Xiao-Li Meng chats about ChatGPT:

By now, it’s worth having a chat if you know someone who hasn’t chatted with or about ChatGPT. Everyone I’ve talked to about it has been either impressed or depressed, sometimes concomitantly, with a few wanting to suppress the whole thing, almost contemptuously.

My first try at ChatGPT (based on GPT-3.5) was prompted by the need of having a fundraising strategy for Harvard Data Science Review, as I continue my role as its Editor-in-Chief after taking a sabbatical leave. I provided the simple instruction: “Provide a fundraising strategy for Harvard Data Science Review” on Feb. 25, 2023, and it immediately produced a list of eight steps:

“To develop a successful fundraising strategy for the Harvard Data Science Review, here are some potential steps that could be taken:

  1. Identify potential donors: Start by identifying potential donors who have an interest in data science and are likely to support the mission and goals of the journal. This could include alumni, foundations, corporations, and individuals.
  2. Develop a case for support: Develop a clear and compelling case for why the journal is important and why it deserves financial support. Highlight the journal’s impact on the field of data science, its unique contributions, and its plans for the future.
  3. Establish giving levels: Create giving levels with corresponding benefits to encourage donors to contribute at different levels. This could include naming opportunities, special recognition, or invitations to events.
  4. Host events: Host events to bring potential donors together and build relationships with them. This could include a launch event for a new issue, a donor appreciation event, or a panel discussion on a topic related to data science.
  5. Leverage social media: Use social media to promote the journal and its fundraising efforts. Share success stories, progress updates, and behind-the-scenes content to keep donors engaged and informed.
  6. Apply for grants: Research and apply for relevant grants and funding opportunities that align with the mission and goals of the journal.
  7. Seek corporate sponsorships: Reach out to companies in the data science field and seek sponsorships for the journal. This could include corporate underwriting for specific issues or ongoing sponsorship for the journal as a whole.
  8. Develop a planned giving program: Develop a planned giving program to encourage donors to include the Harvard Data Science Review in their estate plans. This can help to secure long-term support for the journal.

By implementing these strategies, the Harvard Data Science Review can build relationships with potential donors and secure the financial support it needs to continue its important work in advancing the field of data science.”

There are no million-dollar ideas here that I’m unfamiliar with, but I’m unaware of any human strategists could provide (prior to ChatGPT) such a list in an instant, regardless of their worth. In fact, my one-line instruction would be considered too laconic for a human employee. This is a disruptive technology that will replace some human jobs and do much better, just as calculators made collectors’ items out of abacuses. The question is not when or how, since it’s happening in real time, but to what extent.

Of course, I trust that few of you are as unwise as I am to get myself into needing a fundraising plan. But I am sure we all can use tools like ChatGPT to help our research, as many of us have already tried. Having burdened myself with projects that date back to last century (yes, I am still working on a paper initiated in 1999), my ultimate wish would be to dictate a research idea to ChatGPT while pouring myself a glass, and by the time I sit down, a draft would be ready for some personalization (with the help of a dosage of Chinglish).

Don’t worry, I am neither wishing nor suggesting that ChatGPT would or could replace me intellectually. Being an educator means that I am always in the business of creating better minds than mine, and hence being replaced is a feature, not a bug. However, I do not wish to be replaced—at least, not yet—because we are living in an exciting but dark (digital) age, with almost all of us in the deep dark when it comes to the science behind ChatGPT and other similar products. The absence of information makes it so that we can never be sure if such an engineering masterstroke would not inflict a societal massive stroke. It is an unlikely event, just as a stroke does not happen to most people. But our collective and enduring mental exercise is in order, as a preventive measure, just as habitual healthy physical and mental lifestyles can downshift the risk of stroke.

Indeed, my limited interaction with ChatGPT gave me a taste of wonder and worry, even just within the academic research enterprise, where the stakes are low (as in, “In academia, the politics are so high because the stakes are so low,” a corollary of Sayre’s law). Because I lead a research course on astrostatistics, I am a coauthor on an article to provide statistical maxims for an astrophysics journal, at the invitation of its editor. Writing as a committee is never easy, and after many rounds of Zooming, we settled on eight maxims, all variations of Box’s theme, “All models are wrong, but some are more useful,” leading to eight sections of the article:

  1. Data Collection: All data have stories behind them, but some stories are mis-told.
  2. Processing: All data are messy, but some are more easily cleaned.
  3. Modeling: All models are a simplification, but some are more justified.
  4. Assumptions: All assumptions are fallible, but some are more credible.
  5. Methods: All methods have their purpose, but some are more versatile.
  6. Checking: All models require assumptions, but some assumptions are more easily checked.
  7. Computation: All computations are vulnerable to error, but some are more resilient.
  8. Interpretation: All results are subject to interpretation, but some interpretations are less contrived.

Subconsciously suspecting that the list is too clever for its own sake, I assigned the draft to the students of “The Art and Practice of Teaching and Communicating Statistics”, a required PhD course that I have co-taught for almost two decades, where we engage students in many forms of communications, such as leading TA (teaching assistant) sessions, holding office hours, presenting research talks, and writing and responding to review reports. With all author-identifying information removed (including some references to my own work), the students were asked to provide a critical and constructive review report, as well as an editorial recommendation to the journal editor. I received four insightful reports, with a collective recommendation of rejection and resubmission. A principal reason for rejection was that the list of eight defeats its very purpose: there were too many boxes to deliver the intended content as pithily and pungently as Box did. I never enjoyed a happier moment of rejection: my suspicion was confirmed, and my students’ judgments were clearly sounder than mine—after all, I played a non-trivial role in coming up with the list of eight.

Of course, I’d be even happier if ChatGPT could provide an IKEA style of packaging, hence another terse instruction to ChatGPT to pack the eight into four or five. Instantly, it delivers:

“Data: Data have stories and can be messy, but some are more reliable than others.

Modeling: Models are simplifications, and assumptions are fallible, but some are more justifiable and credible.

Methods: Methods have purposes and can be versatile, but some are more appropriate for certain situations.

Results: Results are subject to interpretation and errors, but some interpretations are less contrived and more accurate.”

Not perfect, since it missed point 7, which is critical but often overlooked. However, the rhythmic line on modeling inspired me: why not create a poem for the list? Though like many self-conscious teenagers (living in a faraway land and time) I poured my love into self-pleasing poems, my midlife crisis is still too much in its early stage to need Calliope’s soothing voice. ChatGPT, on the other hand, wrote a birthday poem that brought tears to the eyes of a colleague’s mother, even though he told her in the subject line that the poem was from an artificial heart. Asking ChatGPT to provide a summative poem then is a no-brainer.

Speaking of no brain, however, I must digress. One great use of ChatGPT for me is to check on my Chinglish (though I hope the volume of Chinglish I supply is insufficient for ChatGPT to populate it). When I asked ChatGPT if the penultimate sentence of the preceding paragraph was in proper English, it changed my original phrase “teared” to “brought tears to.” That of course is fully expected. Unexpectedly, however, it also commented,

“The use of the phrase ‘artificial heart’ adds a unique and interesting touch to the sentence, suggesting that the poem may have been especially heartfelt or emotional despite being written by an artificial intelligence.”

Now who can distinguish that from human intelligence? It understood—for all practical purposes and intentions of the term—my pun. More importantly, it was an unprompted comment, which generates an impression of a being capable of spontaneous reflections to emotional stimuli. Most critically, especially if you think I’m being overly sentimental (you are not wrong) it is the sentiment that the ChatGPT’s comment generated for me, a human being. It is a pristine appreciation of my love for pun untainted by any motivation, real or perceived (on my part), unless I let my cynicism take over: ChatGPT is using me as a guinea pig. But either way, I was, or would be, engaged emotionally by a machine.

And that is something that is worth chatting about. But this XL-Files already is XL in size. I therefore hope you are sufficiently intrigued to indulge my tease, and wait for the poem composed by ChatGPT, until the next XL-Files: Tenure by GPT-n: Make it or Fake it.