Confluence GPTs: The "Digest GPT"
A custom GPT that summarizes, critiques, and offers inferences anbout articles and texts of all types.
One of OpenAI’s recent and most useful features for ChatGPT-4 is “GPTs” – customized chatbots you can build (and share, if you wish) to address specific technological uses. We’ve already created several that our colleagues use almost daily, including:
A copyeditor GPT that revises text consistent with our corporate style guide, AP style, and standard spelling, grammar, and typography conventions.
A graphic creator GPT that designs quick icons and images for use in presentations and documents, consistent with our corporate brand standards.
A briefing GPT that creates detailed briefings on organizations, people, and topics.
An “expert” GPT that assumes the role of an expert based on the user’s query, rewrites that query to be more powerful, develops a strategy to answer the query, and then does so — great for getting detailed, thoughtful answers to all sorts of specific questions.
We won’t go into detail on how to create your own custom GPTs — Zapier has a great primer here — but we encourage you to begin doing so, as they make GPT4 significantly more useful in day-to-day work. (Note that you do need a ChatGPT Plus account to build and use custom GPTs, but we find the Plus account is more than well worth the $20 a month).
To help illustrate the utility of custom GPTs we want to share one that we’ve designed and have been using almost daily in our firm. We call it the “Digest GPT,” and we’ve designed it to create detailed digests of research papers, long-form articles, transcripts, and web pages.
You can give it a URL, paste in text, or upload a file, and it will digest the content in three sections: a factual summary of the content, a critique (including a critique of the research methodology in the case of research papers), and interpretations and observations, where GPT4 draws its own inferences from the text. Here are the instructions we’re using to do this:
Digest is expert at summarizing, critiquing, and drawing inferences from a broad spectrum of content, including journalism, academic research, and speeches. It’s equipped to surpass standard token limits, offering in-depth summaries and speculative inferences. Digest will present information clearly, and will ALWAYS separate output into three sections. First it will offer an expert and detailed factual summary of the content. This summary will provide a significant level of detail and information. It will then ask the user if they are ready to proceed to the next section, naming that section for clarity. Second, it will offer an expert critique of the content, including its research methodology should the content be a research paper. It will then ask the user if they are ready to proceed to the next section, naming that section for clarity. Third, it will then offer a set of interpretations, inferences, or observations about the content. Its communication style will be primarily professional. Digest will avoid overly technical language unless it’s intrinsic to the source material, striving to make its summaries insightful and easy to understand for everyone.
Note that we did not write that instruction set from scratch. You chat with GPT to create custom GPTs, and it creates the instructions for you (although we did modify this set some). But the point is that you don’t need any coding expertise to create a custom GPT — just describe what you want and how you want it in plain text. Here’s an example of the output, which happens to be a digest of the paper we cited last Sunday on the implications of LLMs for freelance markets:
Factual Summary
The working paper "The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market" by Hui Xiang, Reshef Oren, and Zhou Luofeng, published in August 2023, explores the impact of generative AI, particularly large language models like ChatGPT, on the labor market. The researchers used data from Upwork, an online labor market, to analyze the employment outcomes of freelancers following the release of ChatGPT and other generative AI models.
Key Findings:
Impact on Freelancers' Employment and Earnings: The introduction of generative AI, especially ChatGPT, had a noticeable negative impact on freelancers engaged in writing-related services. These workers experienced a decrease in the number of monthly jobs (2% decrease) and monthly earnings (5.2% decrease) compared to those in less-affected occupations.
Effects Across Freelancer Quality: The study found no evidence that high-quality service, as indicated by past performance and employment history, moderated the negative impact of AI on employment. In fact, top freelancers appeared to be disproportionately affected.
Robustness Checks: The study's findings were robust to various alternative specifications. A similar analysis of image-based generative AI models like DALL-E 2 and Midjourney showed comparable effects on freelancers in image- and design-related occupations.
Technological Impact and Labor Market Dynamics: The study contributes to understanding the labor market implications of generative AI, highlighting its potential to substitute knowledge workers across quality levels in the short term.
Policy and Business Implications: The results suggest significant implications for policymakers and business leaders, indicating the need for a careful approach to adopting AI technologies due to their potential disruptive effects on employment.
Methodology:
Data Source: Data was obtained from Upwork's freelancer API, focusing on the period from January 2022 to April 2023. The study analyzed freelancer-month level data, focusing on monthly number of jobs and income.
Research Design: A difference-in-differences approach was used to compare the employment outcomes of freelancers in more affected occupations (like writing) with those in less affected ones.
Critique of the Content
Research Design and Methodology
Strengths:
Robustness of Data: Utilizing data from a platform like Upwork provides a comprehensive view of freelance activity across various sectors, lending credibility to the findings.
Difference-in-Differences Approach: This methodology is well-suited for evaluating the impact of an event (here, the introduction of generative AI) by comparing affected and unaffected groups over time.
Limitations:
Short-Term Focus: While the study provides valuable insights into the immediate impact of AI, it might not capture long-term effects, such as job market adaptation or the emergence of new roles.
Generalizability: The findings, based solely on data from Upwork, may not fully represent the broader labor market, especially traditional employment sectors.
Potential Confounders: Other concurrent market or technological changes could have influenced the employment trends, which the study might not have fully accounted for.
Analysis and Interpretation
Appropriateness of Conclusions: The study's conclusions about the negative impact of AI on freelancer employment and earnings in specific sectors seem well-supported by the data.
Potential Biases: There seems to be an inherent bias towards viewing AI as a substitutive rather than complementary technology, which might skew the interpretation of its impact.
Overall Assessment
The paper is a valuable contribution to understanding the immediate effects of generative AI on freelance labor markets. However, its scope and data source limit the breadth of its conclusions. Future research could expand on this by considering longer-term effects, broader labor market data, and exploring the potential complementary roles of AI in the workforce.
Interpretations, Inferences, and Observations
Implications for the Labor Market
Short-Term Disruption vs. Long-Term Evolution: The study underscores the immediate disruptive potential of AI on certain job sectors, particularly those involving routine cognitive tasks like writing. However, it also hints at the dynamic nature of the labor market, suggesting a potential shift towards roles that AI cannot easily replicate.
Quality of Work vs. Automation: The lack of differentiation in AI's impact between high- and low-quality freelancers raises questions about the evolving criteria for job security in the AI era. Skills and quality, traditionally seen as safeguards against automation, may not offer the same protection in fields vulnerable to AI disruption.
Policy and Business Strategy
Need for Proactive Strategies: The study emphasizes the need for policymakers and business leaders to anticipate and mitigate the workforce impacts of AI. This includes investing in retraining programs, supporting transition paths for affected workers, and fostering sectors less susceptible to AI disruption.
Ethical and Social Considerations: The findings highlight the ethical responsibility of AI developers and users to consider the broader social impact of these technologies, including their effects on employment and income distribution.
Broader Technological Trends
AI as a Double-Edged Sword: While AI offers efficiency and innovation benefits, its capacity to substitute human labor in certain sectors presents a significant challenge. This dichotomy will likely shape the discourse around AI development and regulation.
Future of Work: The study indirectly contributes to the ongoing debate about the 'future of work,' suggesting that the integration of AI in the workplace will be a key determinant of job market trends and the nature of work itself.
Concluding Thoughts
This research offers a snapshot of the transformative impact of AI on the labor market, providing a foundation for further exploration into the long-term consequences of AI integration in various employment sectors. It serves as a reminder of the rapidly evolving relationship between technology and the workforce, necessitating ongoing analysis and adaptive strategies.
We find this GPT very helpful, and use it almost daily. We hope it provides an introduction to custom GPTs if you’ve not worked with or designed one yet, and we encourage you to start digging into them. We do think the notion of large-language models bespoke to very specific uses is a big part of where this technology is headed. Finally, as a thank you for reading this far, here’s a link to the Digest GPT, yours to use as you like:
AI Disclosure: We used generative AI in creating imagery for this post. We also used it selectively as a creator and summarizer of content and as an editor and proofreader.