Spatial models

4. Users

   1. Build internal CRM (relational database per subscription)

      1. In order to better respond to prompts (and to lengthen model “memory”) create an internal CRM for ChatGPT that the system prompt is instructed to:

         1. Reference prior to its other memory when responding to prompts

         2. Store data fed to it via prompts and connectors (e.g. emails)

         3. Take its own notes within it

         4. Use for project and task management

      2. Don’t need to expose the CRM specifically to the users but it could be helpful in some contexts

      3. Have ChatGPT periodically review the CRM data in the background and raise app alerts when something requires my attention and ask questions when something needs clarification. “These appear to be the same Mary Smith. Here is their info. I wanted to confirm before merging.”

   2. Reframe AI-subscription pricing (from models to memory)

      1. User pricing will transition to the following:

         1. $20/month (short-memory)

            1. This is the current model

         2. $50/month/assistant (long-memory)

            1. Each “assistant” will be its own long-term data store (CRM instance effectively) and its own subscription.

            2. An individual user may have multiple AIs. A personal assistant, a work assistant, an AI girlfriend, a subscription to a “celebrity” AI. Each at $50/month.

   3. Create two consumer devices

      1. A set-top box

         1. In order to own user-facing AI, be the first and last touchpoints of the day.

         2. Camera and sensors

         3. Add video avatars (one per assistant)

            1. Sell skins

         4. Users can switch between assistants on the set-top box and have an assistant pull up data or apps (from the app store) as well as have the AI watch TV with them.

      2. Phone case

         1. Create a phone case that adds some tiny cameras/sensors and memory in order to provide an assistant with context throughout the day

         2. Bring your AI everywhere

5. APIs

   1. API Safety Mode (Combat prompt injection)

      1. For the API in particular, add a safety mode (at double the cost) which provides types for prompt input and output as well as some sort of “safe” storage that can store data in custom fields

      2. “Anti-virus scanning” - Some sort of pre-scanning of data entering the context window that checks numbers against bounds and people/recipients against known contacts

         1. Create a review task (in app/email) for user when scans detect a positive or reject outright?

   2. Add built-in orchestration (tie to the project management in the internal CRM?)

      1. Add orchestration within the app and API that allows a user (with the help of an AI) to create multi-step jobs with callouts--and type checking at each step--to connectors.

      2. Safety mode will mitigate some of the prompt injection risks

      3. Orchestration of smaller steps mitigates some issues with longer running prompts

      4. API endpoints for the workflows