I really and truly hope so. I have a lot to learn about the continent, and I love it much. I've done some work on significantly reducing the experiment cost for small experiments on GPT and convnet models (WikiText and CIFAR10, respectively), and one of the big things that I thought of was how it could help Africans in AI specifically because it allows for good research test signals but isn't prohibitively expensive/slow.
On A100s this particular CIFAR10 network trains in under 7 seconds for example and the network (fp16) is <10 MB IIRC, the relative gains over other previous training speeds is even better on older cards IIRC due to it being pretty memory limited at this point. It could be used as a backbone in other networks and is nice because it is good to run experiments on and also can directly deploy at scale pretty easily. In my experience, cycle time is actually one of the biggest resources for most smaller research initiatives.
I just haven't had the opportunity to talk to anyone yet to see if there's a way that I could interface with the African AI community over it. I don't know how Colab works but they work pretty darn well in the free version of Colab, and both are annotated to help a moderately well-motivated learner pick up some of the ropes themselves.
It seems like you know a little about the space. If you do know anyone or have any pointers as to where I could/should look to talk to people about this, please do let me know! I think OSS can do a lot and would love to give a bit to the community, for sure.
I'm really not sure there's much of an investor interest in helping Africans train their own AIs, your data center costs will be way higher in Africa. And the talent pool there to help with AI is still very small.
There will be a large interest in selling AI to Africans, GPT-4 today, lightly modified, can already radically improve African education. The vastly underpaid and overworked teachers no longer need to waste time in preparing lessons, they just need to control the classroom and perform assessments. Also GPT-4 probably dominates your average English tutor in Africa...
The problem is the cost, a GPT-4 subscription is far beyond the means of an average African. The bottleneck is with server-side inference costs, once that is solved, I expect a AI gold rush in Africa, where western companies compete to sell AI subscription services. Your average African already owns a cheap andrioid phone with 3g connection, which is good enough for an AI chatbot.
On A100s this particular CIFAR10 network trains in under 7 seconds for example and the network (fp16) is <10 MB IIRC, the relative gains over other previous training speeds is even better on older cards IIRC due to it being pretty memory limited at this point. It could be used as a backbone in other networks and is nice because it is good to run experiments on and also can directly deploy at scale pretty easily. In my experience, cycle time is actually one of the biggest resources for most smaller research initiatives.
I just haven't had the opportunity to talk to anyone yet to see if there's a way that I could interface with the African AI community over it. I don't know how Colab works but they work pretty darn well in the free version of Colab, and both are annotated to help a moderately well-motivated learner pick up some of the ropes themselves.
It seems like you know a little about the space. If you do know anyone or have any pointers as to where I could/should look to talk to people about this, please do let me know! I think OSS can do a lot and would love to give a bit to the community, for sure.