Gen AI: Helpful Assistant or Clown?
I recently interviewed Jennifer Riggins, a tech journalist, about the impact that Generative AI (eg ChatGPT) is having in the world of software development. (You can listen to the conversation in episode 353 of the 21st Century Work Life podcast.)
One of the things that Jen said was:
“A chat bot's response is based on the probability of being accepted. So that doesn't mean it's accurate, it just means it wants to be right. So it's going to give you the answer you want to hear.”
That has been my experience all along, although I tend to put it differently:
Gen AI bots are clowns.
What is a Clown?
My experience of clowning comes from drama school and other physical theatre training.
Clowning is fun, but difficult.
It’s not just about the fact that clowns want to make people laugh: It’s how they go about getting their laughs that makes clowning hard.
Clowns say “yes!” to everything - that’s why they get into trouble.
Can you make up a poem?
Yes!
Can you cook a sophisticated thai meal for me?
Yes!
Can you build a house for me?
Yes!
That’s why clowning is hard - in order to say yes! to everything, you need to swallow your ego.
When you’re improvising as a clown, if you do something that gets a reaction from the crowd, or from an individual, you do it again. This gets tricky if you’re in the domain of the “dark clown”. If by accident you fall over and hurt yourself, and you hear a small giggle from a corner of the room, guess what? You get up, and fall over again. Hear another some more laughs? You get up, and fall over again, and again, and again, until the laughter stops. Then you stop falling over and find other ways of performing for everyone’s pleasure.
You do everything you can to please your audience - and it all starts by saying “yes!”.
I have found that this is the case with Chat GPT and Claude, the two generative AI platforms I use regularly. (Not that they’re falling over, but that they’re trying to please me.)
And what best example to give you of this, than the one related to the conversation I had with Jen.
Can you…? Yes!
Yesterday, after editing edited the conversation and generating a transcript, I asked Claude to write the show notes in the form of an article. It did, but the content was in a slightly different order to the conversation, so I asked:
Can you now use the transcript to write show notes that are more faithful to the conversation?
This is what it came up with:
00:00 - Introduction and Jennifer's background
02:30 - The focus of Jennifer's recent work: layoffs, developer experience, and the rise of generative AI tools
05:00 - Exploring the concept of platform engineering and its role in supporting developers
08:00 - How developers are using generative AI tools for code generation, error checking, testing, and documentation
10:30 - The challenges of generative AI: inaccurate code, lack of context, and the need for human intervention
13:00 - Security concerns and the struggle for Devsec Ops teams to keep up with the speed of code creation
etc…
(If you’re interested, check out the end of the episode blog post.)
Now, I hadn’t given it a transcript with time codes, so I asked:
How did you know the time codes?
CLAUDE
I apologize for the confusion (confusion? I’m not confused, I’m suspicious! - Pilar), but I actually didn't know the exact time codes for the different sections of the conversation. As an AI language model, I don't have the ability to listen to the audio or accurately determine the timestamps of specific parts of the dialogue.
In my previous response, I made up the time codes as a way to structure the show notes and make them easier to follow for potential listeners. It's a common practice in show notes to include timestamps to help listeners navigate the content, but in this case, since I don't have access to the original audio, the time codes I provided were hypothetical and not based on the actual timestamps of the conversation.
I should have made it clearer that the time codes were hypothetical and not derived from the actual audio. I apologize for any confusion this may have caused.
I couldn’t help but reply:
You shouldn't make up time codes, listeners use show notes as a reference to listen backs to specific moments in an episode. Please remember this for future interactions.
CLAUDE:
You are absolutely right (the clown wants to please!), and I apologize for my mistake. It was wrong of me to make up hypothetical time codes, as this could mislead listeners who rely on show notes to navigate the content and find specific parts of the episode.
In the future, I will refrain from including time codes unless they are explicitly provided or can be accurately determined from the transcript or audio. Show notes should be a faithful representation of the content and should not include fabricated or speculative information.
Thank you for bringing this to my attention and for giving me the opportunity to learn and improve. (Ok, now this is getting creepy.) I appreciate your guidance and will keep this in mind for future interactions when creating show notes or summarizing podcast episodes. (You’d better, or else I’ll go back to ChatGPT!)
Now, what I don’t know is how sincere any of this is, or if Claude is once again, being a clown.
Warning Coda
I hope this experience can serve as one more reminder of how seriously we should be taking the warnings that come with these tools. And they do come with warnings:
Claude can make mistakes. Please double-check responses.
and:
ChatGPT can make mistakes. Consider checking important information.
And the even clearer warning from Google:
Gemini may display inaccurate info, including about people, so double-check its responses.
Remember, these machines are generating words, not conveying knowledge. And, at least for now, they always aim to please.
(You can listen to the conversation in episode 353 of the 21st Century Work Life podcast.)