Comment by Gareth321
18 hours ago
> Put all of these trends together, and I think that in 10-15 years, we are going to have consumer PCs (and phones!) running models doing pretty much anything that frontier models can do right now.
At current pace, we'll have open weight LLMs with frontier intelligence in 6-12 months. The constraint is RAM - both for the model and the context. It's likely that distillation and quantisation and TurboQuant will significantly reduce RAM requirements. I think we'll have Opus 4.8-like performance on 64GB of RAM in two years.
Of course, by then, frontier intelligence will be god-like.
> god-like
So would you say we are months away from full self-driving cars that can out-drive a human being in any situation?
Remember, these cars are running local LLMs, not frontier models. The issue with self-driving cars has been the edge cases. The 0.0001% of situations where the models did not have sufficient training data. This is compounded by the hardware limitations. Onboard RAM in a typical Tesla on the road is 16GB (+16GB for the backup computer). This has to run the existing onboard OS and other operations plus the LLM. These two factors combined means that the cars are currently incapable of negotiating the 0.01% cases, let alone the 0.0001% cases. And this is compounded by the fact that LLMs cannot currently update their weights in real-time, like humans. It takes months to train a new model. Special small models can be very tricky, especially around safety and mission critical applications like FSD.
All that said, current data shows that FSD is already better than human drivers on average. See the recent regulatory decisions by the Dutch and Danish road safety authorities. So we've already crossed the rubicon. All improvements now are icing on the cake. My prediction is that local LLMs will get much better, very fast. How that's operationalised with Tesla (or other) data is yet to be seen. They have at least three new ASCIs/SoCs in the roadmap for improved LLM efficiency and with a lot more RAM. Plus they just announced new technologies allowing the local LLMs to learn from driver intervention and behaviour. Some form of vectorised RAG, which could mitigate a lot of the limitations around real-time learning.
I am very optimistic for the future of self driving. I own a Tesla with FSD now, and it's incredible. It makes mistakes, but fewer than I do, and so far has saved my butt (and my wife's) several times from obstacles and emergencies we would not have seen. The car has undeniably made us safer.
> so far has saved my butt (and my wife's) several times from obstacles and emergencies we would not have seen.
Honestly, you need to reflect on your driving habits. FSD has only been usable for two or three years maybe? And you already encountered MULTIPLE situations requiring active safety intervention to save you during this time?
You cannot rely on the extra safety it provides. A driver with basic competence should be able to avoid most risks through anticipation before they happen.
1 reply →