IIRC Elon has been talking about merging the two stacks since last November (or earlier). James Douma recently explained why there has been such a big delay in combining the old AP stack with the new FSD stack:
The problem is that FSD now models the world with a top down look on a 3D model they call "vector space" (argh). This works great in cities when there is nothing of importance far away. But they are having trouble modelling highways in vector space because it's a much larger physical space and requires too much storage and/or processing.
I think they will need to use a different model for things that are far away. IMO this is what humans do because things that are far away are usually not an immediate threat. I do the same thing with time. Events that will occur 2 or more weeks away are usually put in the "infinity" bin. George Gamow expressed a similar idea in his book One Two Three... Infinity. This is a very natural thing for humans to do.
I imagine the Tesla dev team is struggling with this because they have put so much effort into, and created so many expectations for, a single model of everything. Having two models may essentially be like having two stacks and might be worse because currently they are using only one stack at a time. Creating two models takes more work and merging them when needed also takes more work. This is similar to the problem of merging radar data with vision data.
It's pretty easy to identify the harder problems because there is a long delay between when Elon announces they will be solved and when they actually get solved. IMO it's a mistake to rely on Elon's expectations. A much better measure is to graph Elon's estimates over time. True FSD has been a year away for many years now. Likewise, Elon's estimate for how much better the current hardware will be than a human dropped by a factor of 3 or more last August. My guess is there will be at least two more hardware iterations before true FSD beats humans by a factor large enough to satisfy regulators.
There is an engineering adage that the last 10% of a project takes 90% of the work. I don't think neural net modelling of the world for FSD is immune from this even though other AI projects, with a much more limited domain, such as AlphaZero (which plays well defined games) do seem immune. IMO the 10%/90% rule arises because the problem domain is usually not fully understood until the project seems to be nearly done.
The problem is that FSD now models the world with a top down look on a 3D model they call "vector space" (argh). This works great in cities when there is nothing of importance far away. But they are having trouble modelling highways in vector space because it's a much larger physical space and requires too much storage and/or processing.
I think they will need to use a different model for things that are far away. IMO this is what humans do because things that are far away are usually not an immediate threat. I do the same thing with time. Events that will occur 2 or more weeks away are usually put in the "infinity" bin. George Gamow expressed a similar idea in his book One Two Three... Infinity. This is a very natural thing for humans to do.
I imagine the Tesla dev team is struggling with this because they have put so much effort into, and created so many expectations for, a single model of everything. Having two models may essentially be like having two stacks and might be worse because currently they are using only one stack at a time. Creating two models takes more work and merging them when needed also takes more work. This is similar to the problem of merging radar data with vision data.
It's pretty easy to identify the harder problems because there is a long delay between when Elon announces they will be solved and when they actually get solved. IMO it's a mistake to rely on Elon's expectations. A much better measure is to graph Elon's estimates over time. True FSD has been a year away for many years now. Likewise, Elon's estimate for how much better the current hardware will be than a human dropped by a factor of 3 or more last August. My guess is there will be at least two more hardware iterations before true FSD beats humans by a factor large enough to satisfy regulators.
There is an engineering adage that the last 10% of a project takes 90% of the work. I don't think neural net modelling of the world for FSD is immune from this even though other AI projects, with a much more limited domain, such as AlphaZero (which plays well defined games) do seem immune. IMO the 10%/90% rule arises because the problem domain is usually not fully understood until the project seems to be nearly done.
