[AMG Power]

BYD has announced it will offer self-driving systems in nearly all of its models, even the lowest-priced ones.

By integrating DeepSeek into its autonomous driving technology, the EV maker can provide a far cheaper alternative to systems like Tesla’s Autopilot.

BYD's shares hit a record high in Hong Kong, increasing by 4%.

Following this announcement, Japanese car maker Nissan also announced it will be integrating DeepSeek into its EV’s.

Source

[Turbanator]

One more reason to keep away from such companies. Not that I am worried about privacy or anything but I don’t think most countries will like cars to have such stuff sending back all data to the masters in China.

[AMG Power]

True.

The issue is that other manufacturers like Nissan are also integrating DeepSeek into its EV’s in order to stay competitive. The newly launched N7 which eerily resembles the Verna will be Nissan’s first to offer DeepSeek.

On another note many automotive brands including Zeekr, Dongfeng M-Hero, Smart, Leapmotor, IM Motors, and Arcfox have also announced the incorporation of DeepSeek's AI models in their cockpits.

Nissan’s new N7:

[Sentau]

How is a LLM helping with autonomous driving¿? I am not a software engineer and hence I am genuinely confused.

[SunaPana]

What is Deepseek going to do. LLM's can't drive cars those are done by AI designed for driving cars.

This sounds like they are just jumping on the hype and publicity from Deepseek

[dileepcm]

If you are connected to internet via your phone, your data is not secure for the most part anymore. There are all sorts of ways to track and trace what you do and you have been getting targeted ads already for years.

So what BYD is aiming to do is nothing new in grand scheme of things.
Those who are tech savvy enough will find a way to minimize their digital presence as much as possible and that is going to happen with BYD as well.

Now its not just BYD that has a connected car sending data to elsewhere, Tesla and other electric car makers have live customer/driver data in their systems for years now. If you simply use google maps for navigation, your live data is sent across to elsewhere - its as simple as that.

- So there is nothing new here. What we need is the awareness of how these systems work and ways to ensure our security and safety by asking right questions and by making smart choices. At least Deep seek has opened their source code to the public so that techies out there can really read and understand how it works.

On the other hand, if you have not realized yet, Open AI has 'open' only in their name, their code is not public. This means, Deek seek offers higher level of transparency on how it works and handles data in comparison.

Originally Posted by Sentau

How is a LLM helping with autonomous driving?

Originally Posted by SunaPana

What is Deepseek going to do.

In current times, there are are various types of applications using machine learning algorithms to learn from large amount of data so that it can make accurate predictions and these (self driving, weather prediction etc) are some of those implementation flavors.
My company (who supplies supply chain solutions) has machine learning algorithms that can predict order delivery timelines by learning from millions of updates that are supplied by various carrier services on delivering products to customers.

The more amount of data the algorithms has for processing, the better it can learn and react to variety of situations. So collecting most recent data and making it available to the 'brain' is important to keep up the accuracy.

[Sentau]

Originally Posted by dileepcm

If you are connected to internet via your phone, your data is not secure for the most part anymore. There are all sorts of ways to track and trace what you do and you have been getting targeted ads already for years.

Yeah it is somewhat disgusting how everything in our lives are collecting more and more of our data to show us targeted ads, eroding our privacy in the process.

Originally Posted by dileepcm

- So there is nothing new here. What we need is the awareness of how these systems work and ways to ensure our security and safety by asking right questions and by making smart choices. At least Deep seek has opened their source code to the public so that techies out there can really read and understand how it works.

I thought only the weights were open and not the algorithms or training data?

Originally Posted by dileepcm

In current times, there are are various types of applications using machine learning algorithms to learn from large amount of data so that it can make accurate predictions and these (self driving, weather prediction etc) are some of those implementation flavors.
My company (who supplies supply chain solutions) has machine learning algorithms that can predict order delivery timelines by learning from millions of updates that are supplied by various carrier services on delivering products to customers.

And I am assuming you use specialized models/machine learning algorithms trained on specific relevant datasets. Shouldn't it be the same for assisted/autonomous driving where a specialized model is used? When i was studying astronomy, my fellow researchers from the computational side used to use specialized models which I don't think used similar algorithms to LLMs.

[dileepcm]

Originally Posted by Sentau

I thought only the weights were open and not the algorithms or training data

There is in fact a serious debate on the definition 'open source' when it comes to AI models (Reddit being the main platform). Let me rephrase my statement - Deepkseek is more open than any other similar products. This gives a higher degree of transparency and trust for those who would want to list down pros and cons of such models in order to make an informed decision.

Originally Posted by Sentau

And I am assuming you use specialized models/machine learning algorithms trained on specific relevant datasets. Shouldn't it be the same for assisted/autonomous driving where a specialized model is used?

This is an interesting point. Let me explain what I see happening in any industry.
In a lot of scenarios, a lot of companies directly or indirectly capture lot more data than what is absolutely needed for the given contextual purpose (Gender info is a good example). Why is this the case? Its simple - any data is valuable and there are always more creative opportunities for that data to turn into information that could create business opportunities. In short, the more data and the more creative you are, the better the overall outcome is.

In my company's case - the data (metrics around pick up point, details of goods being transported, drop point etc) is provided by carriers. Not all attribute of the data are really relevant for training the model for predicting delivery lead times which is the main purpose (which is what customer sees on websites while browsing products or while adding something to cart). However the data is still stored either for a different (side) purpose or for future possibilities. This aspect is simply the case - any data is good data, the more you have it, the wider the opportunities are.

[greyhound82]

Originally Posted by Sentau

How is a LLM helping with autonomous driving¿? I am not a software engineer and hence I am genuinely confused.

It's not specifically LLM , they have a multimodal architecture for sensor fusion tasks . Let us say a VLM( vision language model), which help in embedding the vision and language in a feature space can be used for video summarisation and understanding. Similarly you can embedded a time series and 3D- point cloud data or Ultrasound and Visible spectrum data
For example we use multimodal to understand human activity on factory floors for flagging anomalous activities and SOP's using Sensor fusion.

[electric_eel]

Originally Posted by Sentau

How is a LLM helping with autonomous driving¿? I am not a software engineer and hence I am genuinely confused.

Anything and everything can finally be converted to a Language model. But then you get silly things like this (wait till the end of the video to see what chatgpt can do to chess against stockfish the AI engine designed for playing chess)

https://www.youtube.com/watch?v=6_ZuO1fHefo

The problem with LLMs is that they "hallucinate". They are like the "mystics" who make a lot of absurd statements which needs to be sieved to get "gems of wisdom".

[greyhound82]

Originally Posted by electric_eel

Anything and everything can finally be converted to a Language model

Transformers which is a differential compute module, are data/modality agnostic as you can model any data as input tokens and self attention can easily work on these. we need some kind of preprocessing for positional encodings for different data.

We are not converting the videos/ images or say any other modality say x-ray, ultrasound, vibration to language rather they can be embedded in a shared multidimensional feature space which then used for down stream tasks.

In ADAS, we make use of LiDARs, Visible, Radiowaves etc and transformer based LLM architecture is a good fit for sensor fusion and capturing spatio- temporal informations other than using them for language queries

[Simha.k]

I've said it before and I'll say it again. Any data that enters or leaves China/Chinese servers is subject to severe scrutiny by the CCP.

If you're ok with the commies handling your data, kiss your privacy goodbye.

[electric_eel]

Originally Posted by greyhound82

Transformers which is a differential compute module, are data/modality agnostic as you can model any data as input tokens and self attention can easily work on these. we need some kind of preprocessing for positional encodings for different data.

We do agree here. Any task can be seen as a language generation task. For example the "commands" you give to the car can be seen as a language using suitable encoding. and hence all tasks are "language recognising/generation task". LLMs can be use in such context.

In ADAS, we make use of LiDARs, Visible, Radiowaves etc and transformer based LLM architecture is a good fit for sensor fusion and capturing spatio- temporal informations other than using them for language queries

This is where I have to disagree. We do know LLMs have a tendency to hallucinate and they can only be used if there is a "verification" process to their output. The chess example is particularly revealing where the first part of the game it played quite competently till chatgpt goes completely bonkers.

The promising idea is to use LLMs together with an (automated machine checked) verification process. See things like https://reservoir.lean-lang.org/@lean-dojo/LeanCopilot where LLMs are used to search for purported proofs which is okey because they undergo checking by the lean system. See also https://mathstodon.xyz/@tao/111206761117553482.

Without such a verification step I would not recommend its use for mission critical stuff like ADAS.

[Chrome6Boy]

Originally Posted by Turbanator

One more reason to keep away from such companies. Not that I am worried about privacy or anything but I don’t think most countries will like cars to have such stuff sending back all data to the masters in China.

Originally Posted by Simha.k

I've said it before and I'll say it again. Any data that enters or leaves China/Chinese servers is subject to severe scrutiny by the CCP.

If you're ok with the commies handling your data, kiss your privacy goodbye.

There is a bias against China, but not against Western countries. I don't know if I really trust one over the other.

Many of us have been using Google Maps / Facebook / ChatGPT / Google Photos for years now, and all our data has been sent to the masters in the US. For me, it doesn't matter which master scrutinizes my data. If I like the product I use it.

Besides, the Indian government is already restricting BYD's usage of sim cards in their cars. I'll leave it to our government to decide if it is in the best interest of our country or not. I do not want to restrict myself from using new technology because I am scared of losing privacy.