The proto-AGI could wih its long term memory and ability to grow its neural network should be able to programm much better than Codex or alpha-code. While also understanding the software achtitectures much better and thus be able to help create a monolithic (solved in one architecture and not like the proto-AGI that is more like a patchwork of different programms) AGI that is maybe build a little like https://futureai.guru/technologies/brian-simulator-ii-open-source-agi-toolkit/ but much better scalable and usable and thus 2-3 orders of magnitude faster and effective ( maybe even usable for robots by then ).