depends on how you define "step". Engineer a 10x/100x version of what we have in terms of LLM (either by being more efficient and/or more/specialized hardware) and let this thing build novel attempts for AGI algorithms 24/7 in a evolutionary setting.

I guess the challenge is more to agree on a fitness function to measure the "AGI"-progress" against, but thats a different topic. But in general scaling up the current GenAI tech and parallelize/specialize the models in a multi-generational way _should_ be a safe ticket to AGI, but the time scale is inknown of course (since we can't even agree on the goal definition)