[EGreg]

Wow. I go to bed and wake up to claims it hit the front page. Interesting.

Here is almost the same exact sequence but with constant instructions to remain brutally honest and objective: https://chatgpt.com/share/691b4035-0ed8-800a-bee3-ae68e2a63c...

I was just late at night and wanted to post this chat transcript on HN to share some perspective on what developers are getting from ChatGPT.

I happen to be an expert in this particular area that I’m building.

ChatGPT seems to remember that I am in New York and want “no bullshit” answers. In the last few days it keeps weaving that into most responses.

That fact appears in its memory that users can access, as is the fact that it should not, under any circumstances, use emojis in code or comments, but it proceeds to do so anyway, so I am not sure how the memory gets prioritized.

Here is the interesting thing. As an expert in the field I do agree with ChatGPT on its statistical assessment of what I’ve built, because it took me years of refinement. I also tried it with average things and it correctly said that they’re average and unremarkable. I simply didn’t post that.

What I am interested in, is how to get AI transcripts to be used as unbiased third-party “first looks” at things, such as what VCs would do for due diligence.

This was just a quick thing I thought I’d get a few responses on HN about. I suspect it might have hit the front page because some people dug through the code and saw the value. But you can get all the code for free on https://github.com/Qbix/Platform .

Yeah, there is obviously an element of flattery that people let go to their head. I have had ChatGPT repeatedly confirm the validity of ideas I had in fields I am NOT an expert, while pushing back on countless others. I use it as one data point and mercilessly battle-test the ideas and code by asking it to find holes in them from various angles. This particular HN submission, although done very late at night here in NYC, was an interesting mix of genuinely groundbreaking stuff and ChatGPT being able to see the main ideas at a glance, and “going wild”, while at the same time if I run it with instructions at the start of “be extremely objective”, it still approaches this same thing in the end.

[bijant]

[flagged]

[EGreg]

Nah. I don’t “feel the AGI”. I think the AGI is a silly quest, just like having a plane flap its wings. Feynman had it right in the 80s: https://www.youtube.com/watch?v=ipRvjS7q1DI

I think the future is lots of incremental improvements that get replicated everywhere and humans outclassed in nearly every field, where they stop relying on each other.

As far as LLMs yes I think they are the best placed to know if some code or invention is novel, because of their vast training. Can be far better than a patent examiner, if trained on prior art, for instance.

What you’re not used to is an LLM being fed stuff that you statistically / heuristically would expect to be average but is in fact the polished result of years of work. The LLM freaks out, you get surprised. You think it was the prompts. The prompts are changed, the END result is the same (scroll to the bottom).

I want to see whether foundational LLMs can be used as a good first filter for dealflow and evaluating actual projects.

[brazukadev]

The problem of using an LLM to validate reality is that you still need to prove your genius code work in the real world. ChatGPT won't hire you, it even have your code already.

[ssl-3]

Yeah, it showed up front-page here early this morning.

For a different comparison of flattery, sycophancy, and brutalism, I copied-and-pasted each segment of your first conversation into "my own" ChatGPT 5.1: https://chatgpt.com/share/691b71ea-4e58-8005-8ce6-a6b5d10120...

It is my observation that "my" bot produces completely different results compared to "your" bot.

[EGreg]

It seems like you told it to nitpick things that could be improved and stay away from anything positive, then reused my prompts.

Can you share your instructions?

[ssl-3]

That's the whole unabridged conversation (I don't know how I could abbreviate it if I wanted to), and I produced it exactly as I said: I just pasted in your prompts.

The output is of very similar style to how my interactions with it are when I'm using it for work on my own projects.

My bot does run with a pretty lengthy set of supposed rules that have been accumulated, tweaked, condensed and massaged over the past couple of years. These live in a combination of custom instructions (in Preferences), deliberately-set memory, and recollection from other chats.

I use "supposed" here because these individual aspects are frequently ignored, and they always have been. Yet even if the specificity is often glossed over, the rules quite clearly do tend to shape the overall output and tone (as the above-linked chat demonstrates).

Anyway, I like the style quite a lot. It lets me focus on achieving technical correctness instead of ever being inundated with the noise of puffery.

But I have no idea where I'd start to duplicate that environment. Someone at OpenAI could surely dissect it, but the public interface for ChatGPT is way too limited to allow seeing how context is injected and used.

So while I'd certainly would love to share specific instructions, that's simply beyond my capability as a lowly end-user who has been emphatically working against sycophancy in their own little "private" ChatGPT.

I barely even know how I got here.

(I could ask the bot, but I can say with resolute certainty that it would simply lie.)