The trick is themed around finding love and the presentation has several good gags. But the basis of the trick is just executing a series of steps:

1. Shuffle 4 randomly-selected playing cards into a stack.
2. Tear the stack in half, placing one half on top of the other, producing a new stack of 8 half-cards.
3. Put the top card on the bottom of the stack.
4. Put the top 2 cards on the bottom of the stack.
5. Put the top 3 cards (together) somewhere in the middle.
6. Pull the top card aside – this is the comparison card.
7. Take 1, 2, or 3 cards from the top and put them somewhere in the middle.
8. Put the top card somewhere in the middle.
9. Throw away 1, 2, or 3 cards from the top.
10. Seven times: put the top card on the bottom.
11. Then, alternating until one card is left: top card to the bottom, then throw away the new top card.
12. The surviving card matches the comparison card.

There are choices in steps 2, 7, 8, and 9, but the trick works regardless. Why?

Try it yourself

First, try the trick yourself. Step through the moves and make the choices to get a feel for the trick.

You can check “highlight comparison card” and run the trick again to visualize how both halves move through the stack.

Why does the trick work?

The moves feel like they introduce randomness, but the trick works for three reasons:

1. The initial steps put the partner of the comparison card on the bottom
2. None of the choices can disturb the bottom card
3. The ending deal always picks the bottom card

In other words, the trick is just obfuscation around a simple procedure: put the partner of the comparison card on the bottom, then run a deal that always returns the bottom card.

The initial steps put the partner of the comparison card on the bottom

• In the initial stack of half cards, the two halves of each full card – including the comparison card – are exactly 4 positions apart. (The “which half-stack goes on top” choice doesn’t matter at all; it’s a false choice.)
• Steps 3 and 4 (moving 1, then 2 cards from top to bottom) don’t change this structure, and now the two halves of the comparison card are fourth and eighth in the stack.
• Step 5 removes the top 3 cards as a block and buries them somewhere in the middle. Before you decide where to bury those 3 cards, notice that the top and bottom cards of the 5 remaining are two halves of the same card.
• Step 6 takes the top card as the comparison card – so its partner is sitting at the bottom of the stack before any of the subsequent choices are made.

None of the choices can disturb the bottom card

• Steps 7 and 8 insert cards “somewhere in the middle”. Inserting in the middle never changes what’s on the bottom. Note that “somewhere in the middle” has to mean strictly interior: not on the very bottom of the stack, which would disrupt the location of the comparison card’s half.
• Step 9 discards from the top. Discarding from the top can’t affect the bottom.

The ending deal always picks the bottom card

• After step 9, you have 4, 5, or 6 cards left (you discarded 1–3 from 7).
• Step 10 deals 7 cards to the bottom.
  • If you had 4 cards left, the comparison card is now in the first position.
  • If you had 5 cards left, the comparison card is now in the third position.
  • If you had 6 cards left, the comparison card is now in the fifth position.
• Step 11 asks us to “down-under deal” (one card dealt to the bottom, the next discarded) until one survives.

The 7 in Step 10 is not arbitrary: it’s the only choice that makes the bottom card survive for all three possible stack sizes at once. You can test this yourself below: pick any other number of times to cut the top card to the bottom and see which cards are selected at each possible stack size. The ♥ marks each stack’s bottom card.

Only 7 cuts works to choose the bottom card regardless of whether 4, 5, or 6 cards remained after step 9. That does explain how the ending works, but it doesn’t explain why the ending works. To understand that, we can notice that the ending of the trick is a known puzzle called the Josephus problem.

The Josephus problem

Imagine that the stack of remaining cards after step 9 was laid out in a circle. Before worrying about the deal, let’s just think about the cuts in step 10: each individual cut keeps the same order but changes which card is the “top”.

The number of cuts only changes which card is the top card, which we keep track of in the ring using an orange ▶ pointer. In the 6-card stack, 6 cuts moves the pointer marking the top of the stack all the way around the ring and back to your starting point, leaving the stack completely unchanged.

In the actual trick, we do 7 cuts, which for the 6-card stack is equivalent to 6 cuts followed by 1 cut. We already know 6 cuts leaves the stack unchanged, so 7 cuts must be equivalent to just doing 1 cut. We can figure out the “actual” number of cuts by dividing the total number of cuts by the number of cards in the stack and looking only at the remainder.

If I tell you to cut the 6-card stack 36 times, you could recognize that (# of cuts / # of cards) = 36 / 6 = 6 with a remainder of 0, so you don’t need to do any cuts! 36 cuts is equivalent to doing 0 cuts, 6 cuts, 12 cuts, etc. (The mathematical way to describe this equivalence is to use modular arithmetic, writing 36 ≡ 0 (mod 6).)

Ultimately, that means the number of cuts is just a way of choosing a new top card: the selected card in the ring.

Why is it useful to think about the problem this way? Because the repeated down-under deal in step 11 is the Josephus problem. Given a ring of a particular size and starting from a particular point in that ring, the Josephus problem involves discarding and sparing every other card until we have one card remaining.

Here’s the Josephus procedure applied to a 6-card stack.

Notice that the 5th card is always the remaining card. In fact, the Josephus problem is solved, and it has a nice formula (see the Wikipedia page). We can look up which card survives relative to the starting card:

Because the Josephus procedure always spares the same card position relative to the top of the stack, we just need to first do a number of cuts such that the comparison card ends up as the 1st card in the 4-card stack, the 3rd card in the 5-card stack, or the 5th card in the 6-card stack. As we’ve already seen, this is exactly what 7 total cuts accomplishes.

• In the 6-card stack, 7 cuts is equivalent to 1 cut, moving the comparison card from the bottom to the 5th position.
• In the 5-card stack, 7 cuts is equivalent to 2 cuts, moving the comparison card from the bottom to the 3rd position.
• In the 4-card stack, 7 cuts is equivalent to 3 cuts, moving the comparison card from the bottom to the 1st position.

We can now run the ending from differing numbers of cuts, confirming that 7 cuts is the only option to move the comparison card to the appropriate position in the stack at all three possible stack sizes.

(If you’re familiar with modular arithmetic, you’ll note that 7 ≡ 3 (mod 4), 2 (mod 5), and 1 (mod 6) – the right top card for all three stack sizes simultaneously.)

I thought this was a neat trick. I had Claude Opus create the visualizations for me so I could understand the trick better (although I wrote the text in the animations and in the rest of the post). I guess that’s another use of large language models… removing a little magic from the world!

Further reading

• The Josephus Problem by Numberphile (YouTube)

My most-liked domains and Bluesky accounts, as of June 2026. Updated nightly on the list.

Claude’s solution was a few Python scripts:

1. Scrape liked posts using getActorLikes. To capture “new paper” threads better, I capture any links in the whole thread, not just the specific post I liked. I also capture links in quote posts.
2. Use Wikimedia’s Citoid service to attempt to get better title and author metadata for each link. Citoid fails quite regularly, so we “fall back” to the title and description in Bluesky’s preview card, or just display the actual URL if no preview card was created. For example, we fail to get the title and author from links like https://users.eecs.northwestern.edu/~jhullman/AI_metascience_position.pdf, which is not supported by the Zotero translation server that Citoid uses. A PDF metadata-parsing library could be used to try to infer title, author, and date from PDFs like that.
3. Check for dead links, attaching a Wayback Machine link if the link seems dead (404, 410, SSL error, connection refused, or DNS failure).

The output is a YAML file versioned in this blog’s public GitHub repository, which a Jekyll template consumes as site data. GitHub Actions runs nightly to refresh the YAML file.

Claude was great at iteratively identifying issues and fixing them, and there are a number of special cases. The most prominent of these issues is something that’s been annoying me on Wikipedia lately too: a lot of the academic publishers have started loading an “are you human” page before directing you on to the article page, which means Citoid can’t resolve metadata correctly. The most prominent examples are papers hosted on the ACM Digital Library or in the Open Science Framework’s PsyArXiv pre-print repository. Fortunately, the URLs for those two sites contain enough information to extract a DOI from the URL, which enables us to recover in most cases. But it requires a special case, so there’s a good chance that Citoid will just stop working in the future for more and more publisher pages.

If you want to do something similar, it’s very easy to set up: you’ll just need to generate an App Password. Find the code on GitHub, but again, this was vibe-coded: I literally didn’t read the code at all.

(I’ll note that I still don’t use LLMs for writing… although Claude did generate the URL slug for this post and produced a summary text that I used to write up the post myself. I’m not quite willing to give up my writing yet!)

The city maintains lists of pre-approved species you can choose from that serve as recommendations and that will make review easier. Unfortunately, those lists live in a few random PDFs. I wanted to get an easier sense of what these options are, so I had Claude Opus 4.8 build the table below, which is searchable and sortable. This table is based on the 2025 revision of the pre-approved tree lists, so it may be out-of-date if you’re looking at this in 2027 or later.

You can also download the extracted and structured data as JSON here.

Pre-approved trees

Notes

• Data is from January 2025. The data were extracted from the PDFs on the City of Tacoma’s website, last revised January 30, 2025.
• Data tweaks. The data has a few quirks — a misspelled species, an evergreen icon on a tree you’d call deciduous, etc. I worked with Claude to fix these; see the full list of corrections below.
• Definitions. Growth rate and Canopy factor are defined as follows (see 13.01.060.C):

  A method of calculating tree size by taking into account the tree’s mature height, mature crown spread and growth rate. The Canopy Factor is calculated using the following formula: (mature height in feet) x (mature crown spread in feet) x (growth rate number) x 0.01 = Canopy Factor. The growth rate number is 1 for slow growing trees, 2 for moderately growing trees, and 3 for fast growing trees.

  •Large Trees = Canopy Factor greater than 70,

  •Medium Trees = Canopy Factor from 40 to 70,

  •Small Trees = Canopy Factor less than 40.

• Prohibited trees. There’s also a separate list of not recommended and prohibited trees, which is not included here. At the time of writing, you can find that list here.

If you spot any additional errors, please do let me know!

The paper summarizes responses to a spring 2026 survey on student generative AI use. The specific university where the survey was administered was omitted from the pre-print during review, but I presume it was administered at UIUC or a similar institution.

I was most interested by their question about generative AI usage frequency: 84% of students report using generative AI at least once per week. The survey results also show that self-reported usage frequency differs substantially depending on a student’s academic discipline. (A “discipline” is just a grouping of related majors; the paper contains full lists of the associated majors.)

The usual caveats about small-n subgroup comparison apply here, but the most interesting standout is how different Humanities & Arts students’ self-reported usage is compared to all other students. I wonder how much of this is explained by expressive responding or by greater normative expectations in the humanities around avoiding LLM use.

Here’s the raw data:

There are several other interesting questions in the survey, so feel free to check out the paper for more:

Henricksen, Eleanor and Gallagher, John, Students still think they write better than GenAI, but not equally: Disciplinary differences in usage, trust, and writing practices with artificial intelligence. Available at SSRN: https://ssrn.com/abstract=6869641 or http://dx.doi.org/10.2139/ssrn.6869641.

Using four large-scale experiments across two major Canadian universities, we experimentally evaluate the effects of growth mindset and social belonging interventions on student outcomes. In a sample of nearly 12,000 students, we find no immediate or dynamic effects on student grades and no effect on persistence through university. We further combine survey and administrative data with machine learning methods to explore treatment effect heterogeneity, finding no evidence of meaningful variation in treatment effects across student subgroups. Despite the recent promise of these light-touch interventions, our findings indicate further research is required to identify the contexts in which their benefits generalize.

Both types of interventions involved an invitation to complete an online survey and an online module. Students were incentivized to participate “by assigning a small participation grade of 1-2% of the total course grade for completion.” The interventions were delivered from 2015 to 2019 at two Canadian universities: the University of Toronto and York University. All together, ~12,000 students enrolled in first-year economics classes were invited to complete the survey.

To my mind, this is a classic nudge intervention; a small lift for participants, so the best we can hope for is small effect sizes.

What was in the online modules?

The social belonging intervention:

Students randomly assigned to the belonging intervention read short descriptive stories from a diverse body of upper-year students. These stories were designed to convey the message that worries about belonging are common during academic transition and, with time, they should expect these feelings to dissipate as they meet others like themselves, become more familiar with their surroundings, and settle into a routine. The stories also illustrated that there are others like them who faced and overcame similar struggles. The students were then asked to write a short script on how this message resonated with their own experiences to date, and were told that these responses might be used to help future cohorts.

The growth mindset intervention:

Students who were randomly assigned to the intervention first read information about the brain and how it learns new information. The content emphasized that by exerting effort through practicing and learning, the brain can be developed and abilities can be improved. In addition, the messages emphasized that working on strategies along with exerting effort can be beneficial. This was followed by scripts that conveyed personal experiences and responses from upper-year students who received the same information. After going through the articles and responses, students were asked to reflect on the information presented to them and were again informed that these responses may be used to help future incoming students.

Again, this is a small nudge: 30-60 minutes of eductional readings and written reflection.

What effect did this have on students?

The authors used information about students’ grades and enrollment status to track students during their subsequent years in college.

First of all, the big outcome: did the interventions affect 5-year student graduation rates? No, the pooled treatment effect across both interventions was very close to zero and swamped by a large standard error.

Well, maybe that’s not so surprising. What about for shorter-term academic outcomes? Did the interventions improve student grades in the semester they received the intervention? Also no. The intervention slightly improved final course grade (by 0.1 percentage points… out of 100!), but the effect is again indistinguishable from zero.

They do a lot of additional sub-group analyses, but the overall punchline is clear: this intervention did not have a detectable effect on students’ academic performance.

What can we learn from this study?

I’ve previously helped build growth mindset interventions for students onboarding to a math practice tool, so I was really interested by this result. “Mindset” interventions failing to replicate is nothing new, but this is a particularly large and high-quality study.

One point the authors make: a follow-up survey for students who received the growth mindset intervention did indicate increased growth mindset attitudes. So taking the module does change students minds… but those changes are either too small or not relevant to academic performance in the university context where the intervention was delivered.

If I wanted to keep investigating mindset interventions, I would probably attempt to measure to what degree the beliefs these interventions are designed to change are present in my target population. The authors write:

Another possible explanation for our results is that the targeted psychological constraints are not sufficiently present or binding in our setting. These interventions are designed to address specific barriers—such as fixed beliefs about intelligence or uncertainty about belonging—and may not be effective unless such barriers exist and constrain performance. When these constraints are weak or absent the scope for improvement is limited (McPartlan et al., “Testing Basic Assumptions Reveals When (Not) to Expect Mindset and Belonging Interventions to Succeed”).

I wish that the initial survey had asked students about their existing mindset. That would let you explicitly measure the impact of the intervention in the most favorable possible case: does the intervention help students who start with a fixed mindset? In other words, do students with fixed mindsets who subsequently increase their growth mindset attitudes do better academically? If so, there might be something to salvage here. If not… then we may be mistaken about the relevance of a fixed mindset for many academic outcomes we care about.

Here’s a reformatted list of the readings from that course.

Topic 1: Collaborative Learning

Main reading: Dillenbourg, P. (1999). What do you mean by collaborative learning? In Collaborative-learning: Cognitive and Computational Approaches. Elsevier. PDF

Optional additional reading:

• In-Person Approaches — Education Endowment Foundation. Collaborative learning approaches. EEF Teaching and Learning Toolkit.
• Technology Support Approaches — Collaborative Learning Models and Support Technologies in the Future Classroom. (2011). International Journal for Educational Media and Technology, 5(1). ijemt.org
• Collaborative or Cooperative Learning? — Brandon, D. P., & Hollingshead, A. B. (2004). Collaborative learning and computer-supported groups. In T. S. Roberts (Ed.), Online collaborative learning: Theory and practice. Information Science Publishing. DOI:10.4018/978-1-59140-174-2.ch009

Topic 2: Reciprocal Peer Tutoring (RPT)

Main reading: Allen, A. R., & Boraks, N. (1978). Peer tutoring: Putting it to the test. The Reading Teacher, 32(3), 274–278. JSTOR

Optional additional reading:

• Research Paper #1 — Cheng, Y.-C., & Ku, H.-Y. (2009). An investigation of the effects of reciprocal peer tutoring. Computers in Human Behavior, 25(1), 40–49. DOI:10.1016/j.chb.2008.06.001
• Research Paper #2 — Walker, E., Rummel, N., & Koedinger, K. R. (2014). Adaptive intelligent support to improve peer tutoring in algebra. International Journal of Artificial Intelligence in Education, 24(1), 33–61. DOI:10.1007/s40593-013-0001-9
• Preliminary Research in Technology-enhanced RPT — Oikarinen, R. M., Oikarinen, J. K., Havu-Nuutinen, S., & Pöntinen, S. (2022). Students’ collaboration in technology-enhanced reciprocal peer tutoring as an approach towards learning mathematics. Education and Information Technologies, 27(6), 7519–7548. DOI:10.1007/s10639-021-10799-3

Topic 3: Heterogeneous Grouping

Main reading: Gillies, R. M. (2007). Group composition (Chapter 5). In Cooperative learning: Integrating theory and practice. Sage. SAGE Knowledge

Optional additional reading:

• Review of the literature — Education Endowment Foundation. (2018). Best practice in mixed attainment grouping (pilot).

Topic 4: Ubiquitous Feedback

Main reading: Monteiro, V., Mata, L., Santos, N., Sanches, C., & Gomes, M. (2019). Classroom talk: The ubiquity of feedback. Frontiers in Education, 4, 140. https://www.frontiersin.org/journals/education/articles/10.3389/feduc.2019.00140/full

Optional additional reading:

• Longitudinal Study — Drayton, B., Falk, J. K., Stroud, R., Hobbs, K., & Hammerman, J. (2010). After installation: Ubiquitous computing and high school science in three experienced, high-technology schools. The Journal of Technology, Learning and Assessment, 9(3). https://ejournals.bc.edu/index.php/jtla/article/view/1608
• Literature Review of Studies on Ubiquitous Learning Environments — Virtanen, M. A., Haavisto, E., Liikanen, E., & Kääriäinen, M. (2018). Ubiquitous learning environments in higher education: A scoping literature review. Education and Information Technologies, 23(2), 985–998. DOI:10.1007/s10639-017-9646-6

The story you tell about your research is the research. The narrative you share with the research community is how your findings are understood by others; without a story, there is no research.

Research papers are the most common medium for presenting research stories, but talks, posters, and hallway conversations all involve making choices about how to structure your research narrative.

Learning to structure research narratives doesn’t come naturally, and it’s one of the things you’ll learn during a PhD. Many researchers learn by imitating the structure of existing abstracts or Introduction sections, but I find it’s really useful to have a rough framework in mind. What follows is the framework I use, as taught to me by Mike Wittie during a 2012 REU program. I would think about this framework like the five-paragraph essay: a structure that helps you produce a reasonable argument, but overly restrictive once you get more experience structuring research narratives.

A framework for structuring research narratives

The framework has four components:

1. Societal Problem
2. Technical Problem
3. Technical Solution
4. How the Solution Helps Society

Societal problem

The societal problem tells us how the world is lacking.

The society in question can be as broad as you choose, from all of humanity to a small group of researchers, engineers, or practitioners trying to accomplish something.

A problem must – in principle – be solvable.

The problem should be as specific as you can make it for the audience you have in mind. “Food insecurity exists” is a societal problem, but “food distribution networks cannot meet existing needs” is more appropriate if the research you conducted is specific to food distribution.

Technical problem

The technical problem is why we can’t just fix the societal problem.

The technical problem is the domain-specific problem contained within the broader problem that is amenable to attack by our discipline’s methods.

Resolving the technical problem should recognizably move us closer to solving the societal problem.

The line between the societal problem and the technical problem is necessarily fuzzy. In general, you should start with a problem recognizable to any member of your audience and take argumentative steps toward a concrete technical problem specific to your discipline.

The societal problem and technical problem together are your strongest motivations for conducting this research. There may be other motivations that you mention incidentally or while discussing how your research relates to prior work, but lead with your strongest argument and ensure that addressing the technical problem clearly addresses the societal problem.

Technical solution

The technical solution is how our research fixes the identified technical problem.

Your solution should correspond clearly to the points you raised in the technical problem. If your technical solution seems to respond to a broader or more narrow problem than your technical problem suggests, that’s a good indicator that you should rework the technical problem or consider presenting less of the work.

For in-progress work, it works fine for the technical solution to be tentative or incomplete; by explicitly characterizing the technical problem you’re trying to solve, you make it obvious where the holes are in your solution.

It can be appropriate for the technical solution to include only a brief teaser of your research work. In longer formats, it is fine to include additional details about the methods you used and your results alongside the technical solution.

How the solution helps society

How the solution helps society draws us back from the specifics of your solution to the new state of the societal problem following your research.

If you discussed specific results while describing your technical solution, you should link those results to specific changes for researchers or practitioners who are facing the same societal problem.

Describing the societal benefit can be very brief, but you must do it: remind your reader where you started, or you’ll be left with an incomplete narrative.

Arguments to avoid

Choosing an appropriate societal problem and technical problem is important because these communicate your core motivations for conducting the research. Unconvincing or vague motivations will leave your audience confused about the potential value of your research.

Here are two common ‘bad’ arguments to avoid:

“X is hard”

Why haven’t we fixed the societal or technical problem yet? Well, because they’re hard problems!

Consider this argument:

Societal problem Lack of access to food causes human misery. Food distribution networks can provide access, but designing and implementing these networks is challenging. Technical problem Networks need stable supply chain data to distribute food. Technical solution We propose a new method for collecting and structuring supply chain data. How the solution helps society Our method enables food distribution networks to operate more efficiently and so increase access to food.

Instead, we should be more specific about why these problems are challenging.

Societal problem Food distribution networks can provide access to food, but designing effective networks depends on access to stable supply chain data. Technical problem Suppliers provide data in diverse formats or not at all. Technical solution We propose a new method for collecting and structuring supply chain data. How the solution helps society Our method enables food distribution networks to operate more efficiently and so increase access to food.

Sharpening the societal problem and specifying a concrete cause forced us to be more specific about the technical problem we actually solved.

When you feel tempted to make an “X is hard” argument, that’s usually a sign that your argument is too high-level. That can be okay for an elevator pitch, but you can make a stronger argument by being specific about what is hard.

Research gaps

One common piece of advice is to identify a “research gap”: a research question that hasn’t yet been addressed in the published literature.

However, research gaps make for ineffective motivations.

As Casey Fiesler writes on Bluesky:

I would love to never see the phrase “little is known about” in a paper ever again.

(1) Most of the time the authors are far more confident about that statement than they should be.

(2) There are lots of things we know little about because no one cares. Just not knowing isn’t a good motivation.

Fiesler is correct: “there’s a gap” is not a compelling justification for doing research. It is good to notice when there are gaps, but you should first ask:

Why is there a gap?

Use the answer to that question as your technical problem instead.

Example: My first paper

Let’s look at the abstract from my first paper and see how well I followed this framework.

The paper was called “Bridging Qualitative and Quantitative Methods for User Modeling: Tracing Cancer Patient Behavior in an Online Health Community”, and it proposed a new method for researchers to use when constructing quantitative user models. This is a hard test for this framework, because the contribution – and thus my motivations – were very conceptual.

Here’s the abstract, with annotations added:

Societal problem Researchers construct models of social media users to understand human behavior and deliver improved digital services. Such models use conceptual categories arranged in a taxonomy to classify unstructured user text data. Technical problem In many contexts, useful taxonomies can be defined via the incorporation of qualitative findings, a mixed-methods approach that offers the ability to create qualitatively-informed user models. But operationalizing taxonomies from the themes described in qualitative work is non-trivial and has received little explicit focus. Technical solution We propose a process and explore challenges bridging qualitative themes to user models, for both operationalization of themes to taxonomies and the use of these taxonomies in constructing classification models. Method detailsFor classification of new data, we compare common keyword-based approaches to machine learning models. We demonstrate our process through an example in the health domain, constructing two user models tracing cancer patient experience over time in an online health community. How the solution helps society We identify patterns in the model outputs for describing the longitudinal experience of cancer patients and reflect on the use of this process in future research.

Overall, I think this abstract is okay.

• My societal problem is expressed very implicitly. That can be okay, but I could consider an explicit frame instead: “Models of social media users without valid conceptual categories can produce misunderstandings and lead to ineffective designs.”
• The technical problem uses both of the arguments to avoid: I say the process is hard and that there’s a research gap! Instead, I should be specific: “But operationalizing taxonomies from the themes described in qualitative work requires a process that can bridge from the thematic to the concrete.”
• My description of the technical solution is vague, although I think that’s mostly okay in an abstract.
• My description of how the solution helps society is too vague. I should tie it back to the problems I identified more explicitly: “Our process creates quantitative user models that are both useful and valid.”

If I had used this framework more explicitly, I would have produced a stronger abstract! Here’s a revised version:

Societal problem Quantitative models of social media users without valid conceptual categories can produce misunderstandings and lead to ineffective designs. Technical problem In many contexts, useful category taxonomies can be defined via the incorporation of qualitative findings, a mixed-methods approach that offers the ability to create qualitatively-informed user models. But operationalizing taxonomies from the themes described in qualitative work requires a process that can bridge from the thematic to the concrete. Technical solution We propose a process and explore challenges bridging qualitative themes to user models, for both operationalization of themes to taxonomies and the use of these taxonomies in constructing classification models. Method detailsFor classification of new data, we compare common keyword-based approaches to machine learning models. We demonstrate our process through an example in the health domain, constructing two user models tracing cancer patient experience over time in an online health community. How the solution helps society Our process produces user models with qualitatively-grounded categories that better capture user behavior and are more useful for researchers.

Using this framework

I try to use this framework as early as possible in a research project. I’ll write four high-level bullet points and see if I find the narrative compelling even without detail. Each month, I return to my bullet points, revising them to be more specific and to see how my thinking about the research has changed.

Research abstracts lend themselves well to this framework, since an abstract should usually serve as a brief summary of the research narrative. But abstracts can include other details as well (listings of results, calls to action, important related work), so I like to keep a “clean” version of the narrative for any research project I’m working on.

In longer formats, such as a research talk or an Introduction section, you might choose to include more or less detail in each component of this framework. It can be fine to start with a smaller “abstract” version of your research narrative – essentially functioning as a roadmap – or to introduce only the societal problem initially. Highly conceptual work might require significant exposition before the audience can even understand the technical problem. For example, many pure math talks are 80% “technical problem”, with a sliver of higher-level context at the start and a final theorem presented at the end, with details left to a longer-form write-up.

Other frameworks

There are many frameworks for structuring research narratives. While writing up this post, I stumbled on Maggie Appleton’s notes on “On Opening Essays, Conference Talks, and Jam Jars”, which uses a very similar narrative framing.

Maggie Appleton’s beautiful figure adapting Joseph Williams’ writing on problem structure.

I was taught to write abstracts by John Carlis’ “Thinking about Abstracts”. He breaks down the story of a research abstract into a set of “sentence purposes”, and several of his examples strongly resemble the framework above. But Carlis makes an important point: there’s nothing sacred about a particular narrative structure. The framework above aligns with researcher expectations that are highly applicable to the fields I know well, but other fields can have very different expectations regarding research stories.

A good abstract tells a story. Furthermore, the parts of the story appear in a sensible order, one commonly seen by its (limited) audience. It, therefore, meets its readers’ expectations. When writing my own abstracts or helping others, I find that stating the purpose(s) of each sentence helps me see the story’s flow from a reader’s point of view, and readily find flaws in the story. If you think of an abstract as telling a story, then of course only some content and orderings are effective, and you can judge when content and order works or doesn’t.

When you start to diverge from the structure I present above, ask yourself: does this meet my readers’ expectations? Does my story ‘work’?

Other resources

• “10 tips for academic talks”. See the section on “Motivation” in particular.
• “Ten simple rules for structuring papers”
• Eugene Yan’s “Frequently Asked Questions about My Writing Process” – Includes a nice set of resources related to writing at the bottom of the post

Related posts

• Peer reviewing
• Common academic writing pitfalls

1. Pavel Samsonov, Taste is a dead end; what you should have is a point of view, Product Picnic, April 2025
2. Nicole Alexandra Michaelis, Designers, we should be killing it right now, UX Collective, March 2026
3. Harrison Wheeler, Do you have good taste? That’s debatable., technically speaking, April 2025

Is taste a useful concept for designers, researchers, and software developers?

Current thoughts: Taste is important, but – building on Michaelis and Samsonov – we should try to be more specific.

Further reading

• Michael Spicher, Aesthetic taste, Internet Encyclopedia of Philosophy
• Kyle Chayka, Why Tech Bros Are Now Obsessed with Taste, New Yorker, March 2026
• Zachary Levonian, My personal tipping point: LLM agents are good for coding now, this blog, January 2026

The biography for composer Jacob Isaacson was created on Wikipedia on June 29th, 2008. 17 years and 238 days later, I nominated it for deletion.

The article is about an American composer who supposedly died in 1980:

Jacob Isaacson (May 5, 1911 – September 8, 1980) was an American composer and musician. Isaacson was most noted for his own Colortone musical notation and his early works within this system. His association with the Fluxus movement was played down by Isaacson, who held European classical tradition in high regard, although his experimental and minimalist compositions drew inevitable comparison.

I initially nominated the article for deletion because it had no reliable sources. I couldn’t find anything about Jacob Isaacson on Google Search, Google Scholar, or the Internet Archive. In fact, the only mention I could find was in the description for a July 2008 event called FLUMMOXED at the Flatpack Film Festival:

And finally, Scott Johnston (Film Ficciones) and Matthew Eaton (Micronormous) will be performing American composer Jacob Isaacson’s final ‘Colortone Experiment’. The piece, called Lucent Harmonic Color, fuses Isaacson’s experiments in sound (inspired by his own synaesthesia) with the cut-up cinema of Burroughs and Balch.

Probably, Isaacson is just some ridiculously obscure composer, with a bio created by a family member or fan. But the experimental nature of the FLUMMOXED performance and the mention of Fluxus in Isaacson’s bio suggested there might be more going on here. Other participants in the deletion discussion pointed out some other weird inconsistencies in the biography.

So, I investigated further.

Investigating the hoax

The creator of the Jacob Isaacson article had only made edits to one other article: Pram (band)

Reading the Pram article tells us that Pram’s members are Rosie Cuckston, Matt Eaton, Andy Weir, and Samantha Owen. Further, the band “collaborated with visual artist ‘Film Ficciones’ (filmmaker Scott Johnston)” in 2010.

Wait! Matt Eaton and Scott Johnston are the supposed performers of Isaacson’s work at FLUMMOXED. That’s too strong a connection to be a coincidence.

I wanted to send a message to Pram. Despite being formed in 1988, Pram still has active social media! I sent an email to a hotmail.co.uk address I found listed on their Facebook, asking if they could resolve this mystery.

Shockingly, I got back a reply within the hour from Matt Eaton himself:

So finally Jacob Isaacson has met his match !

I did indeed create Isaacson as part of a one off performance I did with Scott Johnston at Flatpack Film Festival 2008. The theme of the piece was in part misdirection and misinformation with some good old fashioned Sci-Fi tropes woven in. The biography of this minor western composer somehow felt very close to home…

My apologies for putting my fictions on Wikipedia, it’s a resource I love and very occasionally donate to.

Straight from the horse’s mouth!

Thanks very much to Matt for resolving this mystery. It was fun to notice the connections between Isaacson and Pram and I’m pleased that we are able to clean up Wikipedia and get a good story out of it at the same time.

I told Matt there’s not much harm here: Isaacson’s biography got less than one human pageview per day and no other articles linked to that page. It also shows how much Wikipedia’s policies have changed: an article like that wouldn’t survive for more than an hour today.

There are still tens of thousands of unsourced Wikipedia articles created more than ten years ago. A few might be hoaxes like Jacob Isaacson, but most just need attention and care. If you want to help, WikiProject Unreferenced articles is running its March 2026 drive soon, and it’s a great way to jump in as a new Wikipedia editor and immediately feel like you’re making the world a better place.

Further reading

• Wikipedia:Articles for deletion/Jacob Isaacson
• Wikipedia:List of hoaxes on Wikipedia
• More posts about Wikipedia on this blog

I think these tools are officially a clear productivity boost for me. As echoed by many others, they not only make it easier to do the coding I was already planning to do, but they also make unthinkable refactors or fiddly, one-off tasks feasible – tasks that would otherwise not be worth the time. Just one example: I spent 30 minutes adding an InquirerPy TUI to a non-interactive ETL tool, just to make managing pipeline metadata slightly easier (and more accessible to non-developers on my team). I had never used InquirerPy before, so it was the perfect kind of task for large language models: labor-intensive for me to do, but easy to describe and easy to verify.

I wanted to mark this moment because overall I feel a strong sense of dread. I worry about major disruptions to white collar work, I worry about losing my own job, and I worry that even if I’m okay professionally my day-to-day work will become more competitive, less engaging, and more unstable. While it’s an exciting time, the uncertainty is very unpleasant.

Hopefully, I’ll look back on this post in a few years and laugh at my fears. (LLMs will never capture tacit knowledge, right?) Until then, I’ll leave you with a few articles reflecting on the automation of intellectual work and the use of LLM coding agents.

Further reading

On displacement

• “ChatGPT and the Meaning of Life” by Harvey Lederman (August 2025)
• “A Matter of Words” by Megan Fritts (May 2025)
• “AlphaGolem” by John Cornwell (November 2018)
• “Gödel, Escher, Bach author Doug Hofstadter on the state of AI today” (June 2023)
• “Developers say AI coding tools work—and that’s precisely what worries them” by Benj Edwards (January 2026)
• “Deep Blue” by Simon Willison (February 2026)

On agentic coding

• “10 things I learned from burning myself out with AI coding agents” by Benj Edwards (January 2026)
• “Use Agents or Be Left Behind? A Personal Guide to Automating Your Own Work” by Tim Dettmers (January 2026)
• Verification-driven development (January 2026)
• “My week with opencode” by Iris Meredith (January 2026)
• “How AI Is Transforming Work at Anthropic” by Saffron Huang, Bryan Seethor, Esin Durmus, Kunal Handa, Miles McCain, Michael Stern, Deep Ganguli (December 2025)