So, what’s the ‘large‘ and ‘small’ model difference?
I really feel that SMARTER ROUTING would make the version I’m on 1000% better. I’ve had it avoid boxes in the road, pull over for emergency vehicles and move over so I didn’t get side swiped all before I saw any of that. Routing is the weak link, dumb routing.
Is the safest job out there the one where guys are paid to dig up and move goalposts?
To answer your question, it's context size and parameter count. The way I understand it is a small model is like a specialized tool like a pocket knife, where it can do a lot of things but can't do any individual one remarkably well. Whereas a large model is like having a dedicated tool box with every tool you would need in it. Or another way to describe it is the small model doesn't "understand" the relationship between a ball rolling into the street and a child likely following it. It just sees "Object: Ball." And a large model doesn't just see a "pedestrian"; it understands that a pedestrian standing near a bus stop with a backpack is likely waiting for the bus, whereas a pedestrian looking over their shoulder at a crosswalk is likely about to step out.
But I get what you're saying. It's frustrating that they hype this up to then only let us down with delays. And to make it worse, I feel that the next goalpost is going to be "Sorry, we can't actually do the 10x parameter count on AI4 because there is not enough compute or memory, so you will need AI5." Then we finally get AI5, and that will suddenly not be enough either. I understand the logic behind not overbuilding for more than you need from a business perspective, but boy does it suck really bad from a consumer perspective, especially when the upgrade costs $50,000 to $60,000 and a lifetime ongoing subscription.
I just know we are likely already reaching the point where they can't really do much more with AI4. Some of the more "open" early access testers have noted this: the limitations of AI4's memory/context window are simply too short to make it work very well in certain areas like parking, as an example. It's why you often see it circling the same area in parking lots over and over instead of moving into a different part of the parking lot to continue searching for an available space. Or when it gets stuck in parking garages trying to exit and just does another infinite loop. It simply doesn't "remember" what it's done or where it's been.
I like to compare it to an AI chatbot, and imagine if you were trying to have a conversation with Grok, Gemini, or ChatGPT, and it only answered individual questions, but if you asked it a follow-up question or tried to hold an actual conversation, it just completely forgot the context of the original question that you asked.
www.notateslaapp.com
