9 posts were merged into an existing topic: Discussion of Covid-19 information (No holds barred)
5 Likes
The robot take over has begun! Elon announces the Tesla-bot.
5 Likes
He’s already a robot 
.
3 Likes
A 9 pflop cpu !
“This is real.”
4 Likes
1 Like
2 Likes
6 Likes
Debatable if you aren’t able to take possession of the asset
2 Likes
If you have a PayPal business account you do not have access to crypto services. Beyond me as to why.
State of the art deep learning increasingly requires specialized supercomputers to train, leaving it pretty much out of reach for anyone but the tech giants. The reason for this is that the algorithm used to train neural nets requires tons of communication between each computing node, so if you need 10 000 GPUs to train a neural net, you can’t just get 10 000 gamers to join some decentralized neural net training app, because they don’t have the required 400 GBps connection between them, so instead they’re all stuffed together in a datacenter with the fastest networking equipment money can buy.
If some new algorithm could be found that could train the neural nets without needing constant high-bandwidth communication, that would open up the possibility for using unused GPUs to join training of more state of the art deep learning models. Hugging Face has recently published steps towards this.
Hugging Face has done some tests in a semi-centralized with an authentication server and whitelist of participants, in other words a permissioned peer-2-peer network.
Getting this to a permissionless decentralized solution requires much of what SAFE is doing like decentralized storage, crypto payments, byzantine fault tolerance, so it would be interesting to see this as a SAFE app or plugin in the future.
13 Likes
nice find. A critical fine point in the future of distributed computing. There are algorithms to partition communication graphs to optimize this problem, but it’s going to be tricky without allowing direct connections between computing units.
3 Likes
wouldn’t that require to execute the tasks multiple times? making it a lot less efficient.
Yes it would. I suppose in this case there could perhaps be other ways around it. Since training a neural network is just updating the weights on each artificial neuron, if some of these updates are random bullshit numbers, it could propbably be averaged out or fixed over time so it doesn’t have much of an impact. Transfering tokens requires much higher tolerance and with neural network training there isn’t anything for malicious users to gain other than to cause damage, so it’s unlikely many will do and it even has a cost for the malicious user, so it will most likely be fine to just run the computations once with more basic fault tolerance.
1 Like
We may be needing something like this too someday.
9 Likes
2 Likes
Yep. Atomic swaps ensure the survival of a cryptocurrency and enable future use on dexes. SAFE has all the primitives needed and will have a few battle hardened examples for implementation by the time it launches. I would rank atomic swap functionality as a pretty high priority for dev time (Bamboo fund?) after launch.
9 Likes
Dogecoin is for the dogs … literally!
1 Like