HW4 & HW5 Discussion

[Usain]

From x.com

Elon: HW4 software will lag HW3 by at least another six months, as our focus needs to be on getting FSD on HW3 working super well and provided internationally.

https://twitter.com/x/status/1695683422667317304

Just as I suspected.

Now that I've switched from HW3 to HW4, it sounds like it's going to be a year or so before I get FSD beta back. Sigh.

But the good news is that if things work out then $TSLA will take off. "HW3 working super well" is exactly what we need to move higher and higher.

 

[Unpilot]

Is it just me or is that thread full of nutters?

Yes.

Many would benefit if they read

The Bitter Lesson​

Rich Sutton​

In the post just above yours.

 

[Singuy]

The best V12 video out there. Shows you a lot of nuance of human behaviors that's better than V11 you may have missed.

 

[dhanson865]

As a bystander observing this I'm loving the reversal.

There were hoards of posts saying V4 or V5 hardware would be needed for FSD and for now V3 is not only currently better than V4 but Elon says they are going to get V3 up and going in multiple countries before focusing on V4 again.

The fact that they are prioritizing V3 over V4 runs counter to the argument that V4 is required to make FSD happen.

 

[aronth5]

As a bystander observing this I'm loving the reversal.

There were hoards of posts saying V4 or V5 hardware would be needed for FSD and for now V3 is not only currently better than V4 but Elon says they are going to get V3 up and going in multiple countries before focusing on V4 again.

The fact that they are prioritizing V3 over V4 runs counter to the argument that V4 is required to make FSD happen.

Now Tesla just needs to figure out how to provide FSD for those owners who just purchased a new HW4 car and no longer have FSD. Waiting 6 months isn't what owners expected so I'm hopeful Tesla will provide a solution sooner rather than later.

 

[Singuy]

As a bystander observing this I'm loving the reversal.

There were hoards of posts saying V4 or V5 hardware would be needed for FSD and for now V3 is not only currently better than V4 but Elon says they are going to get V3 up and going in multiple countries before focusing on V4 again.

The fact that they are prioritizing V3 over V4 runs counter to the argument that V4 is required to make FSD happen.

Extrapolating using what you know today but in a negative fashion with Tesla will most likely end with a wrong result. TslaQ is famous for doing this. Bet Green on twitter is trying to figure out a way to proof himself right by spinning the facts right now.

 

[Singuy]

Now Tesla just needs to figure out how to provide FSD for those owners who just purchased a new HW4 car and no longer have FSD. Waiting 6 months isn't what owners expected so I'm hopeful Tesla will provide a solution sooner rather than later.

If HW4 is delayed 6 months just to get FSD pushed out to international owners then so be it. Six or 8 months wait sucks, but those folks have been waiting for 5+ years.

*Yes, I just traded my HW3 car in for a HW4 car myself*

 

[Buckminster]

https://twitter.com/x/status/1695816363502387672

 

[insaneoctane]

Just a "cautionary" comment on all the v12 excitement (which I'm participating in):

Using Tesla's v12 demo to predict the approach of solving FSD, IMO, is similar to timing your purchases of TSLA in the market. I, for one, admit my timing and ability to read the tea leaves the last few years, has been pretty poor. I know some hear feel the same, while others (often option buyers) often feel they're pretty apt. The fact that Tesla has chosen this new v12 approach, also should highlight that v11 and prior was not the preferred method, probably yet another local maximum. Until we learned of v12 recently, most assumed FSD was imminent via v11 method....that's now, flawed as Tesla goes backwards and takes another fork in the road. So, will v12 need to backtrack after a new epiphany? Are we genuinely closer? Maybe....I'm just pointing out that while we can remain hopeful, and v12 makes a solution look closer than ever, v12 might just be another example that the previous route was inadequate.

Also, I can't lose sight of the fact that Tesla is spending billions learning how not to do FSD (while learning how to properly do it), but 2nd+ place folks are likely learning from Tesla's mistakes, spending pennies on the dollar. First mover advantage and all, but I do worry about how long Tesla's FSD lead actually will turn out to be, finally?

Anyway, there's my carebear speech for the day. I'm still as optimistic as ever, but I'm trying to apply some caution too.

 

[cliff harris]

There is no second place in FSD, if you need NNs and vast training data. The ONLY company on earth which a fleet numbered in the millions collecting training data is Tesla. It doesn't matter if Tesla go and give a ten hour powerpoint presentation on prime time TV explaining EXACTLY how the neural nets work, and giving the specifications of the cameras, and a handy guide to how to set up all of the training infrastructure...

its still absolutely useless without the final weights from the trained data. You basically have 2 options:

1. Sell at least 2 million cars at a likely loss, with all the hardware installed, and wait a few years hoping to scoop up an equivalent data set before you go bankrupt because everyone has just bought a Tesla.
2. License FSD.

Being second-place to robotaxis is not the same as second place to seat belts, or to LED headlights. Its not something that can be simply copied by looking at an existing car. Of course, 99% of traditional auto-execs have zero clue how this stuff works, so they probably think they can copy it. But they cannot. They are screwed. Oh and they don't have the in-house expertise for the hardware or software either.

I am very bearish on robotaxis, although bullish on FSD as being a substantial value-add for every Tesla. But the chances of anybody quickly following Tesla once its achieved is zero. This is one of the few things where I think gary Black is totally wrong.

 

[Singuy]

Just a "cautionary" comment on all the v12 excitement (which I'm participating in):

Using Tesla's v12 demo to predict the approach of solving FSD, IMO, is similar to timing your purchases of TSLA in the market. I, for one, admit my timing and ability to read the tea leaves the last few years, has been pretty poor. I know some hear feel the same, while others (often option buyers) often feel they're pretty apt. The fact that Tesla has chosen this new v12 approach, also should highlight that v11 and prior was not the preferred method, probably yet another local maximum. Until we learned of v12 recently, most assumed FSD was imminent via v11 method....that's now, flawed as Tesla goes backwards and takes another fork in the road. So, will v12 need to backtrack after a new epiphany? Are we genuinely closer? Maybe....I'm just pointing out that while we can remain hopeful, and v12 makes a solution look closer than ever, v12 might just be another example that the previous route was inadequate.

Also, I can't lose sight of the fact that Tesla is spending billions learning how not to do FSD (while learning how to properly do it), but 2nd+ place folks are likely learning from Tesla's mistakes, spending pennies on the dollar. First mover advantage and all, but I do worry about how long Tesla's FSD lead actually will turn out to be, finally?

Anyway, there's my carebear speech for the day. I'm still as optimistic as ever, but I'm trying to apply some caution too.

I hope this explanation helps.

Transformers were introduced in 2017. Prior to this was RNN. Transformers gave ML engineers the ability to expand their NN to a much bigger size, hence the initial development of Dojo. It gave them the ability to train in parallel vs sequentially with RNN, which also drastically reduce training time.

Tesla did a hard code approach, then converted RNN to Transformers based model, and was hitting local maxima until the development of diffusion which happened just a year ago. This gave access to a predictive model in which the AI can reverse engineer a garble mess by predicting what the mess should be. So in combination of diffusion and Transformers, end to end had a chance of becoming real.

So it's not a matter of Tesla not knowing what they were doing hitting local maxima, but they were doing the best they can with occupancy networks and BEV net until diffusion blew those out of the water. Note, diffusion was not developed by Tesla, none of these training models were. So they were grabbing what is avaliable and trying to make self driving work.

So it's not a "well why didn't they do end to end since the beginning of time? Seems obvious". However those training techniques were not available years ago.

 

[M|S|M MYP]

If HW4 is delayed 6 months just to get FSD pushed out to international owners then so be it. Six or 8 months wait sucks, but those folks have been waiting for 5+ years.

*Yes, I just traded my HW3 car in for a HW4 car myself*

If the market wakes up and realizes the true potential of FSD ($$$$) and $TSLA takes off to the moon, you can buy a HW3 equipped car for fun to use FSD now.

 

[Jeeps17]

Every day I use Google Translate or another; they are now better than any human aides we sued only two years go.

I can understand you wanting to use AI if the humans you employed were that bad ...

 

[Electroman]

Is it just me or is that thread full of nutters?

Par for the course for that topic. They have been at it for quite sometime now and will keep complaining without any reference to all the progress made so far. The usual suspects.

 

[Cosmacelf]

Elon: HW4 software will lag HW3 by at least another six months, as our focus needs to be on getting FSD on HW3 working super well and provided internationally.

Ugh. Isn't this kinda of an osborning in that it delays purchases until HW4 works with FSD? I know myself, I'm holding off upgrading until HW4 works with FSD. Maybe most people won't know/care.

 

[MP3Mike]

Ugh. Isn't this kinda of an osborning in that it delays purchases until HW4 works with FSD? I know myself, I'm holding off upgrading until HW4 works with FSD. Maybe most people won't know/care.

Which is really weird, as they are a handful of HW4 vehicles with people running FSDb 11.4.4. I think people misunderstood what Elon said. I expect most HW4 cars to get FSDb soon, but it will perform slightly worse than the same version running on HW3 for the next ~6 months.

Part of my reasoning for this is this post from Elon yesterday:

https://twitter.com/x/status/1695479308729598241

 

[Usain]

Now Tesla just needs to figure out how to provide FSD for those owners who just purchased a new HW4 car and no longer have FSD. Waiting 6 months isn't what owners expected so I'm hopeful Tesla will provide a solution sooner rather than later.

As one of those owners, I am hopeful as well. But I don't have very much hope. I think V12 on HW4 is at least a year off. Maybe two.

If Tesla FSD is close to "solved" on HW3, they will continue to concentrate on that. It sounds like they even want to support multiple countries before they worry about HW4.

I don't see much incentive to concentrate on HW4 until they start pumping out robotaxis in Mexico.

 

[Usain]

Just a "cautionary" comment on all the v12 excitement (which I'm participating in):

Using Tesla's v12 demo to predict the approach of solving FSD, IMO, is similar to timing your purchases of TSLA in the market. I, for one, admit my timing and ability to read the tea leaves the last few years, has been pretty poor. I know some hear feel the same, while others (often option buyers) often feel they're pretty apt. The fact that Tesla has chosen this new v12 approach, also should highlight that v11 and prior was not the preferred method, probably yet another local maximum. Until we learned of v12 recently, most assumed FSD was imminent via v11 method....that's now, flawed as Tesla goes backwards and takes another fork in the road. So, will v12 need to backtrack after a new epiphany? Are we genuinely closer? Maybe....I'm just pointing out that while we can remain hopeful, and v12 makes a solution look closer than ever, v12 might just be another example that the previous route was inadequate.

Also, I can't lose sight of the fact that Tesla is spending billions learning how not to do FSD (while learning how to properly do it), but 2nd+ place folks are likely learning from Tesla's mistakes, spending pennies on the dollar. First mover advantage and all, but I do worry about how long Tesla's FSD lead actually will turn out to be, finally?

Anyway, there's my carebear speech for the day. I'm still as optimistic as ever, but I'm trying to apply some caution too.

Tesla has always known that "end-to-end" was the end game.

See Karpathy's slide on how Software 2.0 is "eating" traditional code.

They have always known that the feast would continue. It looks to me like limiting factor was, and still is, raw computing power.

 

[Mardak]

As one of those owners, I am hopeful as well. But I don't have very much hope. I think V12 on HW4 is at least a year off. Maybe two.

FSD Beta 11.4.x has been running on customer HW4 vehicles since May, so hopefully that means new HW4 owners will be able to at least use FSD Beta before V12 -- potentially as soon as next week. However, Musk and Elluswamy seem to be bundling Actually Smart Summon as part of V12, so that could mean many other features of even $6k Enhanced Autopilot will need to stay as deferred revenue until it's compatible with HW4?

 

[Gigapress]

Bear in mind that Tesla has 10% of the global stationary storage market. Two other players have the same scale as Tesla give-or-take a per cent or so. Seven others are within spitting distance. A further ten have their names on the board. Some of those are also cell suppliers into Tesla so get to win either way.

For now, but that will be changing quickly. Tesla delivered just 3.7 GWh of storage in Q2, but Lathrop alone has nominal capacity for 10 GWh/qtr. And if the IRA lasts until sunsetting, almost all the big profits from it will come in 2027-2032. Per the scale-up plan, Tesla will need to build the equivalent of at least 10 more Lathrops in the US by then. Does anyone else have public plans for TWh scale by 2030?