They include the enwik8 BPC estimates. It may not be strictly comparable - the most obvious issue is that since the HP takes the compression paradigm of intelli

They include the enwik8 BPC estimates. It may not be strictly comparable - the most obvious issue is that since the HP takes the compression paradigm of intelligence, the compressor size is part of the total, and those 1.5b parameters certainly are not cheap. (This is one reason I think the HP is outdated. The corpus is not big enough to allow the superior asymptotics of approaches like RNNs or Transformers to compensate for their far larger binary size. HP is not measuring progress on an intelligence metric we care about, it's sort of measuring a 'demo scene' metric of intelligence.)