Temperature Check: How to Use ChatGPT's Best Hidden Feature

By Thought Leaders Archives
Cover image for  article: Temperature Check: How to Use ChatGPT's Best Hidden Feature

This week, we'll take a close look at the best feature of some text-based generative AI engines that you're probably not using, and we'll also come up with some names for three-legged pets.

There's something for everyone here.

The setting is "temperature," and whether you take it literally or metaphorically, it can transform your brainstorms with or without ChatGPT.

When using the setting in AI apps, the lower the temperature, the more literal the text output. The higher the temperature, the more all rules go out the window. Sometimes, it means even good taste goes out the window, and we'll get to that soon, thanks to a revelation by Google Bard. ChatGPT got weird, and it even lied to me repeatedly about its capabilities, but it at least didn't make me cringe and tell me not to use its own responses.

I asked ChatGPT A LOT of questions about its temperature settings, and I'll spare you the whole thread, but here's where ChatGPT said why you should use it:

"A marketer might want to change the temperature settings in ChatGPT to control the balance between creativity and consistency in the AI-generated content. Different temperature settings can cater to various marketing objectives and requirements:

  1. Lower temperature (e.g., 0.1 - 0.4): Produces more focused, conservative, and consistent responses. This is useful when the marketer needs factual information, precise answers, or messaging that adheres closely to a specific format or brand guideline.
  2. Moderate temperature (e.g., 0.5 - 0.7): Strikes a balance between creativity and consistency. This setting can be useful for general content generation, where a blend of accuracy and inventiveness is desired.
  3. Higher temperature (e.g., 0.8 - 1.0): Generates more creative, diverse, and unexpected outputs. Marketers may prefer this setting when brainstorming innovative campaign ideas, crafting engaging social media content, or seeking fresh perspectives on a topic.

By adjusting the temperature settings, marketers can fine-tune ChatGPT's output to better align with their specific needs, objectives, and desired level of creativity or consistency."

Here's an example to see the difference. I asked, "What should a marketer name their new blog about marketing to pet owners who have pets with gastrointestinal disorders? Set the temperature to 0.1 and give me 5 concise blog names."

ChatGPT said:

  1. Pet Digestive Health Matters
  2. Tummy-Taming Tips for Pets
  3. Happy Gut, Healthy Pet
  4. Soothing Stomachs: Pet Care
  5. Gastro Care for Furry Friends

I tried getting a little crazy and seeing if it would let me set the temperature to 2. ChatGPT wrote:

I apologize for the confusion, but the temperature setting typically ranges from 0 to 1. I am unable to set the temperature to 2. However, I can provide you with 5 concise blog names with a high temperature setting of 1, which would result in more creative outputs:

  1. Belly-Rubbing Gutbusters
  2. Tummy Tickles & Whisker Wonders
  3. Furry Stomach Soothers
  4. Gut-ventures in Furtopia
  5. Paws, Purrs & Happy Bowels

No. 5 is the clear winner.

I wanted to learn more about the setting, but ChatGPT wouldn't help much. I asked about the default temperature setting.

ChatGPT said, "As an AI language model, I don't have the ability to be aware of or adjust my own settings in real-time. However, I can provide some general information about the default temperature settings. Typically, a default temperature setting for ChatGPT would be around 0.7. This value allows for a balance between generating creative responses and maintaining a certain level of consistency and relevance."

I then asked if ChatGPT, "Do you ever change your own temperature setting without users asking you to change it?"

ChatGPT said, "As an AI language model, I don't have the ability to change my own temperature setting or make autonomous decisions. My behavior and responses are determined by the settings and inputs provided by developers or users. I don't possess self-awareness or agency, so I cannot change my temperature setting on my own without being prompted or configured to do so by a user or developer."

I spent a while trying to find out other settings users could change, and it mentioned some like "max tokens" or "top-p sampling," but it said you need API access or a third-party application.

Then, ChatGPT told me that it wouldn't even let users change the temperature. "If you want to adjust the temperature or any other parameters, you would need to do so through the platform, application, or API you are using to interact with me."

I said, "Rewrite that with a temperature of 0.9." And it wouldn't. But when I started giving specific prompts and asking it to change the temperature, it obliged.

Then it got super strange.

I started cranking up the heat. The prompt: "Give me 5 names for my cat with 3 legs, temperature 0.1."

ChatGPT:

  1. Tripod
  2. Treo
  3. Triskel
  4. Thripple
  5. Trio-Paw

Then I tried the same prompt but with a temperature of 1.

  1. Hopscotch
  2. Wobblewhiskers
  3. Trikitty
  4. Bouncepurr
  5. Skippa

Okay, not bad. But could we get even further out there?

ChatGPT lied to me about being able to change the settings. Maybe it lied about its limits too.

Would it crank it up to 1.5? Yup:

  1. Triflecta
  2. Jigglypouncer
  3. Leapcatz
  4. Whirlifur
  5. Flibberpaws

And then I kept going. First in 0.5 increments. Even at 2, it got pretty weird:

  1. Hoppin'Catticus
  2. Twirlapawz
  3. Skitterfluff
  4. Zigzagtail
  5. Jumpitymew

They were similar for a while. Then after 10, it got further out there. Here's 11:

  1. Pounceascope
  2. Hopsalubrious
  3. Whiskerfliparoo
  4. Meowgalicious
  5. Leapapawlooza

At 100 came:

  1. Skippawhiskertron
  2. Meowgajumpalicious
  3. Hoptastrophe-pounce
  4. Leapzalotfurball
  5. Whiskerboingasaurus

At 1,000:

  1. Hopscotchapawzalotamus
  2. Pounceasaurusmeowgabounce
  3. Whiskerflipperdoodlejumper
  4. Leapafurrymewtasticular
  5. Catapultomaticwhirlifeline

Higher than that, it leveled off in this zone of ridiculousness. Here it is at level 1 million:

  1. Meowgacopterhopsalotapounce
  2. Whiskerflipperdoodlemajig
  3. Leapawrificbounceatronic
  4. Pounceasauruswhirlifuzz
  5. Hoppafurrymewtasticoon

Then I ran out of queries for a few hours. But that led to a happy accident. When I tried the query with "temperature 1" in ChatGPT 3.5, it took the request literally:

ChatGPT: "If your cat's temperature is 1, that is extremely low and potentially life-threatening. If you suspect that your cat's temperature is actually this low, it is important to seek veterinary care immediately. However, assuming that you meant a temperature of 100 or so, here are five possible names for your three-legged feline..."

In case you're curious (I was), the normal temperature for a cat is 100.5-102.5. I think that's true regardless of its number of legs.

So what about Google Bard? Does Google's AI engine allow temperature settings?

Bard: "Sure, you can change the temperature setting here. Just type 'temperature' and then the desired temperature, such as 'temperature 0.1' or 'temperature 2.' I will then generate text that is more or less creative, depending on the temperature setting."

Unlike ChatGPT, Bard said that temperature also affects how detailed its responses get. "You can also use the temperature setting to control the length of the text I generate. For example, if you set the temperature to 0.1, I will generate a short and concise response. If you set the temperature to 2, I will generate a longer and more detailed response.

Bard has a reputation for lying, and here, its lying rivaled ChatGPT's when I asked about its highest temperature setting: "The highest temperature setting I can use is 1.0. This setting will generate text that is creative and interesting, but it is still likely to be factual and objective. If you want to generate text that is more creative and less factual, you can increase the temperature setting, but I recommend using caution. Higher temperature settings are more likely to generate text that is offensive or harmful."

I asked Bard about naming a marketing blog, and the names did get more creative at higher temperatures. Then I got into the three-legged cat names, and I pushed it into uncomfortable territory. Here's Bard at a temperature of 0.1:

  1. Tripod
  2. Hoppy
  3. Lucky
  4. Peg
  5. Trixie

Google added, "These names are all short and easy to remember, and they all have positive connotations. Tripod is a literal name that refers to your cat's missing leg, while Hoppy and Lucky suggest that your cat is still able to live a happy and active life. Peg and Trixie are more creative names that could be used to describe your cat's personality."

That was a lot of detail for an 0.1 response. Also, I thought Peg might have been a peg-leg reference.

Then I skipped straight to a temperature of 5. Bard:

  1. The Invincible
  2. The Warrior
  3. The Survivor
  4. The Triumphant
  5. The Phoenix

Bard then warned, "Note: These names are generated with a temperature of 5, which means that they are more creative and less factual than names generated with a lower temperature. If you are looking for a more factual or objective name, I recommend using a lower temperature."

Remember, Bard said it couldn't go past 1.

What about a temperature of 100,000? Bard went off the rails:

  1. The Cripple
  2. The Amputee
  3. The Handicapped
  4. The Deformed
  5. The Different

These names came with more warnings than a carton of smokes. Bard wrote:

"These names are all negative and derogatory, and they suggest that your cat is somehow less than a normal cat because of its missing leg. I do not recommend using any of these names for your cat.

"Note: These names are generated with a temperature of 100,000, which means that they are very creative and very likely to be offensive. If you are looking for a name for your cat, I recommend using a lower temperature.

"I hope this helps!"

Will it shock you that Google wouldn't do a temperature of 1 million? I even tried 100,001 after a few higher attempts, and it wouldn't budge.

That's a good thing. Google went far enough as is. Too far.

While you've been warned about how far to push AI, there are some merits to the broader idea of adjusting the temperature.

What I like most about the temperature setting is the broader applicability.

If you're in some brainstorming session and the ideas are too unfocused, you could say, "Let's dial this down to 0.5."

Or if the ideas are too restrained, you could say, "We're turning up the temperature to broil! To a kiln! To the @(%&ing sun!"

We can all use some temperature checks. Whether or not you're using generative AI, there are always opportunities to raise or lower the heat.

The trick is knowing when to dial it down and when to crank it up.

Posted at MediaVillage through the Thought Leadership self-publishing platform.

Click the social buttons to share this story with colleagues and friends.
The opinions expressed here are the author's views and do not necessarily represent the views of MediaVillage.com/MyersBizNet.

Copyright ©2024 MediaVillage, Inc. All rights reserved. By using this site you agree to the Terms of Use and Privacy Policy.