The posted article isn't particularly fascinating, but for a bit of fun, there's an OpenAI project where they demonstrate that due to the non-linear rounding of

The posted article isn't particularly fascinating, but for a bit of fun, there's an OpenAI project where they demonstrate that due to the non-linear rounding of Float32 values you can actually train "non-linear" linear networks: https://openai.com/blog/nonlinear-computation-in-linear-networks/ https://openai.com/blog/nonlinear-computation-in-linear-netw...