I asked a specific question and Rogue started explaining to like an idiot. I am setting his expectations he knows how to respond to my question.
There is certainly a way to go about asking your question that doesn't imply you are a pompous ivory tower academic. If it is that important for others to have the context of your Ph.D, you can add it as a footnote to your profile, maybe include the topic of your dissertation to boot.
For most members on the forum, that answer was hopefully understandable enough given the likely lack of background in the topic. I am not treating you like an idiot but rather the general response for the general forum member. Having a Ph.D doesn't give excuse to engage in a d*ck swinging contest with random people who don't know your background that you haven't shared.
Given your background, I would expect you to have been exposed at least to some degree on the performance differences of models that specialize vs generalized. BERT models are an excellent example (a type of NLP model aka Natural Language Processing models for others on the forum). We can see from open source models that multi-language models have decent accuracy for most languages covered, but the overall performance is diminished as a result. In comparison, models trained for a specific language almost always score higher than their generalized counterparts. Counterpoints are models that group languages by family (ex. Romance languages). Those share enough similarities and overlap that performance penalties from the generalization are negligible or even non-existent. However, you take languages such as English, Arabic (any of the many, many dialects), and Chinese (any of the 12 dialects) with vastly different linguistical rules and alphabets (using the term loosely), the performance flounders drastically given the drastic differences.
There are other examples we can pull from such as Generative AI, specifically image generation. Those models often get released for generalized use cases that struggle in accuracy and consistency when pushed towards distinct styles via prompts. To counter this, people partially refrain those models on the target specialization (hyper-realism, anime, etc) and produce something called LoRA(s), Low Rank Adaptation. Again, the same pattern emerges. Increased performance in the specialization over the generalized solution for the targeted use case.
In essence, the truism jack of all trades and master of none is applicable to models. Pick one. Be generally good or okay at a number of things, or be exceptional at one thing at the cost of the other things.
In Lucid's case, there are a few areas to specialize. Signal processing vs image processing vs video processing would benefit greatly from different models given the unique characteristics of the electromagnetic spectrum at different frequencies, types of information available at said frequencies, and differences in the data as a whole. Another place for specialization is the decision making portion of Dream Drive Pro. Rather than attempt to cover every feasible driving condition with a singular model (highway, city, rural, mountain roads, flat plains of Idaho, etc.), models specialized in differing scenarios could be developed to increase performance in those situations.
Performance in this regard is accuracy, reliability, and reasonable execution to meet safety standards of autonomous driving. The bar is far higher such that deployment and development are dictated by the worst performance rather than the average (or median or best or amortized and so on) due to regulation.
If Lucid takes the approach of a generalized solution, they are attempting to brute force in many ways the worst performance upwards. Generalized meaning to clump all decision making into a singular model or all sensor types into a singular model. The requirements of such generalized solutions in practice often outstrip their specialized counterparts by orders of magnitude even in aggregate. Related to Nvidia, I suspect Nvidia's platform supplies the building blocks and possibly vast swaths of data to utilize to build the actual functionality rather than the other way around.
www.shm-afeela.com
