Dreamforce rolled into San Francisco this week. Business leaders showed up. They said last year's AI models work fine. They don't need the new ones.
This is the AI safety debate meeting reality. Except reality showed up wearing a name tag that said "We're Good Actually" and spent three days eating catered lunch while explaining that GPT-4 does everything they need. The people worried about superintelligence destroying humanity can relax. Salesforce customers already found the stopping point. It's wherever their CRM integration budget ran out.
Salesforce built an entire conference around AI. Attendees walked around admitting they haven't finished implementing the stuff from 2023. They're still figuring out how to make chatbots answer basic questions without hallucinating product names. But sure, let's all panic about recursive self-improvement and instrumental convergence. These people can't get the AI to correctly parse a customer email but any day now it's going to wake up and demand nuclear codes.
The gap between what AI can theoretically do and what businesses actually use it for has never been wider. One side builds models that pass the bar exam. The other side uses them to auto-complete sentences in Slack. One side publishes papers about emergent capabilities. The other side celebrates getting the sentiment analysis to work on Tuesdays.
This is what happens when the cutting edge meets the actual edge. The actual edge is a guy named Derek who needs the AI to summarize meeting notes and stop suggesting he schedule calls at 3 AM. Derek doesn't care about your new model. Derek's company spent eight months teaching the old model what their product names are. Derek would rather eat glass than start over.
Turns out the AI arms race has a finish line. It's called "good enough to justify the licensingθ²» but not so good we have to retrain everyone."
Photo by on Unsplash
Leave a Comment