In the age of Artificial Intelligence, one thing that everybody keeps talking about is hyperscaling. These massive entities are incentivized to recursively analyze copious amounts of data for their training with tons of garbage collecting, wasted compute, and environmental problems. Society is also making a conscious decision on trusting these firms with lots of economic and creative power. This is the whole problem with centralization and with every other major AI company wanting to get there first.

Nobody is even considering the ethical ramifications on what these AI systems could have on society, at least not to the point of pushing for the transparency the general populus probably are hoping for.

As the founder of Riemann Computing, creator of the decentralized-internet SDK, and a researcher, I want to change that. Riemann Computing is a company centered around building technologies wanting to disrupt telecom and infrastructure as industries, with the core being data compression. We previously won Hackernoon’s Startup of the Year for the Electronic category and invented a new computing paradigm called “Signal Computing”. In the current landscape of AI, one thing stands though, and that is the cost of training.

In the era of Anthropic and Open AI, numerous companies are competing to develop the next frontier model, yet they have not adequately addressed the mathematical issues tied to the enormous training expenses and the challenges associated with that training. Additionally, there exists a significant limitation: rather than allowing large hyperscalers to create or train AIs, I envision democratizing AI training and empowering everyday individuals with that capability.

Initial Training Results

This brief demonstration serves to illustrate the vast potential that Riemann Computing aspires to achieve. That is why recently, I launched the pre-release for “My Name is Riemann”, and the OpenPeer AI line of AI software on Huggingface. If you want to support us, you can give us a follow on Huggingface and check out what we are trying to build.

The pre-release of "My Name is Riemann" utilized $3K worth of hardware during its initial training phase locally, followed by approximately $640 spent on computing resources via a DigitalOcean node. This DigitalOcean node consisted of an 8x AMD MI325X GPU droplet located in nyc2, costing $30.40 per hour, and was employed for stages 2 through 6. In total, it operated for 21.0 node-hours (equating to 168 GPU-hours: 19.1 hours dedicated to pretraining, with the remaining time allocated to growth, controller fitting, setup, and upload), resulting in a total expenditure of around $640.

The idea comes from taking the OpenPeerLLM model and training it up based off of a handful of datasets utilizing the NTK Trainer. The objective was not to develop a fully functional chatbot or a commercially viable multimodal model, but rather to establish a solid foundation of a reasonable size that a contextual layer could implement on top of, or that could be scaled up using previously saved training checkpoints instead of repeatedly altering weights and starting from scratch.

Example Plot of Gate Learning - NTK Trainer

The concept is straightforward: if it's feasible to create a 13B parameter model that produces somewhat reasonably accurate strings on a limited budget, it paves the way for individuals to train and develop AIs locally. This shift could lead to significant decentralization, moving AI models away from being controlled by a select few organizations.

On the other hand, Riemann Computing aims to sell proprietary codecs, and enterprise-level systems. Our vision is to create an ecosystem rather than being the sole authority in the AI revolution. The objective is straightforward: to maintain openness and transparency. The future should not be controlled by a select few; it should belong to the creators. I am Andrew Magdy Kamal, and I hope to see the creator economy. Cheers!