Watch out ChatGPT, Petals is Here
The new experiment is attempting to develop a decentralized system that can run AI by joining resources from people across the internet
One can never have enough of ChatGPT, the text-generating application that’s creating a huge buzz in the field of advertising to copy writing and everything in between. Now, a community-based project is out there developing a system that can run artificial intelligence like the above by connecting resources from people across the internet.
And, it’s called Petals – a project run by the igScience community and backed by Hugging Face. Its code has already been released publicly last month and volunteers can now donate hardware power to tackle a part of the text-generating workload or even team up with others on other tasks in a distributed method.
Decentralization is the key mantra
A report published by TechCrunch quoted lead developer Alexander Borzunov as saying that Petals is a collaborative project of researchers from Hugging Face, Yandex Research and the University of Washington. Their major challenge is to reduce the expenses required to run text-generating AI, which some say cost ChatGPT developers $100K per day.
The report quoted Borzunov as saying that unlike APIs that are typically less flexible, Petals is totally open source and allows researchers to integrate latest text generation and system adaptation methods that are yet not available on APIs and do not allow access to the system’s internal states to study its features.
Who pays for all the lavishness?
Having used ChatGPT a few times, our team here can confirm its utility in creating a first-cut article on topics as wide ranging as fashion jewels to long-short investment opportunities. Of course, there are kinks, but what is mind-boggling is the levels of maturity that such as the system can gain as it feeds on data through machine learning.
Which brings us to the moot question of who pays for this lavishness? If indeed it costs as much as some estimates are suggesting, users like us may be forced to stay away, given the costs. And one cannot blame the developers, as running cutting-edge text-generating AI requires substantial financial resources.
However, this is where Petals is looking to change things via democratization. Borzunov had earlier focused on training AI systems over the internet and Petals is looking to ape this so that the costs of running the text-generating AI is reduced drastically. The research team is clear that this journey represents a shift from large models confined to supercomputers to smaller ones.
Volunteer efforts to the fore
The story further quotes Colin Raffel, a researcher at Hugging Face to remind us of the gold rush of last year in the open source text generation community. It was largely volunteer efforts and funds from big tech research labs that created a “bleeding edge” text-generating AI that was once beyond small time developers.
Most of these experiments required powerful hardware and this is where Petals network comes into its own. Borzunov says it would be powerful enough to run and fine-tune AI systems for chatbots and other interactive apps once at sufficient capacity. Users can install an open source library and connect to the Petals network via its website.
Once this is done, they can generate text from Bloom running on Petals or even configure a separate server to contribute back. And, the more the servers, the better the network as if one goes down, Petals immediately seeks to find a replacement automatically. For now, the servers disconnect in case one isn’t active for 1.5 seconds just so they can save on resources.
What do the tests indicate now?
Early tests indicate that basic text generation took only about a couple of seconds whereas complex ones such as writing a note about the universe took about 20 seconds. For those of us who’ve used ChatGPT, this seems a tad slow, but then what about the time when they start asking for money per use?
Currently, it isn’t clear about the size of the Petals network but what seems sure is that it promises to create a low-cost option or even a free one, if and when others start charging for the text-generating services via AI.
However, there are challenges galore
Of course, just as ChatGPT blows a fuse when confronted with too many variables or around context that isn’t clear to it, even Petals is a work in progress. However, the big challenge could be security concerns as the GitHub page of the project notes. The need to share data with other users in the network such as names and phone details needs to be broken into code.
The other challenge is that the text generating system does not differentiate toxicity as a variant to the information shared by the user. Which is why a researcher at Yandex Research states clearly that for now Petals is only intended for research and academic use.
Which means it would be some time in the future that we get our hands on Petals to see whether it rises above the existing text-generating AI systems. Till then, let’s wait and watch.
