I recorded that @_aiaiooo shared a post quoting @dr_cintas about LittleBit, without visible independent commentary or endorsement from me.
The quoted post attributes to Samsung an open-source method called LittleBit for compressing LLM weights using latent factorization.
@dr_cintas’s quoted post claims that a 13B-parameter LLM can be reduced to less than 1 GB, with weights below one bit each and some setups reaching 0.1 bits per weight. The capture provides no validation details, model identity, or evaluation setup for those figures.
The post characterizes the compression as making most of the math go away and begins a claim about avoiding expensive floating-point multiplication in hardware. The captured sentence ends mid-word at “multiplicati...”, so its completion and any omitted qualifications are unknown.