How Tesla Can SOLVE FSD!

Join this channel to get access to perks:

Tesla's FINAL Piece of FSD Puzzle!!

Get The Elon Musk Mission (I've got two chapters in it) here:

For a limited time, use the code "Knows2021" to get 20% off your entire order!

**You can help support this channel with one click! We have an Amazon Affiliate link in several countries. If you click the link for your country, anything you buy from Amazon in the next several hours gives us a small commission, and costs you nothing. Thank you!

**What do we use to shoot our videos?

**Here are a few products we've found really fun and/or useful:

Tesla Stock: TSLA

**EVANNEX
If you use my discount code, KnowsEVs, you get $10 off any order over $100!

Instagram: @drknowitallknows

Sources:

[stanosu]

You are so right. I am not technical but as a new Tesla driver I notice that construction zones almost always require intervention. FSD waits until the very last minute to try to merge from two lanes to a single lane for through traffic. In fact, any time a merge is required ahead, FSD fails. FSD v 12.3.6. I hope future versions address this. BTW, I just found your channel. Thanks for communicating this technology in a way non technical folks can understand it!

[grahamfowke]

Subscribed. There’s several similarly named channels, looks like it is the one from DrKnowsAI (which may be a more accurate search term to find the channel).

[STEVEF]

I have noticed that V12 does a better job of mid term planning. With a stretch of 4 or 5 traffic lights, V11 would stay in the right lane and wait to pass the last traffic light before trying to get into the left lane for an upcoming left turn lane. I would often have to take over with heavier traffic to insure being in the left lane early enough. V12 now moves into the left lane 4 traffic lights ahead. The one downside that I believe you referenced awhile back is the ability to have local driving knowledge. In this case there are many cars that are in the left lane to try and turn left at the second upcoming light, so staying in the right lane slightly longer would be preferred.

[STxFisherman]

Subscribed. A.I. is a great topic to put on its own platform. Good luck.

[nickmcconnell]

Yippee for new channel. Congrats!
Smart to float off the AI stuff to another channel.

[erwile]

I'm not sure about your point, they surely have an LMM, it's a video LLM. Middle term planning could be solved IMO by tokenizing at multiple scale. 24 token/s (or window (I will say token for the example)), 1tk/s, 1 token/10s and then 1 token/min etc. So the second scale would be a compressed version of the 24 fast token, then one every 10s you can compress your 10 token/s etc. This way you can easily plan ahead with a bigger view

[Digital-Dan]

12.3 has a tendency, when moving into left turn lanes, to start the drift, drift back a bit, and then complete the drift. This is starting to feel like standard behavior. It is less disconcerting once you get used to it, but if is mimicking humans, it is mimicking hesitant and unsure ones.
But this one might be related to your issue: I have a route that involves a left turn into a two lane road followed immediately by a right jog onto a frontage road. The planner takes this turn in two steps, first turning briefly into the left hand (fast) lane and then drifting over. If there is conflicting traffic, the result is a missed opportunity to make the jog and the need for very complex and time-consuming detours.

[IntoTheFray.]

I think Tesla should rethink using outside vendors for navigation maps. I believe they already add some navigation data about potholes and congestion to the data they get from Google. I think they would do so much better if they integrated the nav system with their own navigation maps. this might solve the speed limit problem, the lane changing issues, and allow better real time updates about traffic, construction, and road conditions. If they are committed to using outside data, they should feed it into their own navigation mapping and perhaps integrate into the full FSD package. I don't know the architecture, do they need separate maps for navigation? If so, they should make their own.

[nickmcconnell]

The lane dancing you demonstrated may have been for good reasons of maximum distance between you and a car in another lane. When you review the moves with this in mind they make total sense.

[williamwoo]

Dr. Know it all----> I am having long term covid of the lungs so what has helped is tumeric, ginger, garlic and lots of green veggies, Food is the trick plus your cool plant. I also do 15 mins cardio in the morning and 15 mins cardio at night. Then I practice my martial arts for one full hour. Yes you going to sweat. Then at night I do the sitting meditation by arms out to your side and then I relax with tension. If you aren't relax you will cause it's impossible to hold that position for one hour. Qigong is what I do. I am 74 and most people think I am 50 years old. It's the way i walk and move with speed that most 74 years old can't do. Good Luck. Find a Martial Arts Teacher in Tai Chi. I have a weekly class with my 4 Wing Chun instructor all with 34 years knowledge. I been doing Tai Chi all my life it seems So much more fun learning to fight. Yes it can get dangerous. but that is what makes it so fun.

[yacahumax]

One of the thing fsd does is change lane to the right when the next exit is to the left. Also I noticed the lane changes when there is no need. It seems to prefer the middle lane.

[jimcallahan]

Maybe you have already done this, but might be interesting to have video on how multi-media models are trained.
In other words how are data sets of words and pictures combined in one model?

[techyjames]

Tesla will never have L3 FSD until they solve correctly the speed limit recognition and stop resetting speed limit settings every time they pass a side street. This is both annoying and in some cases very dangerous. At least with v3 hardware it consistently fails to read speed signs, and when it does first side street it resets speed to some assumed speed that is either excessively high like 55 in a 30 or excessively low like 40 in a 55 zone. No one would drive into a business street going over 55 when the posted speed limit is 30. This is a huge safety issue requiring immediate over ride if there isn’t a lead car slowing it down.
Everyone that Tesla will solve FSD by end of year is delusional and for them I have a bridge in London I will sell them for a low price of 100, 000 US dollars.

[jimjsmith]

Would the LMM also be a great place to communicate with people in the car. eg. An accident blocks your route and the bypass will take hours, do you want to continue or go elsewhere? Or maybe the ability to say 'don't take the freeway this time.? Maybe it is a good place for some local memory, like don't slow down for traffic at this corner because there isn't any ever.

[tomcraver]

I wonder if Tesla would do better having each car learning each driver's local area specifically from that driver's interventions.
For most good drivers, this would 'patch' the few remaining failures in people's most common driving, probably giving them another order of magnitude improvement at least.
It has risk of a bad driver training their car to drive badly - but if Tesla's FSD always projects how it would drive WITHOUT that driver's training, they could shield themselves from liability by making the 'patching' a user-enabled feature, and then they could show that in court - i.e. "See, this is how TESLA would have driven it, versus this is how this bad driver's bad driving trained it because he/she chose to enable this feature."

[whodatcatt]

All those lane changes wow it’s obviously taking after the way normal people drive
Especially kids

[Greenspaceservices]

I was driving down the TransCanada highway the other day with no traffic insight around me and the car changed lanes for a while and then changed back to the original lane. I drive this highway 4 days a week and only experienced this one time 🤔

[jamesfirnhaber]

Good video. I think it might require two models running in parallel- safety, basic driving model, and executive planning model which plans ahead based upon the route selected, traffic density, time-of-day, driver patterns and preferences, construction, weather, etc. Then, the output of the models are merged into a decision model. When the executive planning model output differs from the driving model, the decision model asks this binary question: safe or not safe? If safe, the executive decision overrides basic driving decision. For example- a lane change. The basic driving model wants to change lanes for no reason. The executive model wants to stay in the same lane because there is no traffic, no impending turn, and it's more comfortable for passenger(s). The decision model asks if there is any safety reason to change lanes. If the answer is "no", the decision model votes in favor of the executive model. This architecture might require additional in-car processing power than is currently available, but I think might be the only way to get FSD to human level good in the categories of safety, planning, and comfort. Perhaps I am wrong and the basic FSD model can be trained with the planning component, as well as all the basic driving stuff. That would require training to incorporate all of the route information for each session, and then be "graded" on the planning and comfort aspects as well as the "safety" aspect.

[jacksonmatysik]

Nice, your AI content is my favourite

[plop]

Look at how the human brain does it: There are different centers of the brain that take care of different temporal areas. Low level (like breathing) Automatic (take a cup and drink contents), Survival discission (that is a lion - run!), Planning of future actions (If I want to go to A I first need to go to B), Abstract thoughts (that don't really have any temporal dimension). All of these thoughts are determined in different places in the brain, and the brain somehow decides which is the proper decision in the given situation. Something like that also needs to be happening in AI.