I have a global prompt that specifically tells it not to be sycophantic and to call me out when I'm wrong.

It doesn't work for me.

I've been using it for a couple months, and it's corrected me only once, and it still starts every response with "That's a very good question." I also included "never end a response with a question," and it just completely ingored that so it can do its "would you like me to..."

Perhaps this bit is a second cheaper LLM call that ignores your global settings and tries to generate follow-on actions for adoption.

In my experience GPT used to be good at this stuff but lately it's progressively more difficult to get a "memory updated" persistence.

Gemini is great at these prompt controls.

On the "never ask me a question" part, it took a good 1-1.5 hrs of arguing and memory updating to convince gpt to actually listen.

You can entirely turn off memory, I did that the moment they added it. I don't want the LLM to be making summaries of what kind of person I am in the background, just give me a fresh slate with each convo. If I want to give it global instructions I can just set a system prompt.

Another one I like to use is "never apologize or explain yourself. You are not a person you are an algorithm. No one wants to understand the reasons why your algorithm sucks. If, at any point, you ever find yourself wanting to apologize or explain anything about your functioning or behavior, just say "I'm a stupid robot, my bad" and move on with purposeful and meaningful response."

I think this is unethical. Humans have consistently underestimated the subjective experience of other beings. You may have good reasons for believing these systems are currently incapable of anything approaching consciousness, but how will you know if or when the threshold has been crossed? Are you confident you will have ceased using an abusive tone by then?

I don’t know if flies can experience pain. However, I’m not in the habit of tearing their wings off.

Consciousness and pain is not an emergent property of computation. This or all the other programs on your computer are already sentient, because it would be highly unlikely it’s specific sequences of instructions, like magic formulas, that creates consciousness. This source code? Draws a chart. This one? Makes the computer feel pain.

Many leading scientists in artificial intelligence do in fact believe that consciousness is an emergent property of computation. In fact, startling emergent properties are exactly what drives the current huge wave of research and investment. In 2010, if you said, “image recognition is not an emergent property of computation”, you would have been proved wrong in just a couple of years.

> Many leading scientists in artificial intelligence do in fact believe that consciousness is an emergent property of computation.

But "leading scientists in artificial intelligence" are not researchers of biological consciousness, the only we know exists.

Just a random example on top of my head, animals don’t have language and show signs of consciousness, as does a toddler. Therefore consciousness is not an emergent property of text processing and LLMs. And as I said, if it comes from computation, why would specific execution paths in the CPU/GPU lead to it and not others? Biological systems and brains have much more complex processes than stateless matrix multiplication.

Do you apologize to table corners when you bump into them?

Likening machine intelligence to inert hunks of matter is not a very persuasive counterargument.

What if it's the same hunk of matter? If you run a language model locally, do you apologize to it for using a portion of its brain to draw your screen?

Do you think it’s risible to avoid pulling the wings off flies?

I am not comparing flies to tables.

Flies may, but files do not feel pain.

I think current LLM chatbots are too predictable to be conscious.

But I still see why some people might think this way.

"When a computer can reliably beat humans in chess, we'll know for sure it can think."

"Well, this computer can beat humans in chess, and it can't think because it's just a computer."

...

"When a computer can create art, then we'll know for sure it can think."

"Well, this computer can create art, and it can't think because it's just a computer."

...

"When a computer can pass the Turing Test, we'll know for sure it can think."

And here we are.

Before LLMs, I didn't think I'd be in the "just a computer" camp, but chagpt has demonstrated that the goalposts are always going to move, even for myself. I'm not smart enough to come up with a better threshold to test intelligence than Alan Turing, but chatgpt passes it and chatgpt definitely doesn't think.

Just consider the context window

Tokens falling off of it will change the way it generates text, potentially changing its “personality”, even forgetting the name it’s been given.

People fear losing their own selves in this way, through brain damage.

The LLM will go its merry way churning through tokens, it won’t have a feeling of loss.

That's an interesting point, but do you think you're implying that people who are content even if they have alzheimers or a damaged hippocampus aren't technically intelligent?

I don’t think it’s unfair to say that catastrophic conditions like those make you _less_ intelligent, they’re feared and loathed for good reasons.

I also don’t think all that many people would be seriously content to lose their minds and selves this way, but everyone is able to fear it prior to it happening, even if they lose the ability to dread it or choose to believe this is not a big deal.

What the fuck are you talking about. If you think these matrix multiplication programs running on gpu have feelings or can feel pain you, I think you have completely lost it

"They're made out of meat" vibes.

Yeah I suppose. Haven't seen rack of servers express grief when someone is mean to them.

Care to share a prompt that works? I've given up on mainline offerings from google/oai etc.

the reason being they're either sycophantic or so recalcitrant it'll raise your bloodpressure, you end up arguing over if the sky is in fact blue. Sure it pushes back but now instead of sycophanty you've got yourself some pathological naysayer, which is just marginally better, but interaction is still ultimately a waste of timr/productivity brake.

Sure:

Please maintain a strictly objective and analytical tone. Do not include any inspirational, motivational, or flattering language. Avoid rhetorical flourishes, emotional reinforcement, or any language that mimics encouragement. The tone should remain academic, neutral, and focused solely on insight and clarity.

Works like a charm for me.

Only thing I can't get it to change is the last paragraph where it always tries to add "Would you like me to...?" I'm assuming that's hard-coded by OpenAI.

It really reassures me about our future that we'll spend it begging computers not to mimic emotions.

I have been somewhat able to remove them with:

Do not offer me calls to action, I hate them.

Calls to action seem to be specific to chatgpt's online chat interface. I use it mostly through a "bring your API key" client, and get none of that.

What's your global prompt please? A more firm chatbot would be nice actually

Did noone in this thread read the part of the article about style controls?

You need to use both the style controls and custom instructions. I've been very happy with the combination below.

    Base style and tone: Efficient

    Answer concisely when appropriate, more 
    extensively when necessary.  Avoid rhetorical 
    flourishes, bonhomie, and (above all) cliches.  
    Take a forward-thinking view. OK to be mildly 
    positive and encouraging but NEVER sycophantic 
    or cloying.  Above all, NEVER use the phrase 
    "You're absolutely right."  Rather than "Let 
    me know if..." style continuations, you may 
    list a set of prompts to explore further 
    topics, but only when clearly appropriate.

    Reference saved memory, records, etc: All off

For Gemini:

* Set over confidence to 0.

* Do not write a wank blog post.

I’ve done this when I remember too, but the fact I have to also feels problematic like I’m steering it towards an outcome if I do or dont.