And this is why we can’t rely on self driving (eg FSD) features yet

[DuoRivians]

I’m convinced that LLM based “autopilot” systems (which is what fsd is) will never get beyond L3. The driver will always need to be there.

Whether or not it’s easier to just drive yourself (gosh!) versus your supervising a computer to drive is debatable.

For me, I find it easier to just drive myself in most cases. And yes, I have fsd and use it on my MYP.

Sponsored

 

[zefram47]

If I understand your comment correctly, you expected the truck to take the right exit from existing road, with Nav enabled.

Rivian’s D+ doesn’t take turns. It only maintains a lane and speed. It’s a pretty basic L2 driver assist.

If you have Nav set to a destination which requires you to make a deliberate turn off your existing road, it will not do so. Also, I do not believe it will warn you about the situation, so you’ll just blow past the exit and need to double back.

Again, I was established in a lane and it didn't follow that lane. It tried to stay on I-70 despite being in a lane that *only* continued off to the right. I expected it to maintain its lane and it didn't do that.

 

[Tr4ckD4ys]

I’m convinced that LLM based “autopilot” systems (which is what fsd is) will never get beyond L3. The driver will always need to be there.

Whether or not it’s easier to just drive yourself (gosh!) versus your supervising a computer to drive is debatable.

For me, I find it easier to just drive myself in most cases. And yes, I have fsd and use it on my MYP.

Why are you calling it LLM based? You aren’t referring to “Large Language Model”, are you?

I don’t think that is correct. As far as I can tell they may use some combination of architectures including transformer-based ones including its attention mechanism but they are not using LLMs, which would be specifically for text based tokens.
As a “stupid” comparison, Tesla FSD is more based on Midjourney like architectures than on ChatGPT-like ones.

Ok I’m off my soap box.

 

[DuoRivians]

Why are you calling it LLM based? You aren’t referring to “Large Language Model”, are you?

I don’t think that is correct. As far as I can tell they may use some combination of architectures including transformer-based ones including its attention mechanism but they are not using LLMs, which would be specifically for text based tokens.
As a “stupid” comparison, Tesla FSD is more based on Midjourney like architectures than on ChatGPT-like ones.

Ok I’m off my soap box.

It is basically LLM. Image, video decomposed into how ever millions (billions?) of parameters, eg pixel colors, transitions, etc.

Instead of word parameters, they’re video input parameters. LLM doesn’t care, it’s just an algorithm.

“Large language models (LLM) are very large deep learning models that are pre-trained on vast amounts of data. The underlying transformer is a set of neural networks that consist of an encoder and a decoder with self-attention capabilities. The encoder and decoder extract meanings from a sequence of text and understand the relationships between words and phrases in it.”

 

[Tr4ckD4ys]

It is basically LLM. Image, video decomposed into how ever millions (billions?) of parameters, eg pixel colors, transitions, etc.

Instead of word parameters, they’re video input parameters. LLM doesn’t care, it’s just an algorithm.

“Large language models (LLM) are very large deep learning models that are pre-trained on vast amounts of data. The underlying transformer is a set of neural networks that consist of an encoder and a decoder with self-attention capabilities. The encoder and decoder extract meanings from a sequence of text and understand the relationships between words and phrases in it.”

It is not basically an LLM. I would not equate LLMs with Transformers for image recognition. The two architectures and how they are trained, ie the decoding/encoding process is similar yet different between the two.

The above idea that anything that takes an input and generates an output is an LLM, as long as it is a deep learning network with self attention, is flawed. Of course they are all closely related to the “Attention is all you need” paper, but that doesn’t make Dall-E, Midjourney and ChatGPT the same. In fact, even algorithmically and from a training process perspective there are many differences also simply due to the way an image decodes and a text decodes - ie the number of tokens you end up with.

Anecdotally, it would be called Large Everything Model in that case.

Semantics at this point I guess.

 

[mikehmb]

Again, I was established in a lane and it didn't follow that lane. It tried to stay on I-70 despite being in a lane that *only* continued off to the right. I expected it to maintain its lane and it didn't do that.

Aha - definitely a D+ fail then. I’m surprised it didn’t kick itself off. I’ve never been able to keep it active in that circumstance.

 

[DuoRivians]

It is not basically an LLM. I would not equate LLMs with Transformers for image recognition. The two architectures and how they are trained, ie the decoding/encoding process is similar yet different between the two.

The above idea that anything that takes an input and generates an output is an LLM, as long as it is a deep learning network with self attention, is flawed. Of course they are all closely related to the “Attention is all you need” paper, but that doesn’t make Dall-E, Midjourney and ChatGPT the same. In fact, even algorithmically and from a training process perspective there are many differences also simply due to the way an image decodes and a text decodes - ie the number of tokens you end up with.

Anecdotally, it would be called Large Everything Model in that case.

Semantics at this point I guess.

Tokens are just factors, data, information. Abstractly, they can be taken from any mapping, analysis of data, whether it’s text, image, or video.

Yes, there might be different optimization techniques and stylizations depending on the context of the model, but LLMs at an abstract level are similar.

I guess we agree to disagree on this.

 

[Tr4ckD4ys]

Tokens are just factors, data, information. Abstractly, they can be taken from any mapping, analysis of data, whether it’s text, image, or video.

Yes, there might be different optimization techniques and stylizations depending on the context of the model, but LLMs at an abstract level are similar.

I guess we agree to disagree on this.

I think we largely agree to agree. I don’t disagree with the context you are providing. I disagree with the generalization of defining everything happening right now as LLM (in this case defining Teslas FSD as LLM). I just don’t think it’s good for the AI industry and everybody working in it (including myself) to throw this all under the same definition of LLMs. We already see today what happens when research becomes too focused on one problem (ie LLMs) because it’s the big star right now. It is limiting us to get materially closer to AGI because we miss sufficient depth of discoveries in other areas of AI. My only worry with your statement was that: The public generalization and the hype causing everybody to call everything LLM will be limiting to our progress one day, and many in the industry would agree.

 

[HaveBlue]

The one time I used fsd or whatever on the freeway it felt worse than my 15 year old learning to drive. Way too much attention needed vs just driving myself. I resolved I'd never buy a Tesla and pay for that feature. Acc, fine and the rivian comes with it's system. I'll reserve comment until I use it but I wouldn't have paid extra for it.

 

[Robin]

the combination of beta software, inferior hardware and the typical Tesla bro leads to an obvious outcome.

?

 

[Chewy734]

Now, imagine FSD on a Cybertruck while you’re towing infinite mass.

 

[DuoRivians]

Now, imagine FSD on a Cybertruck while you’re towing infinite mass.

 

[SSteveEV]

Oh damn, that was a sketchier situation than I expected.
No hesitation, no surrender!!

 

[blturner]

I agree that FSD is not ready for prime time. Elon said that there were fewer accidents per mile with it on than off. I believe that was a misrepresentation because no one uses FSD for the hard stuff.
At some point it will be true that a future FSD system is better than the average driver statistically. But that only means that instead of 10 people dying from human drivers, that 9 would die from FSD. I don't think we as a society are ready for the moral and legal implications of that. I expect that we will let human drivers kill thousands before we accept even a smaller number of robot caused deaths.

 

[Dark-Fx]

I agree that FSD is not ready for prime time. Elon said that there were fewer accidents per mile with it on than off. I believe that was a misrepresentation because no one uses FSD for the hard stuff..

Tesla also doesn't take responsibility for accidents that occur after handing off to the driver, even if the system handed off in a no-win situation it was responsible for.

Sponsored