0xSero compares today’s reliance on remotely hosted AI with the earlier era of shared computers, arguing that local modelsA local model runs on computing infrastructure controlled directly by the user rather than only through a remote model service. could make useful AI more accessible and give individuals more control.
The presentation introduces model weightsModel weights are numerical parameters learned during training that help determine how a model transforms inputs into outputs. as files containing numbers learned during training and uses next-word prediction to explain how a language model generates textInference is the use of a trained model to produce an output from supplied input..
0xSero describes moving model files from storage into RAM or GPU memory, and explains the trade-off between available memory, model size, speed and hardware cost.
In response to an audience question about cybersecurity, 0xSero argues that local models can offer an alternative when a hosted provider’s safeguards limit a particular request. These claims are presented as the speaker’s perspective, not as a guarantee that any particular security request will be refused.
Watch on YouTube
