Expanding Personal Reading and Connecting with Others through Glasp and Social Web Highlighting

Expanding Personal Reading and Connecting with Others through Glasp and Social Web Highlighting

In the age of the internet, we are constantly bombarded with an overwhelming amount of information. It can be challenging to filter through the noise and find content that truly aligns with our interests. However, there is a unique opportunity presented by Glasp, an online service that allows users to highlight and annotate the social web.

Glasp enables users to locate and connect with content that others have identified as similar to their own interests. By processing the content highlighted and the notes added by other users, individuals can expand their reading and viewing experiences. This not only opens up new avenues for exploration but also fosters connections with like-minded individuals who share similar passions.

On a similar note, OpenAI has introduced a fascinating feature that allows users to fine-tune GPT-3 on data they provide. One exemplary use case is creating a blog post title optimizer using Hacker News data. Hacker News data is particularly useful due to its community validation and the wide variety of idiosyncratic styles in submission titles. Additionally, obtaining Hacker News submission data in bulk from BigQuery is a relatively straightforward process.

However, one challenge in utilizing this data is the imbalance between good and bad titles. With 63k titles that have fewer than 100 points, the dataset is skewed, leading to flawed training results. To address this issue, there are two potential solutions. One approach is to repeat the good posts to roughly equal the number of bad posts. Alternatively, a subset of bad posts can be selected to achieve a similar balance.

Once we have a tool to determine the quality of blog post titles, the next step is generating alternate titles that maintain the same meaning. Surprisingly, GPT-3's latest Instruct model can accomplish this task effectively. Despite the concise input title and the recency of DALL-E 2, InstructGPT can infer that the AI creates something and work from that, showcasing its impressive capabilities.

To optimize a technical blog post title, the process involves selecting the desired title and retrieving up to six alternate titles from InstructGPT. These generated titles are then extracted and cleaned up for further analysis. Each alternate title is evaluated by pinging the finetuned Hacker News GPT-3 model to determine the probability of it being a good title. Sorting the titles by probability in a visually appealing table allows for a quick assessment of their quality.

The results are promising, with most of the alternate titles surpassing the original in terms of predicted probabilities of being a good post, often exceeding 50%. This discovery prompts the consideration of retroactively changing the title for improved SEO, although it may come with a slight dishonor.

In conclusion, the combination of Glasp and social web highlighting offers a unique opportunity to expand personal reading and connect with others who share similar interests. Additionally, the use of GPT-3 and Hacker News data allows for the creation of a blog post title optimizer, enabling users to generate alternate titles that maintain the same meaning while potentially improving their SEO rankings.

Actionable Advice:

1. Explore Glasp and embrace the power of social web highlighting to discover new content and connect with individuals who share your interests.
2. Consider utilizing the finetuning capabilities of GPT-3 on data you provide, such as Hacker News data, to create innovative tools or optimize existing processes.
3. Experiment with generating alternate titles for your blog posts using InstructGPT and the finetuned Hacker News GPT-3 model to improve the quality and discoverability of your content.