By the 24‑month mark (early December 2022), multiple non‑Google/DeepMind groups had publicly demonstrated protein‑folding models with AlphaFold‑comparable accuracy, effectively replicating the AlphaFold2 result.

• HelixFold (Baidu/PaddlePaddle team, July 2022) – HelixFold is an independent reimplementation of AlphaFold2 using PaddlePaddle. Its authors report that, on CASP14 and CAMEO benchmarks, HelixFold’s accuracy is “on par with AlphaFold2” and they released code and a public web service. This is a non‑Google group matching AlphaFold2’s accuracy on standard benchmarks well before December 2022, i.e., a direct replication of the AlphaFold2 result. (arxiv.org)
• RoseTTAFold (Baker lab, University of Washington, Science 2021) – An academic group developed RoseTTAFold, a three‑track neural network that predicts 3D protein structures and protein–protein complexes with high accuracy. Contemporary commentary notes it is based on the AlphaFold2 framework and achieves high‑accuracy structure prediction, though “not quite as accurate as Alphafold2,” which had been kept private at the time. (espeyearbook.org) This still shows an independent deep‑learning model reaching AlphaFold‑like performance in many cases within 2021.
• ESMFold (Meta AI, 2022) – By November 2022, Nature was describing Meta’s ESMFold as “AlphaFold’s new rival” in protein structure prediction; it was characterized as somewhat less accurate than AlphaFold2 but still an accurate, large‑scale structure predictor, and markedly faster. (en.wikipedia.org)

In addition, reviews from 2022 summarize that several deep‑learning methods (including RoseTTAFold and other two‑step approaches) can reach accuracy similar to state‑of‑the‑art AlphaFold2 on many targets, reinforcing that comparable performance had been achieved outside DeepMind within this time window. (academic.oup.com)

Because at least one non‑Google/DeepMind group (notably HelixFold, and arguably RoseTTAFold and ESMFold as well) had publicly demonstrated AlphaFold‑level or near‑AlphaFold‑level protein‑folding accuracy before December 2022, Friedberg’s prediction that “in the next 24 months, someone else will replicate this” is fulfilled.