The crash analysis below shows AI3 FSD v12.6.4 biggest safety weakness: late reaction.
This video uses official dashcam crash footage synchronized with the vehicle’s official telemetry logs obtained from Tesla and visualized in @CamCutApp.
The telemetry shows FSD waiting about 0.67 seconds after the lead vehicle’s brake lights illuminate before it begins braking. From the moment the lead vehicle starts braking until Automatic Emergency Braking intervenes is about 1.5 seconds. During that time, the Tesla slows by only 4 mph, from 62 mph to 58 mph.
By the time AEB takes over, only 132 feet remain before impact. There simply isn’t enough distance left to avoid the collision.
Millions of AI3 vehicles are still running FSD v12.x. Situations like this require the driver to intervene before it becomes obvious that FSD isn’t braking quickly and aggressively enough, and that’s not always easy to judge in real time.
Tesla today announced AI3 v14 Lite which includes “significantly improved safety.” When I get a chance to evaluate v14 Lite, this is exactly what I’ll be testing first. In my opinion, it’s the most important and critical improvement the system needs.
Tesla FSD Vs. @comma_ai 4
Same road. Same car.
Car: Tesla Model Y Performance (2022)
My impression: On a straight highway, I honestly don't feel much difference between Tesla FSD and Comma 4.
The biggest difference is at highway entrances and exits, where FSD is a bit more confident. But that's only about 10% of my daily driving. The other 90% is straightforward highway cruising, where both systems perform similarly.
Would I pay $8,000 or $100/month for that difference? Not anymore.
And city driving on HW3? I think it's overrated. I can drive better than HW3 in most city situations. Unless FSD V14 Lite changes the game, Comma 4 wins this comparison for me.
Sorry for the Thai voiceover in this video. I'll try to add English subtitles later.
Karpathy found a way to reduce token consumption by 90%
The problem is that the LLM re-reads the same files over and over again, loses context between documents, and provides less accurate answers as a result
The solution is called Wiki Layer the LLM cleans, structures, and links all your data once, after which it never works with raw files again
Three folders `raw/` for originals, `wiki/` for a clean knowledge base in Markdown, and files with rules for the agent
Result up to 90% token savings on repeat queries, automatic links between documents, and a visual knowledge graph in Obsidian
Everything stays on your local machine nothing goes to the cloud
@arena@OpenAI it's quite funny that no one in the comment section believe what you said at all.
I have both ChatGPT5.5 Pro and claude 4.7 (pro)
I feel like ChatGPT find me more bugs that claude did.