Just like your LLM has memories, FSD needs memories too. It needs to remember you. The experiences you’ve had together in the past.
I don’t know how Tesla is implementing this, but I imagine it could work much the same way. A video of an intervention could be passed to a model and asked to condense the intervention down to a reason. “Too close to the curb”. You would have to have a model that is very good at figuring out intervention reasons from a video — but I bet Tesla has had something like that driving improvements for quite some time. Maybe you could even just use something like Grok.
Then, this model would simply write a memory, or sort of add to the system prompt, the lesson learned from that intervention. “Don’t drive too close to the curb on Elm street”. The current FSD model can already accept text prompts in and output text in additional to camera and other I/O
This would be a massive paradigm shift for FSD. Instead of having to wait for the next update to come out for your issue to get hopefully solved, you could “patch” the model yourself with your input. The problem would be fixed without requiring a full on model update, retraining and, evaluation / QA
A model that is a major step forward, but with several regressions could potentially have undesirable behaviors patched more easily than with training or fine tuning. This means a faster rate of iteration and faster progress.
And how do you handle a robotaxi network that operates all over the world, with various differences in each city? Do you have to figure out how to train one model that just works out of the box everywhere in the world? That sounds impossible. If you can take the same or similar models and customize them with a set of system prompts or memories for different robotaxi markets, that could be very powerful. When the robotaxi travels to a new market, it probably wouldn’t even need a new model — maybe just a slightly different prompt!
(Right turn on red is legal here. Cops pull people over a lot for speeding on 280, but not so much on 101. Avoid rose street when all the bars are packed on a football game night)
I don’t know. I don’t really know what the fuck i’m talking about but based on elon and ashok’s recent comments it’s clear they’re not resting on their laurels and have some major architectural changes for tesla self-driving in the works. I can’t wait to hear more about these new architectural changes, and to run them in my car
Leftists want AI data centers shut down over electricity.
But they never demand we unplug the factories, assembly lines, and machines that replaced human work with electric power.
Why?
Here are the top 7 voted for questions by investors so far for @Tesla's Q2 earnings call next week:
1) What's the current status of Optimus Gen 3 production ramp, initial deployment in factories, and external sales timeline/volume for 2027? What tasks can we expect the Optimus to perform by end of 2027?
2) What are the main constraints to expanding robotaxi operations faster, and how do you see that lining up with Cybercab production?
3) Tesla has missed short term guidance on robotaxi 3 earnings reports in a row, from 50% coverage of USA by end of 2025 to most recently 7 new cities in 1H26. What is keeping Tesla back from accomplishing these short term goals that they've set for themselves?
4) To reward long term Tesla retail shareholders for their loyalty, can you commit to achieving at least half of the goals outlined in your 2025 compensation plan before considering any offers to acquire or merge Tesla?
5) Why has growth of robotaxi vehicles stalled? When will we see cybercab start customer rides?
6) What are the latest milestones and city expansion plans for unsupervised Robotaxi/Cybercab operations in 2026? How many vehicles are targeted for the fleet by year-end, and what recurring revenue contribution is expected?
7) When can we expect unsupervised FSD with point-to-point summon capability?
Tesla's new six-seat SUV weighs 4,600 pounds and does 325 miles on a charge. A Kia EV9 is almost exactly as long, yet it carries a bigger battery, weighs 1,200 pounds more, and goes 45 miles less.
The gap comes down to wasted energy. At highway speed a big SUV spends most of its battery just shoving air out of the way, so the shape of the body decides how far it goes. Tesla stretched the Model Y by seven inches, added a third row, and raised the roof, then landed a drag coefficient of 0.216. That number measures how cleanly a shape cuts through air, and lower is better. The longer, taller Model Y L cuts through air more cleanly than the smaller Model Y it grew from. It comes in under the Mercedes EQS SUV at 0.26 and the BMW iX at 0.25, two SUVs shaped in a wind tunnel to be as slippery as possible. The boxier EV9 and EX90 are nowhere near as slippery.
Weight is the other half. The Volvo EX90 clears 6,000 pounds and the EV9 sits near 5,850, because a 100-kilowatt-hour battery and the frame around it add a lot of weight. Tesla pours molten aluminum into one enormous press and casts the whole rear floor as one solid piece, in place of about 70 smaller parts that would normally be welded together. Fewer parts, fewer welds, less metal to lug around. The Model Y L comes in at 4,600 pounds, over 1,000 lighter than its three-row rivals.
Then add the heat pump, which warms the cabin by moving heat around instead of running a battery-draining space heater, so range holds up in the cold. Put it together and the Model Y L pulls about 3.7 miles from every kilowatt-hour of battery, where the EV9 gets 2.8 and the EX90 gets 2.9. Being roughly 30 percent more efficient means you need 30 percent less battery to cover the same distance, and none of it costs performance, since the thing still hits 60 mph in 4.4 seconds.
Less battery is the payoff that keeps paying. It means less weight, a lower sticker price, less lithium and cobalt pulled from the ground, and a quicker charge, because there is simply less to fill. Efficiency is the rare number that makes every other number on the spec sheet better. Tesla built a six-seat hauler around using less, and it still goes farther than rivals with bigger batteries. The three-row EVs that do beat it on distance, the Rivian R1S and Lucid Gravity, get there with batteries 40 to 60 percent bigger.
Am converted. Just drove from south San Francisco through Oregon without touching my steering wheel. Tesla full self drive is both magical and life changing, relaxing and maybe even life saving! Thank you @elonmusk for the technology!
Robinhood has officially arrived in Canada. 🇨🇦
We’ve closed our acquisition of WonderFi, marking our entry into Canada through one of the most well-respected crypto platforms in the country.
With 1 million international funded customers, Robinhood’s mission is going global.