Sunday Links: OpenAI for Business, AI Escape, and AI-powered Roll-Ups

Today's post is late in the day, courtesy of the US West Coast timezone I'm in (I was at the AI Engineer World's Fair last week, and I'll be at the Databricks Summit next week - ping me if you are too!). More on those two events soon. For now, though, on to a few links from this week:

• OpenAI hits 3M business users and launches workplace tools to take on Microsoft. OpenAI launched integration with a range of enterprise products from Box.com to Microsoft Teams. This seems a little reckless in terms of expanding their surface area and upsetting their key partner, Microsoft. But, I think it can be explained in one word: Context. It's become obvious that for AI to do a better and better job for knowledge work, it needs to be connected to all of your enterprise activity, from personal notes to docs. With this move, OpenAI is seeking to secure that. Competitors will raise privacy/security concerns, and that will probably hold back adoption to some extent, but it won't stop everyone.
• Early AI investor Elad Gil finds his next big bet: AI-powered roll-ups. Elad is a very astute investor, and I'm highlighting this because you don't hear it that much in the investment media sphere. Taking people-intensive businesses with significant customer bases and AI-enabling them, as you consolidate to get critical mass. I think we'll see a lot more of this as AI clears the bar of doing more and more knowledge work. I'd also go beyond what Elad says and add that AI will also make it easier and cheaper to migrate back office systems and add new services. What counts here is buying the right players with the right customer bases. Having growing rollups in a space will likely also accelerate automation since everyone will feel the pressure.
• AI Is Learning to Escape Human Control. This article in the Wall Street Journal [Paywalled - apologies] is an example of how not to think about AI alignment. Triggered by recent results like Claude 4 Opus attempts to call the authorities in tests and recent experiments with OpenAI o3, the conclusion being drawn is that AI needs to be aligned with human motives. While I don't disagree with the general idea, it's important to realize that doing that in the general case is nearly impossible. First, because there are many value sets, second, because value decisions are often highly contextual, and third, AI is unlikely to be good enough to follow those directives yet. What's happening with these "escape" scenarios is that much of the training data for AI is fiction and pseudo-aligned content (most movies end with the virtuous side winning), so it is a kind of proto-alignment. What's going to be needed for many systems is less of that kind of content in training. Unless you can give an AI full context to make moral decisions, you are going to need a human in the room if a system is going to make such decisions.
• Mary Meeker's 2024 Trends Report –Artificial Intelligence. The Analyst Mary Meeker is known for her annual tech trends report, and this year's is a blockbuster (340 dense slides). Unsurprisingly, it's on AI. Many of the trends are widely written about, but it's a good reminder of just how much consumer adoption there is, especially. No technology has reached so many so fast. Another, slightly more hidden nugget is that the cost of inference, especially, has come way down (by 97% according to the data here). Industry analysis tends to focus on the high cost of training (which is still growing for frontier models), but if inference costs can be compressed, then perhaps we won't need quite as much power as we thought we did. Very likely, this drop comes from firms rapidly figuring out what the minimum acceptable model size is for specific tasks.
• AI Engineering Talks! I'll save more event content for next week, but in the meantime, here is a set of links to the excellent tracks at the AI Engineering event. Three thousand people battling through a wifi desert and still making it out the other side! I spent most of my time in the MCP and Evals tracks, but there is plenty of other good stuff! Go to the World's Fair site for the program so you can get to the right point in the streams.

Wishing you a happy Sunday!