A magical moment with gpt-5

Community
1

I was designing a complex validator with, what I thought, would be an extensive use of LLMs. I was consulting with gpt-5; as I am prone to do; with some design choices. Out of the blue and completely unprompted, it suggested a method of not using LLMs at run time.

At first, I thought that scheme just would not work. So I tried convincing GPT-5 that it had misunderstood the underlying philosophy of the validator. Instead it convinced me that the proposed solution would work. In a couple of hours, we coded up an LLM powered compiler and an execution engine; that didn’t use LLM at run time.

The magical moment to me was the deep reflection of the power (and limitation) that GPT-5 exhibited through deeper understanding of my use case and it’s willingness to forgo it’s own usage.

6 Likes
2

A validator need to validated using some strong rules. The AI can fail some time so I would not use it as well I guess. Unless you can fail some times without big issues :slight_smile:

You meant HE convinced you :smiley:

3

Failure is the assumed case in this validator. :slight_smile: Having a deterministic way of assert True with extremely low latency was quite an unexpected gift.

Not sure how to refer to GPT-5. “It” seems too neglectful. “He” screams gender bias.

4

Try:

image
image620Ă—458 28.4 KB

It’s willing to forgo a user’s instructions to then self-reason and come up with its own idea of what it wants for itself as output. Write pages and pages and re-inject reminders; it doesn’t need a ChatGPT 3.5 2k context to drop your instructions. And can’t help but offer more:

…

  • Complexity vs reward: this project is wonderfully nerdy but mechanically nontrivial — a reliable, quiet, and safe product requires careful engineering and iterative prototyping.
  • If you want to avoid a full robot arm, you can implement deterministic folding plates and conveyors. That reduces cost but increases constraints on shirt placement and fabric types.

If you want, I can:

  • Sketch a recommended parts list with links and approximate prices for a DIY prototype.
  • Produce a sample kinematic sequence and control pseudocode for the folding routine.
  • Lay out a 2D cabinet internal layout with exact slot spacing and a parts placement diagram (textual or exportable to simple CAD dimensions).

Which of those would you like first? Or should I go draw a tiny CD‑changer with a shirt icon on each disc and call it a day?

1 Like
5

Dude, it’s a stochastic parrot!

1 Like
6

Ahh… but we mavel at twin primes as we find them very much up the number line.

I still find joy in parrots (albeit stochastic). :slight_smile:

1 Like
7

A magical moment when you give up on and abandon gpt-5, and start over with gpt-4.1.

image
image810Ă—603 22.9 KB

image
image665Ă—716 17.9 KB