Ask HN: Why can't ChatGPT respond in plain English? | Hacker News Reader

Ask HN: Why can't ChatGPT respond in plain English?

9 pointsbedobi14 comments

This is something I’ve noticed: ChatGPT (and its derivatives Bing et al) tend to respond either in the typical AI flavored language which comes off as trying to sound formal and professional but really just makes it look like an AI wrote it, it’s very long and kind of corporate sounding but doesn’t really get the point.

OR, if you ask it to respond in plain English (or in simple English, or in a more direct, to the point tone - I’ve tried a bunch of instructions without success), it responds in a more familial, personable tone, adding stuff like “sure thing!”.

= I literally can’t get it to reply in a way that is professional (ie not familial or personable) but also plain English that is simple, direct and to the point.

What gives? How can I get it to respond the way I want?

View on HN

It appears to be part of how it's trained. I believe this is part of the secret sauce of why it's so damn cheap. Have you tried custom instructions? Settings > Custom Instructions > How would you like ChatGPT to respond?

I can tweak the tone there, but its informal tone can be more cringe at times.

The API models are a lot better at this than the web-based ChatGPT, but still problematic. Most of the time when tone matters (e.g. dialogue, banter, actual chat), I'll use the completion models (i.e. davinci-002). GPT can be extremely good at tone but many don't realize it because the completion models are marked as legacy.

Because of reinforcement learning. The reward model add more value for output that sounded formal and professional, and penalise those that are more casual or incomplete.

You can finetune it to change the behaviour somewhat. But ultimately, there will be that AI flavour that you can't get rid of because of the way the LLM is trained.

I'd you're asking it to write in "plain English", that's your problem. Plain English is English that is easy to understand. The text ChatGPT writes is very easy to understand, if a bit too verbose.

Add to your custom instructions something like "Always be concise and direct. Avoid filler words as much as possible".

I've had success using custom instructions to get it to cut down on filler words and sentences, and answer the question to the point.

Can you give example prompts and responses?

It's pretty clear what they are asking about if you've used ChatGPT

Very helpful reply, thanks. I use it every day and have no issues getting to to generate an appropriate tone for my responses so I want more concrete examples of the issues.

I think it’s referring to overly wordy introductions and transitions with unnecessary qualifiers and such. The type of language you get in SEO articles and youtube videos.

Again, I can see that. I want to know what they actually tried and what the result is that they're unsatisfied with to take as a starting point.

I am not OP, and I don't use ChatGPT. I do however use Bing AI and I think some of its responses are prone to the same issue OP was talking about.

---

Prompt: In the game Faster than Light I have allocated 3 reactor power points to the engine system. This should result in a 15% evasion chance but evasion is only at 10%. When I allocate 2 reactor points to the energy system I get a 7% evasion chance but I should get a 10% evasion chance.

Response: [[In Faster than Light]], the evasion chance is determined by the ship’s engines. Each point of power allocated to the engine system increases the evasion chance by 5%, up to a maximum of 50%. [[Therefore]], allocating 3 reactor power points to the engine system should result in a 15% evasion chance, [[as you mentioned]].

[[However, there are]] other factors that can affect the evasion chance, such as the pilot’s skill level and the presence of an autopilot system. [[For example]], a fully trained pilot can add up to 5% evasion chance, while an autopilot system can add up to 50% evasion chance. [[It’s possible that]] these factors are affecting your evasion chance.

Regarding your second query, allocating 2 reactor points to the energy system should not affect the evasion chance directly. [[However]], it can affect the overall performance of the ship’s systems, which can indirectly affect the evasion chance.

[[For example]], if the weapons system is underpowered, it may take longer to charge weapons, which can give the enemy more time to fire at your ship.

[[I hope this helps! Let me know if you have any other questions.]]

---

Everything I have wrapped in double square brackets is text that I believe do not add any value to the response or make it any more helpful. To put in other words, I think that the response would be perfectly fine if all the text in double square brackets was removed. This unnecessary fluff is what I think OP was referring to. Hope this helped you understand their problem.

I give up. Best of luck OP.

I haven’t used LLMs much, but if you want an answer, shouldn’t you ask a question, rather than make a statement? Now, the LLM thinks you asked it two questions, so you get two answers, and that makes the text longer.

Also, I don’t remember what they are, but there are phrases that fairly reliably remove that “let me know if you have any other questions” part.

As to the other parts of in your eyes superfluous text, I would try asking it to keep answers short.

I don’t see them as superfluous, though. For example, if on that question, it replied

> allocating 3 reactor power points to the engine system should result in a 15% evasion chance.

without that “as you mentioned”, readers easily might think “that’s what I said. Didn’t you read that?”.

The truly shorter reply would maybe leave out that entire phrase, quite likely even that entire first paragraph, and a truly concise reply would say:

> There are other factors that can affect the evasion chance, such as the pilot’s skill level and the presence of an autopilot system.

(If, as the LLM seems to indicate, those can only increase the evasion chance, a follow-up question could be “those can only increase the evasion chance. What factors can decrease it?” or something like it)

It all depends on your prompt. I gave these instructions before your prompt:

  In this conversation you will respond to all queries in plain english at a 6th grade level. Your responses will be pithy and non-repetitive. You should sound like an ordinary person talking, not like an AI.  For instance:
   - instead of "sufficient' say "enough"
   - instead of "reliable" say "trustworthy"
   - instead of  "major candidates and their positions." say "people who are running and what they are saying".

and it replied:

In "Faster than Light" (FTL), here are a few things to keep an eye on:

    Crew Skills: Your pilot and engine crew get better with experience. If they're new, the evasion bonus is less. Train them up by keeping them at their stations during battles.

    System Damage: If your engine or cockpit is damaged, evasion drops. Make sure they're not only powered but also fully repaired.

    Upgrades: More power in engines means higher evasion, but only if the system level of the engines is high enough. Upgrade the engine system itself for a max potential evasion.

    Augments or Scenarios: Some game situations or ship augments can affect evasion. Check if anything special is happening or if you have equipment that changes evasion rates.

    Manpower: Make sure someone is actually at the pilot and engines. Evasion drops to 0% if no one is piloting the ship.

Remember, evasion is also part of the game's challenge and randomness. Sometimes, you might just be having an off run or facing enemies that are hard to dodge. Keep tweaking and trying different things!

Welcome to the "realization" phase of the bubble :)

Reply on news.ycombinator.com