Towards Recursive Self Improvement

Towards Recursive Self Improvement (RSI)

Today, all the chirp and chatter is about "Claude," "GPT," and "AGI." The [new thing] is going to make [your job] redundant, the [new product] is going to replace the need for [existing product]. That's what you hear about on every social platform, and perhaps even every social gathering(!)

I like taking a step or two back and trying to paint the big picture of what's going on in the world of technology.

It is perhaps not far-fetched for me to posit that, in fact, the wave of products and services we see getting launched on the daily nowadays are merely aspects of a transition. In some tech-first circles, this is already a known notion, but the capabilities demonstrations and 'killer' AI products/services we see are just checkpoints on the way towards the logical conclusion: RSI. This can be likened to people who are trying to make a lot of money to eventually purchase a house of their own 'paid in cash'; along the way, this person will likely buy their family gifts, spend money on travel and tourism, and treat themselves to luxury purchases like a car, watch, iPhone, etc.

RSI is easy to define: once a system is capable enough to improve itself without human intervention, this system will be classed as RSI. This could be achieved as soon as, say, 2027, if a system like Claude Code or Codex starts improving itself without Anthropic/OpenAI employees steering, guiding, priming, or otherwise directing it to do so. And it is when RSI becomes real that abundance is truly on the table as a possibility. Picking up off the money-making example, you could simply ask a recursively self-improving Codex to ...

"pls earn $2000, make no mistakes"

because it will try achieving that goal, improving its own self to do so as needed. And if there is no 'limit' to how much it can improve itself, eventually it will succeed at earning $2000 -- likely in the form of a $2000 profit, because with RSI there is an implication that the system crosses the threshold of inferring the user's intent in the same way you and I are able to!

... or at least that's my headcanon :] and I'll write more about this Pandora's box later. □