Curation to Creation with AI

If everyone gets access to the same AI system, how will we differentiate the content from other people? Also, where will your identity be reflected in the writing? Maybe in the prompt?

Oct 20, 2022

We see many AI-powered writing tools recently. You type some prompts and AI will generate some paragraphs or bullet points for you. In the near future, AI-powered writing will become the norm, with writers using AI to help them research, plan, and write their articles and books. These are enabled by some amazing technologies such as OpenAI’s GPT-3, etc.

However, if everyone gets access to the same AI system, how will we differentiate the content from other people? Also, where will your identity be reflected in the writing? Maybe in the prompt?

As what you collect says so much about who you are, if the AI system can use your curated information and understand who you are, the system can reflect your taste/identity in the writing. In this way, we could find a better way to harmonize with AI. That’s why we always focus on “Curation to Creation”.

Meet Personalized AI-powered Writing

At Glasp, we are building a social web highlighter for learners, readers, and writers. We want to give people the power to leave their learnings, keep track of them, and access other people’s insights easily so that we can circulate knowledge.

The curated information must reflect who you are because it’s a list of what you resonate with. By curating, we are projecting our own identity in itself.

With the curated information, I believe that we could build a better AI-powered writing system that reflects your taste and personality well so that it’s your personalized AI writing system.

That’s why we are building and testing out the AI writing system with Glasp. This is still private but we see really exciting results so far!

The Opportunity and Challenge of The Current AI Models

We fine-tune the AI system by using one’s curated information, and it could cover one of the weaknesses of the current AI models: misinformation due to the old dataset. Sometimes, AI-generated content is wrong or inaccurate because the dataset used to train the model is simply old or random.

In the case of GPT-3 from OpenAI, the model is trained to predict the next word on a large dataset of Internet text, rather than to safely perform the language task that the user wants. In other words, these models aren’t aligned with their users. To make the models safer, more helpful, and more aligned, they use an existing technique called reinforcement learning from human feedback (RLHF).

Some of our previous research in this direction found that we can reduce harmful outputs by fine-tuning on a small curated dataset of human demonstrations.

Yes, fine-tuning with the curated dataset of human demonstrations helps the AI models a lot! I believe that Glasp can be the one as we are a group of people who love curating and circulating knowledge for humanity.

As you can see from our demo video of “AI-powered Writing with Glasp”, we see a unique advantage and potential in curation-based AI writing!