The DESI DR2 claim of evolving dark energy rests on a frequentist analysis of seven BAO data points in a two-parameter model. I will present a comprehensive Bayesian reanalysis using the unimpeded nested sampling database — 248 dataset & model combinations spanning Planck, DESI, DES, Pantheon, ACT, SPT and SH0ES — showing that the evidence for dynamical dark energy is weaker than advertised and largely driven by inter-dataset tension rather than genuine evolution. Moving beyond w0wa, I will show flexible dark energy reconstructions using transdimensional flexknot models in JAX on GPU, where the data themselves select the model complexity, and argue that a supernova magnitude offset provides a better fit than exotic dark energy. More broadly, I will make the case that GPU-accelerated classical statistical methods — nested sampling, HMC, Laplace approximation — are competitive with neural network approaches across astronomy, from gravitational waves to 21-cm cosmology, and that large language models are transforming how we build and verify these analyses without replacing the science itself.