While I can’t verify the claims on training. But I can run their R1 reasoning model locally with just average AMD GPU. That is a huge break though compared to the o1 model of OpenAI where there is no way to run it locally on a home PC. R1 reasoning model exceeds o1 on most AI benchmarks.

 

As 703 mentioned above, the whole problem for OpenAI and Nvidia is being able to run models on much lower powered processors. Development costs were around $5M...yes, only five million while companies like OpenAI, Microsoft and others collectively spent probably over a trillion. Which is why Nvidia took another 5 percent hit today

 

Yep, very true (with few exceptions of course). I've always liked cars, never really loved them so I never felt passionate about certain things like I noticed out of others who truly LOVED cars. Here on the forum I am constantly reminded of how little I actually know about many aspects of cars. I know cars visually/cosmetically, probably 9 times out of 10 I can identify a car make and model at night just by their headlights or tail lights (one of very few things that still impresses my kids). But that's about it. As I've gotten older I've definitely started to care significantly less about almost everything other than travel and experiences.

 

Some of you guys are way more technical than me from what I can tell. But... I'm pretty sure this Deepseek all-in development was still in/near/or even above the $100m mark. I *believe* the $5m figure was just on the latest round of development/enhancement and was spun as only costing $5m total. What implication (if any) does this type of work and new technology approach have on things like FSD? I know it fundamentally applies to LLM's, but can some of the approach to the "potentially better mouse trap" get applied to something like FSD and have it "learn" and improve faster... and in a less expensive way for a company like TSLA?

 

I'm not on the development side of IT, so I can only deduce from what I read and hear from other experts. Maybe they are lying about the costs, but even if they spent $100M, that's still way, way below what other companies have spent, which is in the Billions

 

unless I'm missing something there's a big difference between USING the LLM, and BUILDING the LLM. LLM's can be used in 'limited' memory, cpu, etc. (phones are doing this.) i don't know much about deepseek but maybe it reduced a massive LLM into a small one so it could run in limited resources and achieve 'similar results' to something using a huge LLM - e.g., chatgpt which can't run on a phone stand alone - the app does the round trip to openai's servers. a nice accomplishment but I doubt it's going to eliminate the need for giant compute. back to Tesla, they can run FSD 'stand alone' in the car's resources but the model it's using requires vast compute capability which is not likely changing any time soon.

in neuroscience we're still learning about how the brain does what it does, but it's a pretty amazing 7lb computer, but it does 'guess' and approximate on a lot of things and clearly gets a lot of things wrong, lol. ai will evolve too.

as for the giant ai budgets... hey vc companies are showering them in it, so why not.

 

Per the FT well over $500m was spent on NVIDIA GPUs by DeepSeek. Actual costs are way higher than reported.

 

I think Nvidia's compute power will still be needed regardless, for example, how does Deepseek handle and render video? Plus, I have a stake in Nvida, so there's that

 

That's probably the biggest reason why it dropped

 

 

They didn't reduce the size of the LLM, it has 681 billion total parameters, but when its inferencing, it only needs to activate 37 billion, which makes it computationally more efficient and what we don't know is how they are able to get to the same outcome surpassing bigger models like OpenAI o1 in benchmarks. Meta's (Facebook)'s llama open source model uses 405b parameters when inferencing and produces vastly inferior results - thus computationally more intensive and worse off. Rumors is that even Meta's next generation model doesn't perform as well - hence the war room happening right now within Meta. Sam Altman even say he is impressed and will speed up release of OpenAI's next model (o3).

Anyway, the key take away is that Deepseek has broken the business model of LLM's from all the major AI players in the US, as businesses now being able to run it locally with minimal hardware footprint without being locked into using (expensive) OpenAI, or OpenAI hosted in Azure etc. Most enterprises don't train models, they just use them and ground them on understanding their own data (often called Retrival Augmented Generation).

All the noise about Chinese censorship on political answer doesn't apply to the open source model once you download it. The censoring happens after the LLM produces the results on their public website - hence you can see it in real-time how they have built some kind of sematic filter that erases answers from the chatbot as the chatbot produces the political answers (quite amusing/funny). But at the same time, the LLM is not woke like the US ones, it can openly swear, have racial / gender / appearance debate or jokes.

 

Ford's EV division lost $1.4 billion in Q4 2024 must be all those posts on X and funny moustache man salutes.

 

Wow, they've had EV'S since 2004?? 🤣

 

yes

 

Pond scum