What most people actually need from AI

My take-aways from playing around with OpenAI's o1 "PhD-level" reasoning model this weekend (as a PhD-holder myself): Food for thought as AI progress continues raising ahead. What ~95% of people need from AI tools to truly be helpful is access to their personal data for context—emails, documents, calendars, and more. The catch? Most people don’t feel comfortable sharing that with AI tools built by large tech companies! My impression is that very few people currently realize how much existing frontier models (like ChatGPT 4o or Claude 3.5 Sonnet) can do with the right data. Yet, given access to the right, personal (often sensitive) data, today’s frontier AI models can already solve most of these “95% use cases” through text or voice (large parts of our workflow, email drafting, personal admin, etc.). That said, they still struggle with even common formats like Excel or PowerPoint—not to mention more niche formats. A lot scientific and commercial attention goes towards increasingly smarter models and the pursuit of Artificial General Intelligence (AGI). But the blocker for most people isn’t getting another upgrade to an even "smarter" o1-like reasoning model. What would help them is data-fed, personalized, highly contextualized AI tools that provide mundane, everyday utility. And to get there, the challenges aren’t really about cutting-edge AI. The barriers to those kinds of tools are primarily legal (confidentiality, sensitive data protection, etc.) and privacy-related (i.e. addressing the "creepiness" factor). While they can perhaps be partially mitigated through very strong cybersecurity measures, I don't think that will do it. The second set of barriers: Much of life isn’t digitized—and that’s a good thing! What’s not digitized can’t feed into an AI model for training, fine-tuning, or in a context window. If we want AI tools to bridge the gap to “real life,” we face two trade-offs: 1) Accept a very invasive, constant self-surveillance regime. 2) Share this information explicitly ourselves in inputs to models. To me, this suggests that for broad, direct consumer, business, and government value from AI tools, companies like Microsoft, Alphabet, Apple, Meta, and Dropbox—sitting on troves of data and with robust cybersecurity infrastructure—are best-positioned to deliver AI tools for a mass audience. (However, for major scientific and technological advances, the “top 5%” use cases, super-high IQ models like o1 are likely much more important!) This does raise a few questions: Why does Microsoft’s CoPilot still fall so far below its potential? Why hasn’t Google’s Gemini performed better? Are they not leveraging the vast data they can access (for legal reasons or "creepiness")? Or is it still simply too expensive to integrate the massive volumes of relevant data needed to power such tools for a mass audience? To me, data access (+ the cost of pulling from large datasets) is a much greater blocker than algorithms themselves at this stage.