In today’s episode, you’ll understand why generative AI still requires scaffolding and human guidance for large-scale projects. You’ll learn about the limitations of training data and token output that impact AI’s ability to produce complex and lengthy works. Additionally, you’ll gain insights on the importance of human intervention in providing outlines, breaking down tasks into manageable chunks, and guiding AI through the creative process.
Can’t see anything? Watch it on YouTube here.
Listen to the audio here:
- Take my new Generative AI course!
- Got a question for You Ask, I’ll Answer? Submit it here!
- Subscribe to my weekly newsletter for more useful marketing tips.
- Subscribe to Inbox Insights, the Trust Insights newsletter for weekly fresh takes and data.
- Find older episodes of You Ask, I Answer on my YouTube channel.
- Need help with your company’s data and analytics? Let me know!
- Join my free Slack group for marketers interested in analytics!
Machine-Generated Transcript
What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for watching the video.
Christopher Penn: Let’s talk about scaffolding.
What are we talking about? Scaffolding is making outlines.
One legitimate concern people have with generative AI is that it’s just going to start doing everything and take away all the jobs.
That concern—for now, with an emphasis on “for now”—that concern is bigger than it is; it’s worse than reality.
Generative AI does really well with short stuff, right? Write a blog post, write a summary, write an email, write a grocery list—it’s got those down.
Big stuff, heavy stuff, bulky stuff, less so.
Why? Well, it’s because of examples.
Generative AI out of the box can’t produce big works—well, it can produce them, but it requires a lot of handholding.
The other day, on a whim, I was talking with my friend, Brooke Sellas, who runs B Squared Media, about how Google Gemini helped write a comprehensive outline for a social customer care book.
What isn’t obvious when you look at the outline and the chapters is that it required a ton of co-piloting on my part, a half dozen priming prompts, a writing framework, and a writing personality just to start the outline.
Left to its own devices, Gemini produced an okay, but not great, outline and okay, but not great, writing.
Why?
There are two reasons: training data and tokens.
There are, as of April this year, I think, 38 million books in the Library of Congress.
There are, according to WordStream, an estimated 95 million Instagram posts per day, right? So Library of Congress—38 million books, period.
Two-sentence social media captions—95 million of those a day.
Just by virtue of available knowledge, any large language model, any language that you’re familiar with, right, any generative AI system will have had less big stuff to train on and learn from, and so it’s naturally going to be better at shorter tasks.
We just have more data for that.
The second is token output.
Most models have limits about how many tokens—which are parts of words, three to five character parts of words—at any time.
It’s hidden in the consumer interface.
If you’re in ChatGPT, it’ll hide it, but if you’re in the developer interfaces, it will tell you, like, “Yeah, you’re limited to this number.” Most models will not let you generate more than a few 1000 at a time, anywhere from 700 to 1500 words.
You have experienced this with ChatGPT or the many other systems—to write 800 words or 1200 words, it never gets past a certain point; it always stops.
Encoding, the same thing applies.
These tools, for now, have to have everything broken down into steps and chunks to do it well.
That means a lot of scaffolding, a lot of outlining, a lot of handholding.
Handing models your big ideas isn’t great yet.
So you need to be the provider of the big ideas and the hand-holder for the execution.
In the social customer care book, I had it write an outline of the 10 chapters, and then I had to have it write an outline for each chapter, and then I fed it each chapter’s components piece by piece, so that for any given chapter, there were maybe eight subsections, and each subsection would be 500 words.
Now you’ve got a chapter that’s 4000 words; put that together with 10 chapters, and you’ve got a 40,000-word book, but you have pieces because generative AI can’t do big stuff.
What does this mean? It means you’re still needed; you’re still necessary for anything big in size or scope.
These machines right now can’t handle the big stuff; they need it fed to them.
And they’re very good at it once you do this food feeding.
So, if you’re really good at spoon-feeding, right, you can get these tools to do some pretty incredible stuff today.
But their ability to build coherent, big works is still, it’s still not within easy reach.
It’s still not just one push of a button, and it goes and does its thing.
That is the promise of technologies like agents and things like that, but we’re not there yet.
What defines your success in the era of generative AI is the quality and quantity of your ideas; whoever has the most, best, biggest ideas wins because the tools can fill in one spoonful at a time.
But you have to be the one who has the big idea.
That’s going to do it for today’s episode.
Thanks for tuning in.
We’ll talk to you soon.
If you enjoyed this video, please hit the like button.
Subscribe to my channel if you haven’t already, and if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.
You might also enjoy:
- Mind Readings: You Need Passwords for Life in the Age of Generative AI Fraud
- Almost Timely News, January 14, 2024: The Future of Generative AI is Open
- Almost Timely News, January 7, 2024: Should You Buy a Custom GPT?
- You Ask, I Answer: AI Music Collaborations and Copyright?
- You Ask, I Answer: Legality of Works in Custom GPTs?
Want to read more like this from Christopher Penn? Get updates here:
 Take my Generative AI for Marketers course! |
Leave a Reply