chrisalbon.com

What is Chris Albon?

Chris Albon is a personal website best known as the online home of Chris Albon, a data scientist and author. Its current front page is unusual: rather than a standard blog or portfolio, it presents an interactive project that turns congressional hearing transcripts into a shared semantic map. As you browse, a small language model (all-MiniLM-L6-v2) runs locally in your browser, converting passages into 384-number embeddings. Those embeddings are compared to group similar passages across hearings, saved to a shared collection, and projected onto a two-dimensional map made by visitors. You can inspect the numbers, download the embeddings, or pause the model at any time.

H3 Practical uses

• Explore how remarks on different subjects relate across congressional hearings.
• See a working example of in-browser embeddings and dimensionality reduction.
• Download embeddings for your own analysis or visualization.

H3 Who it suits

• Researchers and journalists looking for a quick way to spot thematic connections in hearing transcripts.
• Data science learners who want to see a small model run client-side.
• Anyone curious about crowdsourced, browser-based data collection.

H3 Trade-offs

• The map reflects only passages processed by visitors, so coverage grows gradually and prioritizes new transcripts.
• The two-dimensional projection is a simplification of 384 dimensions, so close points are suggestive rather than definitive.
• Running the model uses your computer's resources, though you can stop it.

A practical next step: open the site, let a few passages process, then pause and inspect the embedding numbers to judge whether the model's grouping matches your own reading of the transcripts. For related work on data science and machine learning, see Chris Albon.

How does my visit help build a map of congressional hearings?

Your visit is a small act of distributed data processing. While the page is open, a compact language model called all-MiniLM-L6-v2 runs inside your browser and converts passages from congressional hearings into embeddings — lists of 384 numbers that represent a passage's meaning. Those numbers are compared against others to find similar passages across different hearings, and the results appear as points on a shared map. Each passage you complete is added to a collective collection, so the map grows through cumulative visitor effort rather than a central scraping job.

H3 What you actually contribute

• Compute, not content. Your machine does the embedding work; you are not submitting opinions or transcriptions.
• A passage at a time. The site works through the archive incrementally, with new transcripts getting priority as visitors process them.
• A shared result. Completed embeddings are saved to a common collection, and the map is described as "made by its visitors."

H3 Why the design matters

Running the model client-side means the archive can be processed without a single large server doing all the work, and it gives you direct visibility into the mechanics. You can inspect the numbers, download the embeddings, or pause the model whenever you like — useful if you are on a metered connection, a slow device, or simply curious about what the model is doing.

H3 A practical scenario

Suppose you are researching how members of Congress discuss a topic across many hearings. A keyword search only catches exact wording. Embedding-based comparison can surface passages that mean something similar even when the phrasing differs. Your visit is one increment toward that cross-hearing map.

H3 Deciding whether to leave it running

Leave the tab open if you want to contribute and do not mind the background compute. Pause it if you are on battery, working on something demanding, or want to browse without the model running. Either way, the on-page controls let you stop at any point without losing what has already been added.

For more on the person behind the project, see Chris Albon.

Do I need to install anything to run the embedding model on my computer?

No. The page describes the model as running in your browser on your computer while you browse, so there is nothing to download or install to take part. You can also pause it whenever you want, and the page offers downloads of the resulting embeddings if you'd rather work with the data separately.

If you want to check it works for you, open the site and watch the "Preparing the model..." indicator. Once it clears, points should start appearing as passages are processed. If your machine is older or on a metered connection, pause the model and come back later; the page says new transcripts get priority as visitors work through the archive.

A practical way to think about the trade-off:

• Browser-based processing: no install, works on most machines, but speed depends on your hardware and the tab has to stay open.
• Downloading embeddings: better if you want to analyze the data offline or feed it into your own tools, but you get a static snapshot rather than contributing new passages.

For a quick look at the map, just visit the site. For research work, let it run a while first, then download the collection.

Can I download the embeddings or inspect the numbers that my browser generates?

Yes. On Chris Albon, you can inspect the numbers and download the embeddings your browser generates, and you can pause the model at any time.

How it works in practice

• Each passage is turned into a 384-number embedding by all-MiniLM-L6-v2, running locally in your browser.
• The page shows the numbers so you can examine them directly.
• Completed embeddings are saved to a shared collection, so your visit contributes to a collaborative map of congressional hearings.
• New transcripts get priority as visitors work through the archive.
• You retain control: you can stop the model whenever you want.

What you can do with it

• Verify how a particular passage is represented numerically.
• Download embeddings for your own analysis or experiments.
• Compare embeddings to see which passages are semantically close across hearings.

A practical reader scenario

Suppose you’re a researcher studying how members of Congress discuss a topic across different hearings. You could let the page generate embeddings for the passages you care about, download them, and then cluster or search them locally. If you only need a quick visual sense of similarity, the 2D projection on the page may be enough; if you need exact values, inspect or download the numbers.

Decision criterion

Choose inspection and download when you need reproducible data or want to audit the local model’s output. Pause the model when you want to browse without contributing compute or adding passages to the shared collection.

How does the site decide which new transcripts get priority?

New transcripts are prioritized by visitor activity rather than a fixed editorial schedule. According to the site's own description, "new transcripts get priority as visitors work through the archive," and the interface shows a "follow new passages" option alongside a running count of passages added by your visit. The map is built from the passages visitors actually process, so the queue advances as people browse.

In practice, that means the site behaves less like a curated index and more like a crowdsourced work queue: whatever is newest and not yet embedded tends to get attention first, while older material waits until visitors reach it. The site also notes the first passages will appear on the map only after processing begins, so early in a session you may see nothing before results accumulate.

If your goal is to steer what gets covered, the useful next step is simply to spend time on the hearings you care about; your session's embeddings are saved to a shared collection and can be downloaded or paused. A reasonable decision criterion: use the site when you want broad, current coverage of recent hearings and don't mind uneven depth, and look elsewhere if you need a specific older transcript processed on demand.

What do the 384 dimensions and the two-dimensional map actually represent?

The 384 dimensions are the raw output of the all-MiniLM-L6-v2 sentence-embedding model that runs locally in your browser on Chris Albon. Each passage of a congressional hearing is converted into a list of 384 numbers; those numbers encode the passage's meaning as the model has learned it, not human-labeled topics or metadata. Two passages with similar lists are treated as semantically similar.

The two-dimensional map is a separate visualization step: the site projects those 384 numbers down to 2 coordinates so every passage becomes a point you can plot. The projection is a lossy compression for display only — the actual comparison between passages still uses all 384 values. Points that cluster on the map are passages the model judges alike; distance between points is a rough visual cue, not an exact similarity score.

H3 Practical consequences

• You can inspect the numbers and download the embeddings, so the 384-value vectors are reusable data, while the map is a navigational aid.
• Because the map flattens 384 dimensions to 2, some close neighbors in meaning may look far apart, and some visually adjacent points may be only loosely related. Treat the map as an overview, not a measurement.
• The embeddings come from selected congressional members' remarks across subjects, sourced from GovInfo, so the map reflects that corpus rather than all hearings.

H3 A concrete reader scenario

Suppose you are a researcher looking for how different members discussed a topic over time. On the map, you would start near a point you already know is relevant, then open the underlying passages and check their 384-number vectors or similarity values to confirm the match before citing it. If two points look close but the passages differ in substance, trust the passage text and the vector comparison over the plotted distance.

H3 Decision criterion

Use the 384-dimensional embeddings when you need to compute, sort, or compare similarity; use the 2D map when you need to explore the collection visually and form hypotheses. For related public-data work, the official sources include GovInfo and Congress.gov.