My best ideas are my old ideas

LLM’s make so many previous ideas I didn’t think were possible before a lot more possible. It’s like a new lens to the world or some kind of new superpower.

I have noticed with friends actively building on top of the OpenAI API that they often will see workarounds to challenges that even I miss, thanks to LLMs, which anyone would have seen as blockers in the past. To put it simply, a lot more is just possible now. No hurdle is quite the same.

The capabilities are also improving it appears every 18 months or so. Ideas which weren’t possible with GPT-3 may be possible with GPT-4 and so on and so forth.

I noticed with the DALL-E 2 release that the timing of a prompt really mattered for the best artistic results, but maybe we are better off just either waiting, keeping track of our ideas, and simply retrying as chips improve and AI progresses.

Likewise, in the corporate IT world, you might as well we better off extending your project a couple more months until something like GPT-5 comes out before going to management and fully calling it quits. It may not only be able to do the task at hand soon, but could have new capabilities as well which you could not have imagined. The performance in the task may jump from something unreliable to world class in that time as well.

It’s a lot like software programmers in the 90’s just waiting for hardware to get better, except now we are often better off simply waiting until the GPT series makes the next leap to a higher level of intelligence.

You may also be wasting a lot of your own effort in the next few months working while you are waiting between GPT-4 and GPT-5. Imagine working astronomically harder to optimize and get GPT-4 to do something (including curating labelled fine tuning training datasets or some kind of advanced prompt chaining) when you could have just waited and seen superhuman levels of performance out of the box with GPT-5.

I often think there is something unique and special about our initial ideas, our first inclinations towards something. I would say the challenge nowadays is waiting until the exponential curve of AI progress catches up with our imagination and underlying ambitions. In fact, I wonder how far back I should really be going. Should I be going back to my best ideas, even as a kid?