Software 1.0 was code | Software 2.0 was Neural Networks
Now we are in the LLM space, were we can send prompts to neural networks that can generate programs, that is wild idea
We need to create systems around LLMs, like interface to interact LLMs, He puts it like we are in the 1960s era of Computing. Where the computers were not personal, they were mainframe, large machines that fit on factories.
However the LLMs today are like those big gigantic computers that live in the cloud running in multiple clusters of GPUs, though we have local LLMs, the quality and accuracy of high parameter models is not near the local small sized models (they are improving). So we are reliant on Cloud inference like ChatGPT, Claude, Mistral, LLaMA, Gemini, etc.
He relates the transition of LLMs with the augmentation to an autonomous agent. This is compared to Iron Man’s suite, which is a mixture of human control along with AI assistance. When he is outside the augmented mode, its his instinct that help him navigate, however, in the agent mode, the AI decides the track.
Also the vibe coding term was emphasised which has a Wikipedia entry. Vibe coding is easy to do because we don’t care about the code, we care about the outcome, we can call it product driven prompting. And it makes sense when he said, we need to increase the Human-verification and AI-generation loop. It doesn’t mean, you’ll tell AI to one-shot the entire thing, instead go bit by bit, one thing and one action and feedback at a time.
