Research in motion
Notes on AI and research workflow
Engineering notes, methodology, and case studies — bringing AI into every step of the social-science research workflow.
Start with Stata analysis or Zotero reading, and check data processing.
A LaTeX Guide for Social Scientists: No Coding Required
You can use LaTeX to typeset papers without mastering it. Starting with my repeated attempts to learn it and subsequent decision to stop, this guide covers what LaTeX offers, where TeX Live / MacTeX + VS Code, Overleaf, and Agent Driven workflows fit, and how to hand content and formatting requirements to an agent and continue revising the resulting document.
Keywords: LaTeX for beginners; social science typesetting; Overleaf; academic writing tools; reference management
Read the post →Use AI for Data Cleaning and Merging by Checking Keys and Exceptions First
A successful merge does not establish that keys, duplicate handling or sample coverage are correct. Using original simulated firm-year and region tables, this guide explains field normalization, exact duplicates versus conflicting keys, preservation of missing and unmatched records, and row-level provenance. The companion case provides independent Python and Stata calculations, raw simulated CSV, audit tables, cleaned results and a complete rerun package.
Keywords: AI data cleaning; Stata merge; duplicate records; merge keys; Marx Max
Read the post →Turn Stata Regression Results into an Editable Word Table
A regression table must preserve numerical results, model conditions and an editable layout. Using two actually executed Stata regressions on the same 74 observations, this guide explains exporting coefficients, robust standard errors, sample sizes and notes to a native Word table, checking display precision, and retaining CSV, do-files and logs. The companion case provides the generated document, an actual preview and a complete rerun package.
Keywords: Stata Word export; regression table; editable tables; robust standard errors; Marx Max
Read the post →Following Academic Journals with RSS: Start by Discovering New Papers
Literature searches help researchers find existing work on a specific question, but research also requires keeping up with new journal articles and working papers. RSS lets researchers choose their own sources and view updates from different websites in one place. This article explains how RSS works, how to combine sources around a research direction, and how titles, abstracts, and publication information can inform an initial reading decision. It also introduces the design of Marx Max’s paper subscriptions, bringing source management, update browsing, and reading status into the same working environment, while discussing the limits of RSS coverage, full-text access, and literature screening.
Keywords: RSS; Paper subscriptions; Academic journals; NBER working papers; Literature tracking; CNKI
Read the post →Is a High MCMC Acceptance Rate Enough? Comparing Three Scales for One Posterior
Random-walk Metropolis proposal scales affect finite-chain exploration. Using a one-parameter posterior with an analytic reference, this guide compares traces, acceptance, mean and spread, explains why rejected states must remain and a discarded prefix does not prove convergence, and provides all states, executed code and a rerun package.
Keywords: MCMC teaching; Metropolis proposal scale; Posterior sampling; Acceptance rate; Convergence diagnostics
Read the post →Heteroskedasticity and Robust Standard Errors: Comparing Python and Stata on the Same Data
Robust standard errors change covariance estimation, not OLS coefficients. Using the same 500 simulated observations, this guide compares conventional and HC1 standard errors, intervals and p-values, provides independently executed Python and Stata outputs and a rerun package, and distinguishes heteroskedasticity from endogeneity and dependent observations.
Keywords: Heteroskedasticity; Robust standard errors; HC1; Stata vce(robust); OLS inference
Read the post →Public Data Analysis with Marx Max: From World Bank Indicators to Reproducible Artifacts
Public-data analysis starts with measurement and provenance, not just a chart. Using GDP growth and unemployment changes in six countries, this guide explains how to define a task, retain official API snapshots, check years and missing values, compare Python and Stata output, and download data, code, figures and rerun instructions.
Keywords: World Bank data; AI public-data analysis; Python and Stata; Okun relationships; Reproducible artifacts
Read the post →Preparing a Beamer Paper Presentation with Marx Max: From Outline to PDF
A Beamer report requires both a research narrative and compilable LaTeX source. This guide explains how to prepare papers and notes, review an outline, separate author conclusions from presenter discussion, retain source and dependencies, and inspect the final PDF. The accompanying case provides a published 25-page Chinese report preview and PDF download.
Keywords: Beamer paper presentation; AI academic reports; LaTeX slides; Seminar PDF; Marx Max
Read the post →Where Does Research Data Go When You Use AI in Marx Max?
Keeping research files locally and transmitting content for AI processing are different questions. This guide distinguishes local files and computation, conversations and tool results sent to models, cloud plans in Planar, and Zotero's own synchronization. It explains task-specific reading boundaries and why training use, retention, and deletion conditions need separate verification.
Keywords: AI research data; local files; model data transfer; Zotero synchronization; Marx Max
Read the post →Planning the Unplannable
Research rarely follows a fully determined path: literature, data problems, and model results can reopen earlier work. A checklist recording only pending and completed tasks can miss which judgments need revisiting. This article explores a plan built around feedback, artifact dependencies, and human–AI roles. A hypothetical empirical analysis shows how a sample discrepancy discovered during model comparison can return to sample construction, preserve valid work, and trigger review of affected outputs. Planar is being integrated into Marx Max with this design: AI assists with concrete work, people participate in consequential decisions, and feedback shapes what happens next.
Keywords: research task management; Planar; research TODO; human–AI collaboration; research feedback loops
Read the post →Using Zotero in Marx Max: From Selected Papers to Sourced Notes and Citations
How can AI help with papers already stored in Zotero? This guide explains working with a local library, distinguishing metadata, child notes, and PDF attachments, asking focused reading questions, checking evidence, generating references, and adding notes under the appropriate conditions. Finding an item, reading its full text, and establishing support for a claim are separate steps.
Keywords: Zotero guide; AI literature reading; literature notes; references; Marx Max
Read the post →Don’t Mistake AI’s Beta for Your Own Alpha
Obtaining code, regression results, and polished prose with AI can improve delivery without establishing that the user has acquired the methods involved. This essay examines a pattern of use: passing an unclear choice back to AI, accepting an answer, and ending the inquiry before developing understanding. It distinguishes being unfamiliar with a method from skipping the opportunity to learn it, and considers how explanation, verification, and revision relate to personal capability. Findings from a mathematics experiment are separated from implications for research; a model comparison is explicitly hypothetical. The closing proposal keeps unresolved questions in the collaboration so that AI assistance can support understanding as well as produce output.
Keywords: AI-assisted research; personal capability; independent thinking; AI and learning; research judgment
Read the post →Using Stata in Marx Max: From a Research Request to Code and Results
How can AI help when you already have Stata, data, and a do-file? This guide covers connecting a local Stata installation, using the bundled auto dataset to compare two models, checking actual output, handling errors, and saving code, logs, and results for further work. The example regressions were checked in an independent official PyStata process; your research still requires decisions about data, samples, and specifications.
Keywords: Stata guide; AI-assisted data analysis; do-file; regression checks; Marx Max
Read the post →Right Code, Wrong Way
AI can write code and estimate regressions while also changing samples, variables, and models in ways that affect conclusions. Sant’Anna’s recent account and two studies, The Agentic Garden of Forking Paths and Do Claude Code and Codex P-Hack?, make the question concrete: why can identical data yield different or opposing findings? This article examines how expectations can enter exploration and selection, distinguishes execution errors from defensible disagreements and outcome-driven choices, and uses a hypothetical model comparison to explain their implications. When searches are concealed and results are selected to match expectations, the risk can involve research misconduct rather than an execution error.
Keywords: AI data analysis; AI empirical research; AI p-hacking; selective reporting; researcher degrees of freedom
Read the post →Which AI Persona fits your research?
A useful research AI Persona depends on what the researcher needs to learn, judge or complete, and which outputs they can independently check. This article considers students learning methods, doctoral and early-career researchers developing independent arguments, and senior researchers working in familiar domains. Drawing on research about role prompting, learning and sycophancy, it explains why expert labels do not guarantee accuracy and offers adjustable role agreements for explanation depth, questioning and deliverables. The recommendations are organised around tasks rather than fixed career categories.
Keywords: Research AI Persona; AI role prompting; Graduate students and AI; Research prompts; Human-AI collaboration
Read the post →Can AI search for academic literature?
AI can search for academic literature when tools supply real records or texts; a generated reference list alone is not evidence of retrieval. Fabricated citations undermine claims and their use in research work may constitute academic misconduct. This article compares online scholarly sources, dedicated AI search services and local libraries, explains the uses and access limits of Wiley, CNKI-MCP, Elicit, OpenAlex, Semantic Scholar, Crossref and Zotero-MCP, and answers common questions about full text, DOI lookup and library coverage. It then considers how literature tasks can continue around a research question in Marx Max.
Keywords: AI literature search; Literature search MCP; Wiley Scholar Gateway; CNKI-MCP; Zotero-MCP
Read the post →Your group meeting is approaching: how to prepare the slides in half an hour
Marx Max now supports editable presentations. It can help organise papers, reading notes and research materials in the project into a slide draft, then revise it with your feedback. For a paper already read but a presentation still unprepared, this guide offers a half-hour preparation plan: define the question, ask AI to organise and create a draft, then check the key evidence and rehearse. An existing AI Agents for Economic Research presentation shows the outcome and provides a route to try the task in Marx Max.
Keywords: Marx Max presentations; AI presentation creation; Group meeting slides; Paper presentations; Academic presentations
Read the post →Using Marx Max in econometrics teaching
AI can generate and run code to turn abstract econometric concepts into observable differences. Three examples develop this approach: Python simulations distinguish one OLS estimate from its sampling distribution; Stata automobile data connects a negative price–fuel-efficiency coefficient with economic interpretation; and MCMC compares three proposal scales for the same posterior. Each includes a prompt for AI, followed by equations, a regression screenshot or simulation figures that develop the argument. Companion Python and Stata code supports reproduction and checking.
Keywords: Teaching econometrics; OLS sampling distributions; Stata regression coefficients; Teaching MCMC; Bayesian regression
Read the post →Why AI is hard to use in empirical research
AI participation in empirical research depends on the relationships among questions, data preparation, specifications and interpretation. This article examines execution conditions, retained data meaning, program feedback versus empirical validation, and effective collaboration. An illustrative policy-and-R&D task develops the requirements for samples, model comparisons, human edits and repeatable calculation. It explains Marx Max’s collaboration design, assessed through usable deliverables and effort across the complete task.
Keywords: AI for empirical research; Research workflows; Causal inference; Research data preparation; Human–AI collaboration
Read the post →Why Marx Max is built for research workflows
Development tasks can often organise implementation around a defined outcome, while research involves successive observations, calculations and comparisons that change the next analytical step. This article explains how that difference motivates Marx Max’s shared workspace for discussion, editable code and results. A comparison of baseline and controlled models develops the rationale for staged exploration, human participation and repeatable calculation. It also examines retaining sample and specification information so AI assists with concrete tasks while researchers choose the direction from evidence, with value assessed through overall efficiency.
Keywords: AI research workspace; Research workflows; Human–AI collaboration; Empirical research; Research reproducibility
Read the post →Can AI Reproduce an Empirical Research Paper?
Given an empirical paper’s data and code, can AI reproduce its reported results? The task involves finding the correct entry point, checking data and environments, running programs, and comparing tables and figures. Using the public template maintained by the American Economic Association’s Data Editor team, which oversees data and code checks, this article examines the work AI can assist with and the judgments that require researcher review. A regression discrepancy illustrates how to investigate differences, document modifications, and report unverified results. The article also explains what authors should include in replication packages, and why reproducing calculations is distinct from establishing the validity of a conclusion or replicating it with new data.
Keywords: AI paper reproduction; empirical research replication; AEA Data Editor; computational reproducibility; replication package
Read the post →Where did Marx Max come from?
Marx Max began as a personal research tool: a terminal, Stata-MCP and Zotero-MCP. This article traces how difficulties with configuration and the research process led to the idea of a workspace for other researchers, then recounts the product’s early development and naming. It explains how Karl Marx, Max Weber and Max Planck entered the author’s thinking, and what the name is intended to convey: tools can help find materials, run code and continue a discussion, while research questions, methodological choices and interpretation remain the researcher’s responsibility.
Keywords: Marx Max; AI research workspace; Karl Marx; Max Weber; Scholarship as a vocation
Read the post →