Parasma

Modern AI compute relies heavily on GPUs for tasks such as predicting the next word in a sentence or learning actions through reinforcement learning. Parasma frames this as an area where biological systems could serve as an alternative substrate for computation. The site does not describe who specifically is burdened by GPU dependence or quantify the cost of current approaches, only that biological neurons could carry out comparable functions.

Parasma trains brain cells to perform computational tasks by converting inputs into electrical stimulation patterns delivered to living neurons. For token prediction, a sentence is converted into stimulation, the neurons carry that context, and their responses are decoded directly into output tokens, with the company stating no GPU is needed for this process. For reinforcement learning, each frame of an environment is turned into a new stimulation pattern, the neurons select an action, and the resulting outcome is measured and fed back into the neurons within a changing environment. These two capabilities, token prediction and reinforcement learning via neuron stimulation and decoding, are the specific features described on the site.