The city is studying improvements ahead of a 2027 repaving. Here’s the project page with traffic data, crash history, and a timeline for community input—the next open house is tomorrow (April 13) at the Carver Rec Center from 5:30–7:30pm!

Looking at the project materials, many people noticed two concerns that I think deserve attention.

“Safety” vs. “volume” is a false choice#

The page frames safety improvements and traffic volume as two separate considerations to balance. But they don’t have to be, and in fact they can go hand in hand.

This framing quietly assumes that “traffic volume” means “how many cars can move through”—a textbook case of motonormativity, the invisible bias that treats car traffic as the default and only kind of traffic that counts.

When we consider “traffic volume” we need to think about the number of people, not the number of cars! Personal vehicles are just one of the transportation options. A person on a bike takes a fraction of the space of a person in a car. A bus carries 40 people in the footprint of a few SUVs. Sidewalks and light-rails are supremely more space efficient and carry 10 times or more people if we dedicate the street space to them.

When you measure volume in people rather than vehicles, safety improvements like bike lanes and transit priority increase the street’s capacity while making the streets much safer. Furthermore, as streets get safer with other modes, the more people use those other options, reducing the traffic as a result.

Safer streets are more efficient streets; safer streets can move much higher people-traffic volume.

The real trade-off isn’t safety vs. volume. It’s safety vs. the speed and convenience of a single mode—driving. And when you make that trade-off explicit, the answer becomes obvious.

The projection is a self-fulfilling prophecy#

The project page projects Rose Hill’s traffic will grow from ~7,600 to ~9,900 vehicles per day by 2055—roughly 1% annual growth over 30 years. Where does this number come from? The methodology isn’t documented on the project page, but it’s consistent with a standard compound growth factor applied to current counts.

This is really problematic. And it is not about crude projection. There is a fundamental issue that cannot be addressed by sophisticated traffic modeling.

The projection implies that future traffic is something that happens to a road. But traffic demand is not weather. Future traffic depends overwhelmingly on how the road is designed. This is induced demand at work: build for more cars, and more cars will come; build for people walking and biking, and people will walk and bike.

Although the projection is a self-fulfilling prophecy, it has been used as a justification to widen the road and buldoze low-income neighborhoods.

If you design Rose Hill Drive for 9,900 cars a day, you’ll get 9,900 cars a day. If you design it for safe, comfortable cycling and walking, you’ll get more people cycling and walking—and fewer cars. The projection doesn’t tell us what will happen. It tells us what will happen if we keep designing for cars. That’s not a forecast; it’s a choice disguised as a forecast.

The double standard#

And notice the asymmetry: some projections argue that the US ebike market is growing at ~15% CAGR—far faster than the 1% car traffic growth projected here. Yet no one is using that projection to justify a massive expansion of bikeways. That’s motonormativity again: car traffic growth is treated as inevitable, what will happen to us, and worth planning for, while the explosive growth of other modes is completely ignored.

What Rose Hill Drive should be#

Rose Hill currently carries 6,000–7,600 vehicles per day. With 35 crashes in the past 8 years—including incidents involving pedestrians and cyclists—the 2027 repaving is an opportunity to build for the future. The road that is most convenient for driving—wide, multi-lane, no traffic calming—is the least safe and least efficient option, and it will induce more traffic. The road where transit, cycling, and walking are prioritized is the safest and most efficient option, carrying far more people while reducing traffic and crashes everywhere. The choice should be clear.

If you’re in Charlottesville, consider showing up tomorrow and communicating these messages. These decisions shape the neighborhood for decades.

yy

I posted this on Bluesky too, if you want to chime in.

The fix: describe the underlying problem (X) before asking about a candidate solution (Y) and provide more context to your question.

• https://en.wikipedia.org/wiki/XY_problem
• https://xyproblem.info/
• https://www.threads.com/@nomad__surfer_97/post/DW2UvmQkuQE

Measuring disruption#

Our paper “Uncovering simultaneous breakthroughs with a robust measure of disruptiveness” just came out in Science Advances. The disruption index has an elegant idea: if a paper is truly disruptive, future work cites it but stops citing its references. But because it relies on local citation topology, simultaneous discoveries break it. When multiple papers make the same breakthrough, they cite each other, and those mutual citations make each paper look “consolidating” rather than disruptive.

The Higgs mechanism is a perfect example. Three independent teams proposed it in 1964, yet the disruption index ranks one paper near the top and the others at the very bottom. We kept finding this pattern across reverse transcriptase, the charm quark, asymptotic freedom, and other famous simultaneous discoveries.

We introduce an embedding-based measure that resolves this. Each paper gets two vectors, one for its prior context and one for its future context, and the distance between them captures how much a paper redirected the field. Nobel Prize-winning and milestone papers consistently score high on our measure.

But a really interesting thing is that the vectors themselves let us discover simultaneous breakthroughs. The future vectors roughly capture how a paper is used in subsequent literature. If two papers that came out around the same time have very similar future vectors and high disruptiveness scores, they are likely simultaneous discoveries. If they cite each other, the disruption index would tank, but our measure still sees them as important. By combining these signals, we can systematically identify simultaneous disruption events across science. Munjung Kim’s blog post walks through the study in detail. See also her posts on Bluesky and LinkedIn.

Paper | Code | PyPI

Claude scholar plugin#

I’ve been building Claude Code skills for academic workflows and packaged them as a plugin called Claude Scholar.

• claude-scholar (on GitHub)

The key design principle is that these skills don’t try to do the thinking, researching, or writing for you. For those tasks, there’s a whole can of worms (ethical, pedagogical, reliability), and even setting those aside, there’s a vast space of how to prompt and use AI. It’s counterproductive to try to consolidate that into a fixed skill when the underlying models and systems keep advancing rapidly. For this type of fluid tasks, you can very easily over-engineer and mis-step.

Instead, the plugin focuses on concrete, verifiable tasks that are tedious to do by hand but where correctness is easy to check: fetching arXiv metadata, grabbing a BibTeX entry from a DOI, querying OpenAlex’s 240M+ scholarly works, verifying each step of a math derivation with SymPy, checking your references against databases, auditing figures for colorblind risk and resolution issues, walking through a checklist for arXiv submission. These are the kinds of tasks that don’t require judgment calls—they just need to get done right.

The skills are also composable: each does one thing and can be chained by other skills or agents. For instance, the skill for presubmission checks launches multiple agents in parallel to verify references, audit figures, check LaTeX formatting, and review front matter, then presents a unified report. You should also be able to easily build your own skills on top of these building blocks.

If you use Claude Code for academic work, give it a try! More details on my wiki post.

That’s it for now. More soon!

yy

Imagine yourself starting your car. Would you go to a smaller nearby store with fewer items, or drive 5 more minutes to Costco or other big box stores? The marginal cost of driving a bit further is tiny once you’ve started the car. Ample car parking at an urban grocery store is competing in a game it can’t win—at the expense of space that could serve far more customers as bike parking, expand the store itself, or simply lower prices for everyone.

The best customers an urban grocery store can attract are those who live nearby and walk, bike, or ride transit—people who stop by on the way home rather than buying in bulk.

Yet, because of inertia and the legacy of car-centric urban design in North America, I believe there are still too many car parking spots and too few bike parking spots at urban grocery stores.

Here is a list of (incomplete) cases for more bike parking and fewer car parking spaces.

You don’t need a car for groceries#

A common objection: “But I need a car to carry groceries.” A regular bike with panniers handles a typical grocery run. A cargo bike or an ebike with a basket can carry a week’s worth of groceries for a family, and then some. Check out the Carry Shit Olympics on Instagram for what people haul on bikes. The “you need a car for groceries” assumption is a relic of the bulk-buying, suburban big-box model. For an urban store where people shop more frequently and buy less per trip, a bike is more than enough.

Grocery access is an equity issue#

Car ownership costs a lot. The lowest income quintile spends roughly 30% of after-tax income on transportation—largely because car-centric infrastructure leaves them no choice but to drive for their work and grocery shopping. A nearby grocery store with abundant, convenient bike parking gives low-income residents a way to meet a basic need without a car, and can help reduce that transportation burden significantly. Bike parking is equity infrastructure.

Car parking is expensive and subsidizes the rich#

Car parking isn’t free. Paving alone runs $1,000–$2,600 per surface space; add site prep, grading, drainage, and striping, and the total reaches $5,000–$10,000+ per space. And that’s just construction.

The opportunity cost of the land is even higher. A typical parking space takes 300–350 sq ft including its share of the driving aisle. Urban commercial land typically runs $50–$500+/sq ft depending on the city, meaning each parking space sits on $15,000–$175,000+ worth of land—land that could be generating revenue, housing, or community space instead of storing idle cars.

When parking is “free,” these costs don’t disappear—they’re passed on to all customers through higher prices. “Free” parking can add ~$225/month to apartment rents, paid by every tenant regardless of whether they own a car. Since car ownership skews wealthier, this means people without cars—disproportionately lower-income—are subsidizing parking for those who can afford to drive.

We can have 10 more customer parking, per spot#

One car parking space can easily fit 5-10 bikes. Every car space converted to bike parking means 10 more customers the store can serve. Given the space efficiency (10x), and the fact that bike racks don’t require as much work to install, it’s simply not comparable to the cost of car parking.

Induced demand works both ways#

Induced demand is real: more car parking induces more driving to the store, creating more car traffic around the store. The surrounding streets bear the burden—congestion, noise, and reduced safety for pedestrians and cyclists. We all know the burden of traffic too well.

But induced demand works both ways. More bike parking (and fewer car parking spaces) encourages cycling and discourages driving. While the strongest evidence comes from bike lane infrastructure rather than parking alone—e.g., in Cambridge, MA, separated bike lanes led to a 500%+ surge in bike volumes, and across Boston, new bike lanes saw car traffic drop 9–15%—the principle applies to bike parking too. When you make cycling convenient at the destination, you reinforce the same mode shift. Reduced car capacity can actually reduce total traffic through traffic evaporation, not just shift it elsewhere.

This creates a virtuous cycle: more bike parking → more people cycling → cycling feels normal → even more people cycling. The opposite is equally true: more car parking → more driving → feels like you need even more car parking. We get to choose which cycle to invest in.

Bike infrastructure delivers 3:1 to 5:1 returns on investment through healthcare savings, productivity gains, and local business activity. Every person who cycles instead of driving means less road damage, fewer serious crashes, lower emissions, more revenue for local businesses, and a healthier community. Bike parking is one of the cheapest ways to unlock these gains—in fact, it’s not really “costing” us money at all. Every spot allocated to bikes instead of cars saves construction costs, frees up land, and serves far more customers. It saves money.

Spillover benefits#

Business owners often assume drivers are their most valuable customers. The evidence says otherwise: people on foot or bikes spend as much or more at local businesses than drivers—30–40% more in some studies.

People arriving by bike are also more likely to stop at nearby businesses on the same trip. A bike-friendly grocery store strengthens the whole nearby commercial corridor.

Build for the future, not the present#

Last but not least, climate change demands a shift toward more efficient and sustainable transportation. For instance, the ebike revolution is already accelerating this—the US ebike market is growing at ~15% CAGR. Ebikes flatten hills, extend range, and make cargo trips practical, turning grocery runs by bike from niche to no-brainer. The shift from driving to cycling and other sustainable modes is happening globally—sometimes gradual, sometimes drastic (e.g., Paris & London), but the direction is clear and irreversible. Parking infrastructure built today will be around for decades. Overbuilding car parking locks in car dependency; investing in bike parking bets on where our cities are headed!

So, what should we do?#

1. Make it plentiful. Err on the side of too many racks, not too few. Nobody complains loudly about full bike racks the way they do about full parking lots—motonormativity makes us far more sensitive to the lack of car parking. But people adapt to what’s convenient. Over time, less car parking doesn’t mean frustrated drivers; it means more people walking, biking, and riding transit.
2. Make it convenient. Bike parking should be closer to the entrance than car parking—right at the front door. If it’s hidden behind the building or across the lot, people won’t use it, bikes will get stolen, and you lose all the benefits. Every small nudge that makes cycling easier and driving less necessary matters.
3. Install good racks. Use the most boring inverted-U (Sheffield) racks that support the frame and allow U-lock use for both “normal” and cargo bikes. Avoid wheel-bender racks (the ones that only grip a wheel)—they damage bikes and don’t prevent theft. See Essentials of Bike Parking (APBP) for detailed guidelines.
4. Accommodate all bikes. Cargo bikes, ebikes, trikes, adaptive bikes, and bikes with trailers are increasingly common, especially for grocery trips. Include some wider/longer spaces.
5. Cover the racks if possible. A roof keeps bikes (and groceries being loaded) dry and protects from sun. Solar panel canopies can do double duty—generating electricity while sheltering bikes. Covered parking also sends a clear signal: this store takes cycling seriously.
6. Make it visible and well-lit. Bike parking in plain view of the entrance deters theft and makes cyclists feel secure.
7. Connect to the bike network. A bridge that leads to nowhere is never used. Bike parking without a safe route to get there is just decoration. Advocate for protected connections between the store and the surrounding bike network and important destinations.
8. Build for the future, not the present. Demand for bike parking will only grow. The infrastructure decisions made today will shape how people get to this store for decades.

I posted this on Bluesky and Mastodon too, if you want to chime in.

Related#

• Climate and Transit — The Problem With Bike Parking (and how we can fix it)

I’ve been gradually building Claude Code skills for academic workflows and packaged them as claude-scholar. These skills may be useful for you.

What’s in it#

Eleven skills (for now). Some can be invoked as a slash command and some are for Claude. Note that skills can be composed. If you create more atomic skills, then other skills or agents can compose them to do their jobs. In other words, principles of software engineering are useful (obviously?).

Tools for Claude (mainly)#

These skills are usually invoked by other skills that I have, which are more customized to my workflow (and thus not part of this plugin). For instance, for my wiki, I have a /wikify skill that uses whatever information I throw at it (e.g., paper URL, paper title, arXiv ID, DOI, …) to find metadata about the paper and create a page in my wiki. These kinds of skills use some of the following skills to get verified metadata about the paper.

• /arxiv-metadata — fetch structured metadata for arXiv papers (title, authors, date, categories, DOI). Handles multiple ID formats. I let Claude use this skill to find paper metadata when filing a paper into my wiki or other tasks.
• /doi-bibtex — given a DOI, fetches the BibTeX entry and appends it to the project’s .bib file, checking for duplicates. I don’t really invoke it myself. Once I have a DOI, Claude Code can download and add the bib entry to your .bib file.
• /openalex — query 240M+ scholarly works via the OpenAlex API for literature searches, citation analysis, and bibliometric queries. The helper script handles rate limiting, retries, and pagination. No API key required.
• /verify-math — step-by-step verification of mathematical derivations using SymPy. Each derivation step is validated programmatically. Using this skill, Claude can check each step of a mathematical derivation.

Commands that you can invoke#

• /check-refs — verify every citation in your LaTeX paper against academic databases, flag suspicious entries, and suggest missing DOIs/URLs. Uses bibsleuth. Can run standalone or as part of /presubmit-checks.
• /latex-cleanup — systematic review for common LaTeX issues: style consistency, typography, cross-references, draft artifacts, my LaTeX pet peeves, etc. Identifies all .tex files in the project (including \input/\include targets).
• /presubmit-checks — pre-submission checklist that orchestrates multiple checks in parallel: references (check-refs), LaTeX cleanup (latex-cleanup), build verification, figure format checks (flags bitmap figures that should be vector), and front matter review (affiliations, acknowledgements, data availability). Presents a unified report organized by severity.
• /arxiv-prep — automates arXiv submission packaging: cleans source files, removes cruft, verifies compilation, extracts metadata for the submission form, and produces a ready-to-upload tarball. Uses Google’s arxiv-latex-cleaner under the hood.
• /critique-manuscript — structured self-review of your own manuscript. Evaluates the paper across seven review criteria (literature, methodology, causal claims, data quality, generalizability, mechanism, clarity) and produces a report with top risks, what to preserve, anticipated reviewer questions, and detailed comments anchored to specific sections. Optionally searches OpenAlex for missing literature. Designed strictly for self-review; declines to review others’ unpublished work and explains why.
• /critique-figures — critiques academic figures using a mix of programmatic checks and vision analysis. Checks file format and resolution, colorblind risk, excessive categories, font legibility, overplotted scatterplots, and dynamite plots. Handles multi-panel figures by evaluating each panel individually. Report-only — does not modify figures.
• /accessible-pdf — transforms a LaTeX document to produce tagged, screen-reader-friendly PDFs (PDF/UA-1, PDF/A-2b). Asks whether to edit in place or create a copy, applies preamble changes, switches to LuaLaTeX, and verifies tagging. See Accessible PDF from LaTeX.

Installation & Usage#

/plugin marketplace add yy/claude-scholar
/plugin install claude-scholar@yy-claude-scholar
/reload-plugins

You may need to restart Claude Code to make it function properly. Please note that these skills are early-stage products and not comprehensively tested, so proceed with caution. And if you use it for your paper, please check your venue’s AI use policy.

And of course, feedback is welcome!

Acknowledgements#

• K-Dense-AI/claude-scientific-skills — inspiration and starting point for the OpenAlex skill
• Sadamori Kojaku — inspiration for the math verification skill
• Trevor Campbell’s arXiv checklist — inspiration for the arXiv prep skill
• Google’s arxiv-latex-cleaner — used by the arXiv prep skill for source cleaning

See also#

• Kaicheng Yang — Claude Code Skills for Research — links several similar efforts

Voice-to-text dictation app for macOS. Transcribes speech directly into any text field. Sometimes useful with Claude Code for quick prompts.

• https://cmux.com
• GitHub: https://github.com/manaflow-ai/cmux

Why I switched#

I went from bare Ghostty splits → tmux → back to bare Ghostty → cmux. The key thing cmux adds over bare Ghostty is a vertical tab sidebar with stable naming (one tab per project), in-app web browser, plus a notification system that tells me when an agent needs attention. Ghostty’s native splits are fine but it renames tabs based on whatever process is running in the focused pane, which is annoying when each tab is a project. cmux solves this while keeping Ghostty’s speed and GPU-accelerated rendering.

See How I Use It for how this fits into my workflow.

Key features#

• Vertical tab sidebar — shows git branch, working directory, listening ports, and latest notification per workspace
• Notifications — panes show a visual indicator when agents need input; Cmd+Shift+U jumps to the most recent unread; cmux notify CLI integrates with Claude Code hooks
• Splits — horizontal and vertical, with tabs created within a split (not the other way around)
• Ghostty config compatible — reads ~/.config/ghostty/config for themes, fonts, colors
• Scriptable — CLI and socket API for creating workspaces, splitting panes, sending keystrokes
• Embedded browser — scriptable in-app browser that agents can interact with (snapshot accessibility trees, click, fill forms)

Install#

brew install --cask cmux

Life After Cars — Charlottesville, March 30 — On Monday, March 30 from 6:30–7:30pm, there will be a conversation about what life after cars could look like in Charlottesville, featuring Sarah Goodyear and Doug Gordon from The War on Cars and Jamelle Bouie as a special guest. This should be an informative event. So come join! RSVP here. Shoutout to Livable Cville for organizing.

Downsizing fire trucks for safer streets — I came across a 2018 USDOT/NACTO report on how oversized fire trucks shape our streets. Cities design roads to accommodate the largest vehicles, which means wider lanes and larger turning radii—exactly the features that make streets dangerous for pedestrians and cyclists. The report argues that downsizing emergency vehicles is a practical step toward Vision Zero. Not Just Bikes’ video “How American Fire Departments are Getting People Killed” covers this well—comparing the massive North American fire trucks to the smaller, more maneuverable ones used in Europe.

This conflict between better fire service and safe street infrastructure is a frequent challenge with sharp, difficult trade-offs and regulations. Recently, Richmond’s fire department opposed physical barriers on the Mayo Bridge bike lanes, arguing fire trucks need cars to pull into the bike lane to let them pass.

Also, I’ve been updating my Ebike guide page if you’re curious about ebikes or thinking about getting one.

And if you’re in Charlottesville, the Ebike Library Cville lets you borrow and test ebikes—a great way to try before you buy.

One of the best things about walkable and bikeable cities is that they give independence to people who can’t or shouldn’t drive.

yy

LLMs Can Infer Political Alignment from Online Conversations — A new paper from our group. We show that LLMs can infer political alignment from general online conversations. The combination of LLMs’ capacity and widely available online social data means your hidden traits like political leaning can be readily inferred (“inferential privacy violation”).

How to Do Peer Review — I’ve expanded my guide on peer review. The core message: think of yourself not as a judge but as an anonymous sage who helps the authors improve the paper. The guide covers mindset, structure, and common pitfalls.

cmux — My current terminal setup. It’s a native macOS terminal built on Ghostty with a vertical tab sidebar and a notification system for AI coding agents. The killer feature for me: each tab is a project with a stable name, and it tells me when Claude Code needs attention in another tab. I went from bare Ghostty → tmux → bare Ghostty → cmux, and this is the best fit so far for agent-heavy workflows.

Accessible PDF from LaTeX — With the ADA Title II compliance deadline coming up in April 2026, you may need to make an accessible CV, which means tagged, screen-reader-friendly PDFs. The fix is surprisingly minimal—small changes to your preamble, switch to LuaLaTeX, and your existing layout is preserved. If you use LaTeX for your CV, this guide may be handy.

Claude Code (updated) — I’ve split my Claude Code notes into separate pages: Claude Code (overview), Skills (the skill system), Tips (practical tips), and How I Use It (my personal setup). The biggest additions: custom skills and agents I’ve built for my workflow (filing papers, gardening the wiki, processing inboxes), and my current terminal setup with cmux.

ofocus — A Python CLI I built for OmniFocus on macOS. It talks to OmniFocus via JXA, so no plugins or API keys needed. The real motivation is that I wanted agents like Claude Code to be able to manage my tasks directly from the terminal. Available on PyPI (uvx ofocus).

That’s it for now. More soon!

yy

• Traffic evaporation - 2026-04-12
• Traffic calming - 2026-04-11
• Cost of car ownership: the million-dollar decision you're not thinking about - 2026-04-07
• Aiyappa et al., "Implicit degree bias in the link prediction task" (2025) - 2026-04-03
• Aiyappa et al., "Zero-shot stance detection in practice" (2026) - 2026-04-03
• Boucherie et al., "Adaptive cut reveals multiscale complexity in networks" (2025) - 2026-04-03
• Constantino et al., "Representing the Disciplinary Structure of Physics" (2025) - 2026-04-03
• DeVerna et al., "Modeling the amplification of epidemic spread by misinformation" (2025) - 2026-04-03
• Frank et al., "AI exposure predicts unemployment risk" (2025) - 2026-04-03
• Hébert-Dufresne et al., "One pathogen does not an epidemic make" (2025) - 2026-04-03
• Jing et al., "Sameness Entices, but Novelty Enchants in Fanfiction Online" (2025) - 2026-04-03
• Jo et al., "Extracting hierarchical backbones from bipartite networks" (2023) - 2026-04-03
• Kim et al., "The Potential Impact of Disruptive AI Innovations on U.S. Occupations" (2025) - 2026-04-03
• Kim et al., "Uncovering simultaneous breakthroughs with a robust measure of disruptiveness" (2026) - 2026-04-03
• Kojaku et al., "Community-centric modeling of citation dynamics explains collective citation patterns in science, law, and patents" (2025) - 2026-04-03
• Lee et al., "A semantic embedding space based on large language models for modelling human beliefs" (2025) - 2026-04-03
• Lee et al., "LLMs Can Infer Political Alignment from Online Conversations" (2026) - 2026-04-03
• Miao et al., "Cooperation and interdependence in global science funding" (2025) - 2026-04-03
• Miao et al., "Persistent Hierarchy in Contemporary International Collaboration" (2025) - 2026-04-03
• Park et al., "People, Places, and Ties" (2023) - 2026-04-03
• Park et al., "Rap Music as a Social Reflection" (2026) - 2026-04-03
• Seckin et al., "Emergence of Stereotypes and Affective Polarization from Belief Network Dynamics" (2026) - 2026-04-03
• Weng et al., "Beyond Distance: Mobility Neural Embeddings Reveal Visible and Invisible Barriers in Urban Space" (2025) - 2026-04-03
• Wright et al., "Cognitive Linguistic Identity Fusion Score (CLIFS)" (2025) - 2026-04-03
• Protein intake - 2026-04-02
• Accessible PDF from LaTeX - 2026-04-02
• ArXiv - 2026-03-31
• Claude Code: How I Use It - 2026-03-27
• foliate - 2026-03-25
• YYiki - 2026-03-25
• Claude Code: Skills - 2026-03-24
• Todd et al., "In-Context Algebra" (2025) - 2026-03-23
• Zimmerman, "Tokens, the oft-overlooked appetizer: Large language models, the distributional hypothesis, and meaning" (2024) - 2026-03-23
• Kapoor & Narayanan, "Leakage and the Reproducibility Crisis in ML-based Science" (2022) - 2026-03-23
• Hofman et al., "Pre-registration for Predictive Modeling" (2023) - 2026-03-23
• Gloaguen et al., "Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?" (2026) - 2026-03-23
• Julia - 2026-03-22
• Research advice - 2026-03-21
• 연결망 임베딩 - 2026-03-20
• Claude Code: Tips - 2026-03-20
• Claude Code - 2026-03-17