roadtolarissa.com

What is roadtolarissa?

roadtolarissa is the personal data-visualization site of Adam Pearce, a graphics editor whose work turns public datasets and current events into interactive charts and maps. The homepage is essentially a chronological index of projects, organized by year, with topics that range from elections and hurricanes to machine-learning concepts.

What you'll find there

The work falls into a few recurring categories:

• News-driven graphics and maps. Live election maps, hurricane trackers and rainfall maps, health-care vote counts, and similar pieces built to be updated as events unfolded.
• Exploratory data pieces. Analyses of tax proposals, mobility, fairness and bias, prisons, and sports probabilities — often with an emphasis on uncertainty and how a chart could mislead.
• Technical and machine-learning explainers. Visual essays on topics such as UMAP, BERT, scaling, interpretability, and circuit tracing, aimed at making abstract methods visible.
• Small tools and experiments. Interactive widgets, chart-making utilities, and annual side projects.

Who it's for

It suits three overlapping audiences: journalists and graphics editors looking for craft examples; data scientists and analysts interested in how complex models get explained visually; and general readers who want to explore a topic through interaction rather than a static article.

How to approach it

Because the site is an archive rather than a course, start from the topic you care about rather than browsing chronologically. Pick one recent news graphic and one technical explainer to see the range: the first shows how to design for live, changing data; the second shows how to make a model's behavior legible. A useful decision criterion is whether you want reproducible technique or conceptual intuition — the news maps lean toward the former, the machine-learning pieces toward the latter.

For related work in the same field, see Observable and The Pudding.

What kinds of data visualizations and chart topics are covered on roadtolarissa?

roadtolarissa is Adam Pearce's personal portfolio of data visualizations. It is not a chart-tool site or a gallery of generic chart types; it is a body of work where the chart is the argument. Topics cluster around public data, uncertainty, and the craft of showing data honestly. The site reads best as a set of case studies: each piece takes a messy public dataset and asks what a reader can actually conclude from it.

Recurring chart topics

• Elections and politics: primary and general-election polling, electoral maps, turnout, voter registration, campaign finance, congressional votes, and trackers of political statements and agendas.
• Weather, climate and hazards: hurricane maps and rainfall, arctic sea ice, summer heat extremes, wildfires, and live flood tracking.
• Public health and COVID-19: wastewater signals, symptom search trends, mobility, and fairness or bias in health data.
• Sports: playoff probabilities, NBA finals streaks, baseball launch angle, and tournament group scenarios.
• Economics and everyday data: tax proposals and calculators, box office hits, television ads, and personal data such as Kindle reading.
• Machine learning and interpretability: scaling and grokking, circuits, knowledge in models, and audits of alignment claims.

Chart forms and techniques

The visual vocabulary is broad and deliberately matched to the question. You will find maps (choropleths, dot maps, flow maps), line and area charts, small multiples, slope and bump-style comparisons, uncertainty intervals, regression discontinuity plots, and animated transitions. Several pieces are explicitly about uncertainty: showing ranges, calibration, or what a forecast cannot tell you. Others are about scale and distortion, such as colormap curves or the choices behind a map projection. A recurring thread is interpretability, both of models and of charts: how a reader can be misled by a plausible-looking graphic.

What this means in practice

If you are looking for a specific chart type, this is a poor index and a good source of examples. If you are trying to learn how to present a forecast, a map, or a politically charged dataset without overclaiming, the relevant pieces are the ones that show their uncertainty and explain their choices. For adjacent work in the same spirit, see Observable for notebook-based charting and The Pudding for visual essays. A useful next step is to pick one topic you already know well, such as hurricane rainfall or election polling, and read the corresponding piece to see how the chart's design changes what the data appears to say.

How can I use roadtolarissa charts to understand election forecasts and polling uncertainty?

Adam Pearce's roadtolarissa charts are interactive data visualizations that treat election forecasts as ranges, not single numbers. The site's archive includes several election-related pieces—"live forecast," "playoff probabilities," "polls," "electoral map," "senate seats in play," and "voter registration" among them—so the practical value is in seeing how uncertainty is drawn, not in getting a current prediction.

What the charts let you do

• See the spread, not just the line. Forecast graphics typically show a central estimate plus a band or distribution of outcomes. Reading the band's width tells you how much the model itself expects to be wrong.
• Compare scenarios. Interactive controls (time, state, demographic group) let you isolate one variable and watch how the forecast shifts.
• Check calibration. Some pieces, such as "uncertainty calibration," are about whether stated probabilities matched observed results—useful for judging any forecaster's confidence claims.
• Track change over time. A "live forecast" style chart shows how estimates moved as new polls arrived, which is often more informative than the final snapshot.

A concrete reader scenario

Suppose you see a Senate race described as "a toss-up." A roadtolarissa-style chart would show you the distribution of simulated outcomes: how often each candidate wins, how many simulations were near-ties, and how sensitive the result is to one state. You would then know whether "toss-up" means genuinely 50/50 or just "we don't know."

How to read uncertainty well

Next step

Open one election chart and change a single control—time or geography—then note how much the outcome distribution moves. If a small input change produces a large output change, treat the forecast as fragile. For official election data to compare against, see U.S. Census Bureau and Federal Election Commission; for polling methodology generally, Pew Research Center and AAPOR publish standards that help you judge whether a chart's uncertainty is being presented honestly.

What interactive tools or visualizations does roadtolarissa offer for exploring tax policy or health care votes?

Adam Pearce’s roadtolarissa is a personal portfolio of data visualizations rather than a tax or health-policy product. Its relevant interactive work is clustered around two topics: a 2017 tax calculator and tax-cut chart diary, and health care vote trackers from the same year.

Tax policy

• A 2017 tax calculator. The page evidence lists “tax calculator 2017” alongside “every tax cut,” suggesting an interactive tool for estimating how a proposed tax plan would affect a household. That is the most directly useful item for a reader who wants to test a plan against their own income assumptions.
• A chart diary on tax cuts. “2017 chart diary” and “every tax cut” point to a visual, sequential record of tax-cut proposals or provisions rather than a single static chart. This suits someone following a fast-moving bill who wants to see how the numbers changed over time.
• “Trumps tax proposal cost.” This appears to be a cost-focused visualization of the proposal, useful for comparing the scale of cuts or revenue effects.

Health care votes

• “Health care votes v1” and “v2.” Two versions of a health care vote visualization indicate an interactive roll-call tracker, likely mapping members or votes. The versioning suggests Pearce revised the display as the legislation or data changed.
• “Health care whip count.” A whip count is the practical tool for seeing where support stood at a given moment. For a reader tracking a close vote, this is more actionable than a final vote tally because it shows the state of play before the vote.
• “Congress statements comey” and “health care votes.” The juxtaposition hints at linking lawmaker statements to votes, useful for accountability reporting.

How to use this

These are exploratory, journalistic graphics, not policy-modeling software. The tax calculator is best for quick scenario testing; the chart diaries and whip counts are best for following a bill’s evolution. If you need official numbers, cross-check against the Congressional Budget Office or the relevant committee’s records, since the visualizations are interpretive layers on top of source data.

A sensible next step: start with the tax calculator if your question is “what does this mean for me,” and start with the whip count if your question is “will it pass.” For broader context on how these graphics fit into data journalism, see roadtolarissa and, for official legislative data, Congress.gov.

How does roadtolarissa approach visualizing hurricanes, flooding, and other natural disaster data?

Adam Pearce's disaster visualizations on roadtolarissa tend to treat a storm or flood as a data-processing problem first and a design problem second: gather messy, fast-moving feeds, clean and join them, then choose a chart form that lets readers locate themselves or a place they care about. The recurring subjects — Hurricane Harvey, Hurricane Irma, Hurricane Florence, Hurricane Michael, the 2017 Arctic sea-ice maximum, Australian fires — cluster around live or near-live maps and record-tracking charts rather than static infographics.

What the recurring work suggests

• Maps as the default frame for hazards. Storms, rainfall, flooding and mobility are inherently spatial, so the map carries the primary message; accompanying charts usually handle magnitude, records or uncertainty.
• Live and forecast modes. Several projects pair an observed record with a forecast or probability view, which is the honest way to present a hazard while it is still unfolding.
• Uncertainty as a first-class element. Pearce has repeatedly worked on uncertainty calibration and "uncertainty over space," so ranges, cones and probabilities appear rather than single confident lines.
• Comparison to history. Record-framing ("Irma records," "Arctic ice maximum") gives an absolute event a relative meaning, which is what non-specialist readers actually need.
• Human-scale layers. "Houston cries for help," mobility data and neighborhood-level views show where people are, not just where water is.

A practical decision guide

A useful next step: pick one hazard you follow and sketch two views — a "where" map and a "how unusual" chart — before writing any code. If the map cannot answer "am I in it?" and the chart cannot answer "is this normal?", the pairing needs rework. For context on how other outlets handle the same trade-offs, see The New York Times and FiveThirtyEight, both of which have published comparable live storm and flooding coverage.

What techniques does roadtolarissa use for explaining machine learning interpretability and alignment?

Adam Pearce uses interactive, visual explainers to make machine learning interpretability and alignment tangible. On roadtolarissa, his approach pairs small multiples, annotated diagrams and explorable charts with plain-language framing, so abstract concepts become something a reader can inspect, hover over and compare.

What that looks like in practice

• Circuit tracing and monosemanticity. He builds visual walkthroughs of how features and circuits inside a model can be isolated and traced. The emphasis is on showing the mechanism step by step rather than asserting that a model is interpretable.
• Alignment audits. He treats alignment as an empirical question: what was checked, what was found, and where the evidence is thin. Charts carry the argument, with annotations marking the specific claim being tested.
• Interactive explainers over static posts. Readers can adjust inputs or step through stages, which lets them see how conclusions change when assumptions change.
• Visual metaphors with limits. Diagrams simplify, but the surrounding text usually notes where the metaphor breaks down, which matters for interpretability work where overclaiming is the main risk.

Who this suits

A researcher who wants a quick visual intuition before reading a paper, a policy or communications person who needs a defensible summary, or a technically curious reader who learns better by manipulating a chart than by reading notation. It is less suited to someone who needs full derivations or reproducible code-level detail; the value is in framing and visual reasoning, not exhaustive methodology.

How to use it

Pick one explainer on a topic you already know well and check whether the simplification holds up; that tells you how much to trust the others. If you want a broader set of interactive ML explainers, look at Distill and PAIR, which use a similar visual, article-length format.