The promise of Foundation models and where we are now
Unlike traditional models that are designed for specific tasks, Foundation models are vast, pre-trained models that serve as a base for many applications across different domains. Stanford’s Center for Research on FMs, define it as “Train one model on a huge amount of data and adapt it to many applications.”. Many foundation models are also fine tuned models on top of larger pre-trained base models. Links below use different terms to point to such models FMs or Large Foundational Models (LFMs), Frontier Models, or Large X Model (LXM) etc.
The rise of Foundation Models came with a lot promises. This post is a brief survey to find where we are with those promises.
So first, what were those promises? These models will:
a. apply to wide ranging real-world applications within and across many domains,
b. be easy to adapt and extend to different modalities,
c. scale efficiently,
d. reduce the need for high quality, labeled data and
e. generally enable rapid development in AI applications.
There are many other factors that are leading the industry further and further towards FMs:
·????? computation challenge (exponentially more to build many speciality models),
·????? applicability & usage (surge in the number of use cases that demands the need for ML based augmentation),
·????? pace of adoption (pace of adoption across the industries),
·????? cost effectiveness (economics in serving multiple applications),
·????? business opportunities and competition (e.g. data and compute accessibility has become a moat for big tech companies).
Of course, this is in addition to the architectures that became available.
I was curious to know what’s happening with this promise. So, I decided to look up and make a list.
领英推荐
As I went through this exercise, I was amazed by what’s being put out there. I don’t know how good they are in real world applications but it appears like we are in a very exciting journey.
However, extending Foundation models to other modalities and applications across various domains presents several challenges - governance and regulation to ensure responsible development and deployment, potential misuse such as misleading content, managing biases, applications with unintended impact. The important part is that these models, especially OSS, cannot be stopped from proliferation making them a difficult challenge to manage.
Conclusion:
1.??? Foundation models are rapidly growing, and new sub-modalities are emerging quite rapidly. The original promise remains alive.
2.??? Combining these FMs with intelligent routers, like with MoEs, creates powerful systems by dynamically selecting the most relevant models for specific tasks. E.g. projects like MIT’s HiP, highlights their potential to revolutionize complex planning and decision-making processes in robotics and beyond.
3.??? Interesting new modalities, such as emotions, haptic feedback, olfactory data, and biosignal processing, opens up exciting possibilities their applications and impact across diverse fields.
4. The path to AGI/ASI, whenever we see it, is likely paved through the advances and integration of these foundation models.
In any case, a very exciting space to watch out for.
Useful References:
Director Technical Architecture
6 个月Good share and point of view Rajesh Chitharanjan Excellent consolidation of modalities and relevant models. Thanks for sharing resources as well.