Back to Main
This Week in AI — 📝 How to get AI to write longer and better
AI

This Week in AI — 📝 How to get AI to write longer and better

December 12, 2023 | By: Dave Isaac

Headlines You Should Know

 

Google’s Gemini Launch … Could Have Gone Better

Google’s chatbot, Bard, had been running on a large language model called LaMDA that was released in 2021. Last week Google released Gemini, a new LLM that now powers Bard and one the company hoped would overtake OpenAI’s GPT-4.

The Gemini demo video was deemed “fake” by news outlets because it wasn’t an accurate depiction of what the model can do in real-time. TechCrunch said, “In actuality, it was a series of carefully tuned text prompts with still images, clearly selected and shortened to misrepresent what the interaction is actually like.”

As for comparisons to ChatGPT, Business Insider’s Aaron Mok took the new Bard for a test drive and found that Gemini had better access to current information but is still flawed. Gemini refused to answer some of the questions, generated factual errors, and Mok said he preferred ChatGPT’s writing style.

It may be worth giving Gemini a spin for yourself, but realize that its specialty seems to be current events. This is likely a better research tool than a content-generation tool — just be sure to double-check the facts at the source.

EU Wins Race to Regulation with AI Act

The European Union beat Congress to the punch with the world’s first AI legislation. Starting in 2025, tech companies will be legally obliged to tell consumers when they’re interacting with AI chatbots, label AI-generated content, and design their systems in a way that AI-generated output can be detected.

Pundits recognize the law isn’t perfect — a stricter set of rules applies to the AI models that need more computing power to train them, and only the developers would know how to quantify that. A European Commission official couldn’t even confirm if GPT-4 or Gemini make the cut. Here in the U.S., Congress is likely to hold more AI forums, after briefing nine times since September.

Elsewhere …

Tips and Tricks

✍️ Getting AI to write longer and better

What’s happening: A common complaint about AI’s ability to generate content is that outputs are light on both word count and substance. Asking chatbots to produce a certain amount of words almost never works out, and some AI tools (looking at you, ChatGPT) fill in text with empty phrases that seem like they belong but don’t hold much weight.

A new strategy: Instead of simple prompts with brief summaries of the topic you want AI to expand upon, flood it with pertinent information. Last week we went over the outline-to-draft method, which is a good first step in getting better outputs because seeing a framework at an early stage gives you more control. Fleshing each part out individually also forces the AI to focus on a particular section, which can improve the quality and lead to a longer piece when you tie all those sections together.

Try this: Flood the AI with as much solid background information as you have. This is helpful context for the chatbot, but you probably still need to point it in the right direction. Wharton Professor Ethan Mollick highlighted a particularly simple way to do this: Directly tell the AI “Here is the most relevant sentence in the context.”

We’ve also found success with similar approaches, like directing the chatbot to a certain time range in a transcript with timestamps or pointing out particular statistics to weave into the copy. If the chatbot doesn’t pick up on this the first time, repeat the request throughout your prompts for extra reminders — sometimes that second or third time is key.

Quote of the Week

“Whatever the victories may have been in these final negotiations, the fact remains that huge flaws will remain in this final text.”

— Daniel Leufer, Senior Policy Analyst at the digital rights group Access Now to the AP on the EU AI Act