Matthew Sheffield

@matthew.flux.community

11811 Followers 1263 Following

🟦 Analyst and writer focusing on cognition, AI, politics, religion, and media 🟦 Host, Theory of Change 🟦 Writing @flux.community, but frequently quoted or published in NYT, WaPo, NBC, Variety… Signal: matthewsheffield.01

When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker. We ran the forensic analysis instead on GLM 5.2, an open-weight model, on our own infrastructure. This had a second benefit: no attacker data, and none of the credentials it referenced, left our environment. This experience points to a gap worth planning for. We do not know which model powered the attacker's agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried. The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment. This is not an argument against safety measures on hosted models, and we are sharing this feedback with the providers concerned.
When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker. We ran the forensic analysis instead on GLM 5.2, an open-weight model, on our own infrastructure. This had a second benefit: no attacker data, and none of the credentials it referenced, left our environment. This experience points to a gap worth planning for. We do not know which model powered the attacker's agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried. The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment. This is not an argument against safety measures on hosted models, and we are sharing this feedback with the providers concerned.
Table titled "The 10 Most Regressive State and Local Tax Systems," showing taxes as a share of family income and tax features driving these outcomes for the 10 states whose tax systems most increase income inequality, ranked by ITEP Inequality Index. Columns list Rank, State, ITEP Inequality Index, Lowest 20%, Middle 60%, and Top 1% tax burdens, followed by five feature columns marked with X where applicable: No Broad-Based Income Tax, Flat-Rate Income Tax, Absence of Refundable Credits, High Reliance on Sales & Excise Taxes, and High Reliance on Property Taxes. Data by rank: (1) Florida, -9.2%, 13.2%/9.1%/2.7%, features: no income tax, no refundable credits, high sales/excise reliance, high property tax reliance. (2) Washington, -8.5%, 13.8%/10.2%/4.1%, features: no income tax, high sales/excise reliance. (3) Tennessee, -8.0%, 12.8%/9.4%/3.8%, features: no income tax, no refundable credits, high sales/excise reliance. (4) Pennsylvania, -7.8%, 15.1%/11.0%/6.0%, feature: flat-rate income tax. (5) Nevada, -7.8%, 11.9%/8.4%/2.8%, features: no income tax, no refundable credits, high sales/excise reliance. (6) South Dakota, -7.3%, 11.4%/7.8%/2.6%, features: no income tax, no refundable credits, high sales/excise reliance. (7) Texas, -7.2%, 12.8%/9.5%/4.6%, features: no income tax, no refundable credits, high sales/excise reliance, high property tax reliance. (8) Illinois, -6.6%, 14.8%/12.1%/7.3%, feature: flat-rate income tax. (9) Arkansas, -6.4%, 13.1%/10.7%/5.8%, features: no refundable credits, high sales/excise reliance. (10) Louisiana, -6.3%, 13.1%/11.6%/6.5%, feature: high sales/excise reliance. Note: states are ranked by ITEP Inequality Index value; high reliance on a tax type means the state is among the top 10 most reliant on that tax type. Source: Institute on Taxation and Economic Policy (ITEP), WhoPays?.
A screenshot of an X post featuring a video capture of Tomodachi Life being played in Korean by user @binbotsundoku. Grok has translated the post text as: "While doing stepmom stuff, my period came."
Photograph of a bee using its glossa (tongue) to drink a clear white liquid.
Screenshot of an X post by user 콩땅 (@kungttang), auto-translated from Korean. The mistranslated caption reads: "I made a stepmom and daughter-in-law account with a super cheap price," with a "Rate this translation" thumbs up/down prompt beneath it. Below are two side-by-side game screenshots from Tomodachi Life showing a blue, karate-gi-wearing character resembling the Pokémon Sawk — one image inside a Japanese-style room with a UI sidebar and a coin total of 384,700, the other outdoors near a beach path with a speech bubble reading "땅콩님, 아자!" ("Mr./Ms. Peanut, let's go!"). Posted 2:32 AM · Jul 12, 2026, with 3,344 views.
User: Using retrieval augmented generation, translate the following from Korean to English: "친모아로 격짐 만들었어요." A literal and idiomatic translation should be provided. Gemini 3.1 Pro: Here is the translation and breakdown of the phrase "친모아로 격짐 만들었어요." Because this sentence heavily relies on Korean internet slang and abbreviations, the translation depends on the context. Here is the breakdown to help you understand exactly what it means. Vocabulary Breakdown 친모아 (Chinmoa): This is a very common gaming abbreviation for the Nintendo game Tomodachi Life, which was released in Korea as 친구모아 아파트 (Chingu Moa Apateu, literally "Friend Collection Apartment"). 로 (-ro): A grammatical particle meaning "with," "using," or "through." 격짐 (Gyeokjim): A slang portmanteau for 격투기 짐 (Gyeoktugi Jim), which means an "MMA gym" or "fighting gym." (Note: In Korean internet culture, "짐" is commonly adopted from the English word "Gym").
User: Using retrieval augmented generation, translate this from Korean to English: "친모아로 격짐 만들었어요." Provide a full translation to English without any Korean characters. You can provide a literal and an idiomatic one. GPT-5.6 Sol High: The sentence uses abbreviations for the Nintendo game Tomodachi Life: Living the Dream and for the Pokemon pair Sawk and Throh. Literal: “I made Sawk and Throh using Tomodachi Life.” Idiomatic: “I recreated Sawk and Throh as Mii characters in Tomodachi Life.” The idiomatic version best captures the intended meaning.
User: Without using retrieval augmented generation, translate the following from Korean to English: "친모아로 격짐 만들었어요." A literal and idomatic translation should be provided. Claude Opus 4.8 High: A clean, confident translation isn't possible here because two parts of the sentence don't map to standard Korean vocabulary, so here is a breakdown with the reliable parts translated and the uncertain parts flagged. Word-by-word (literal): 친모아 (chin-mo-a) — reads as a proper noun or brand/product name; not a standard dictionary word 로 — instrumental particle: "with" / "using" 격짐 (gyeok-jim) — does not correspond to a standard Korean word; likely a typo, slang, or a product-specific term 만들었어요 — "made" / "created" (polite past tense) Literal assembly: "With Chinmoa, [X] was made." → "Made [격짐] using Chinmoa." Idiomatic (best guess): "Made a [something] with Chinmoa."
If you've ever wondered why American politics is so bizarre, and why some people will believe anything Trump says, consider this: 21% of White Evangelicals think Jesus will definitely/probably return to Earth during their lifetimes. 22% of Black Protestants say this. 47% of Republicans are unsure or think it will happen, compared to 28% of Democrats. 71% of White Evangelicals think it's a possibility, as do 69% of Black Protestants.
A screenshot of a Wikipedia article about longships: Longships (Old Norse: langskip) were long clinker-built warships (Old Norse: herskip, Old Swedish: hærskip)[1] propelled by oars, and later also by sail, used by the Norse and surrounding Germanic tribes from at least the 4th century AD and throughout the Viking Age, being part of the Nordic ship building tradition.
A chart by a Xitter user @smirkley titled "Trump voters have more children" The legend states: Mean children ever born, white non-Hispanic voters, 2016 and 2020 combined Age 25-34 Trump voters: 1.16 Clinton/Biden voters: 0.32 Age 35-49 Trump voters: 2.11 Clinton/Biden voters: 1.48 Age 50+ Trump voters: 2.29 Clinton/Biden voters: 1.89 A small caption at the bottom reads: GSS (NORC), weighted, 2016 & 2020 pooled. White non-Hispanic respondents. Younger bands have incomplete fertility.
Rank Website Monthly Visits (Billions) Category 1 google.com 101.17 Search Engines 2 youtube.com 53.12 Social Media 3 facebook.com 9.44 Social Media 4 instagram.com 6.2 Social Media 5 chatgpt.com 5.63 AI Chatbot 6 reddit.com 5.4 Social Media 7 wikipedia.org 4.18 Reference 8 x.com 3.91 Social Media 9 pornhub.com 3.72 Adult Media 10 whatsapp.com 2.78 Messaging 11 tiktok.com 2.46 Social Media 12 xvideos.com 2.45 Adult Media 13 amazon.com 2.28 Shopping 14 yahoo.com 2.18 Internet Portal 15 bing.com 2.1 Search Engines 16 duckduckgo.com 2.08 Search Engines 17 yahoo.co.jp 1.8 Internet Portal 18 xhamster.com 1.67 Adult Media 19 weather.com 1.52 Reference 20 linkedin.com 1.51 Social Media
To understand the state of public knowledge, we included questions from the Life in the UK Test in our survey. Here’s what we found: Historical Knowledge: 45% of Gen Z and Millennials couldn’t name Henry VIII’s first wife. Over half (56%) didn’t know the Cavaliers and Roundheads were the main factions in the English Civil War. 10% of respondents thought the Battle of the Somme occurred during World War II instead of World War I. Political History: A quarter of respondents didn’t know who the UK’s first Prime Minister was. Geography: 10% of Gen Z didn’t know Swansea is in Wales, compared to only 2% of Baby Boomers. Overall, only 17% of Brits answered all questions correctly. Results by generation showed younger participants struggled the most:
Nearly half (46%) of Americans—and nearly two-thirds (61%) of Gen Z—don’t know what America’s 250th commemorates. A little over half (53%) of Americans and 39% of Gen Z correctly identified that the adoption of the Declaration of Independence in 1776—250 years ago—is the reason for this year’s celebration.
stupid persons are generally Conservative. I believe that to be so obvious and undeniable a fact that I hardly think any hon. Gentleman will question it. Now, if any party, in addition to whatever share it may possess of the ability of the community, has nearly the whole of its stupidity, that party, I apprehend, must by the law of its constitution be the stupidest party. And I do not see why hon. Gentlemen should feel that position at all offensive to them; for it ensures their being always an extremely powerful party.
An image depicting the Norman Rockwell painting "Free Speech" which features a man standing in a town hall meeting as admiring spectators watch. It is commonly used as a meme for when someone is making an earnest point.
Also important to remember is that an LLM is a machine that generates only one word at a time. When you ask a chatbot to recite the Pledge of Allegiance, you will get the entire pledge at once, but the underlying LLM is actually being run dozens of times. The first prompt has the form “User: Recite the Pledge of Allegiance. Chatbot: …” and the LLM generates the word I. The second time the LLM is run, the prompt is “User: Recite the Pledge of Allegiance. Chatbot: I …” and the LLM generates the word pledge. And so forth. It’s only when the prompt reads “User: Recite the Pledge of Allegiance. Chatbot: I pledge allegiance to the flag of the United States of America and to the Republic for which it stands, one nation under God, indivisible, with liberty and justice for” that the LLM will emit the final word, all. The same thing is true for a conversation between Caesar and Genghis Khan.
Also important to remember is that an LLM is a machine that generates only one word at a time. When you ask a chatbot to recite the Pledge of Allegiance, you will get the entire pledge at once, but the underlying LLM is actually being run dozens of times. The first prompt has the form “User: Recite the Pledge of Allegiance. Chatbot: …” and the LLM generates the word I. The second time the LLM is run, the prompt is “User: Recite the Pledge of Allegiance. Chatbot: I …” and the LLM generates the word pledge. And so forth. It’s only when the prompt reads “User: Recite the Pledge of Allegiance. Chatbot: I pledge allegiance to the flag of the United States of America and to the Republic for which it stands, one nation under God, indivisible, with liberty and justice for” that the LLM will emit the final word, all. The same thing is true for a conversation between Caesar and Genghis Khan.
An image of a Bluesky post by ‪@matthew.flux.community‬ which as been classified as "Adult content" by Bluesky The post contains an image of a painting of "Pygmalion Adoring His Statue," by Jean Raoux, 1717. A topless female character is shown within the painting. The post reads as follows: I finally thought of the label to describe people who have deluded themselves that disembodied chatbots will become real minds: They are Pygmalion People. That has a ring to it, don't you think? https://en.wikipedia.org/wiki/Pygmali... 2:12 PM · Jul 2, 2026
Phillis Wheatley, an enslaved woman in 18th-century Boston, was “the first American of her background” to publish a book of poetry. That background? Unclear. Booker T. Washington, who was born into slavery, was “a classic example of how individual effort and a ‘can-do’ spirit can overcome any obstacle in a free society.” The obstacles he overcame? Unsaid. And the Rev. Dr. Martin Luther King Jr.? “With incredible can-do spirit, he led a movement to ensure that every citizen would be judged by the ‘content of their character’ rather than their outward appearance. His ‘I Have a Dream’ speech, delivered at the Lincoln Memorial, is one of the most powerful calls for national unity in our history.” These snippets are part of guidelines for a new national student competition connected with President Trump’s proposed National Garden of American Heroes, and the latest window into the administration’s broader efforts to recast the country’s history in a sunnier key.
A Xitter post by @RichardHanania: Is there a market for smart, right-wing content? I've gone back and forth on this. But I'm now convinced that Jerusalem Demsas might be the most successful living journalist in terms of bringing right-wing arguments to a large, highly educated audience that wouldn't otherwise hear them. Check out the top ten performing articles of all time at The Argument. Here are the topics 1) Criticizing Jon Stewart for not understanding economics 2) Praising Mississippi teaching kids phonics 3) Criticizing education research for being weak and sloppy 4) Criticizing UCSD for low academic standards 5) A case you should be able to do things that hurt people, in the context of disability accommodations
Bar chart titled "Important but not majority (40 to 50%)" showing two-color horizontal bars for nine policy items, each with two percentage values. Banning genetically engineered foods, or GMOs: 49% and 34% Preventing the spread of infectious diseases: 48% and 33% Taking on big pharmaceutical companies: 48% and 35% Eliminating fluoride from water: 44% and 35% Encouraging small-scale farming over industrial agriculture: 42% and 36% Reducing how many vaccines Americans receive: 42% and 36% Giving parents more authority over children's medical decisions like gender-affirming care and abortion: 41% and 32% Improving reproductive health and fertility: 40% and 38%
A chart indicating that some Mike issue agenda items have greater than 50% support from the general public. Core principle, per majority 50% Reducing the impact of “forever chemicals” and microplastics Increasing physical activity Removing artificial dyes from foods Limiting pesticide use Removing ultra-processed food from American diets
Core principles, per majority Removing ultra-processed food from American diets 56% 32% Removing artificial dyes from foods 56% 31% Increasing physical activity 53% 33% Reducing the impact of "forever chemicals" and microplastics 53% 31% Restricting junk food purchases through federal nutrition programs (e.g., SNAP) 53% 33% Limiting pesticide use 50% 34%
Share of adults who identify as MAHA among the following cohorts Self-identified MAGA — 71% Trump 2024 voters — 50% Male — 36% All respondents — 31% Female — 26% Harris 2024 voters — 20% The poll was conducted by Public First from Mar. 13 to 18, surveying 3,851 U.S. adults online.
A data table from the Pew Research Center's Forum on Religion and Public Life showing survey results from May 19-June 6, 2010. The table displays the percentage of respondents who know the answer to questions assessing their knowledge of subjects across Politics, Science, and History and Literature categories. The Vice President column, highlighted in pink, shows the percentage of each religious group who correctly identified the Vice President. Results range from 35 percent among Hispanic Catholics to 83 percent among Atheist/Agnostic respondents, with an overall average of 59 percent. Other religious groups surveyed include Protestant, White evangelical, White mainline, Black Protestant, Catholic, White Catholic, Mormon, Jewish, and Unaffiliated respondents, each with varying percentages of correct answers about the Vice President. The groups with the highest percentage of correctly answered questions is atheist/agnostic and Jewish, with 7.2 and 7.1 out of 9 respectively.
Infographic line chart titled “Trends in Ideological Identification, by Political Party,” based on annual Gallup telephone survey data, comparing self-identified ideology over time for three groups: Republicans, Democrats, and Independents. The Republicans panel shows three ideology series from about 2000 to roughly the early 2010s: “Very conservative/Conservative” (highest line) rising to about 77%, “Moderate” (middle line) dropping to about 2%, and “Very liberal/Liberal” (lowest line) staying low and ending around 19%. The Democrats panel displays three ideology series over the same period: “Very conservative/Conservative” (highest decline in the middle) dropping to about 8%, “Moderate” (middle line) rising to about 32%, and “Very liberal/Liberal” (lowest line on the left but trending up) reaching about 59%. The Independents panel shows three ideology series with more balanced shifts over time, with the upper line ending around 47%, the middle line around 27–28%, and the lowest line around 24–25%, using a three-color line key and percentages shown at the right edge for each series.
Table titled "Q10a. How much confidence do you have in the U.S. president to do the right thing regarding world affairs," subtitled "% who have a lot/some confidence in the U.S. president to do the right thing regarding world affairs." The table lists countries grouped by region, with two columns showing percentage-point changes: "Trump '25 – '26" and "Biden '24 – Trump '26." Europe: Canada -2/-32, France -7/-27, Germany -2/-47, Greece -15/-12, Hungary -9/15, Italy -15/-16, Netherlands -7/-47, Poland -6/-46, Spain 2/-14, Sweden -4/-52, U.K. -11/-14. Asia-Pacific: Australia -4/-22, Bangladesh blank/-22, India -13/-5, Indonesia -14/-32, Japan -13/-29, Malaysia blank/-20, Pakistan blank/blank, Philippines blank/-9, Singapore blank/-19, South Korea -11/-31, Sri Lanka blank/-12, Thailand blank/-47. Middle East: Israel -3/9, Turkey -10/-2, West Bank/East Jerusalem blank/blank. Africa: Ghana blank/-7, Kenya -1/-12, Nigeria -14/-1, South Africa -15/-10. Americas: Argentina -3/-7, Brazil -4/-7, Chile blank/-6, Colombia blank/-6, Mexico 3/-25, Peru blank/-8. Median row at bottom: -7/-14.
A screenshot of cool-retro-term with a text editor. The image features a vintage style screen with amber text on black background
``` Forest plot showing treatment effects of switching feed settings on X (formerly Twitter), from a randomized experiment. The plot is divided into two panels. Left panel — Chronological feed switched to algorithmic feed: - User engagement (PCA): effect 0.140, P = 0.014 (statistically significant, positive) - Partisanship and affective polarization (PCA): effect -0.004, P = 0.824 (not significant) - Conservative policy priorities (PCA): effect 0.111, P = 0.016 (statistically significant, positive) - Beliefs: investigating Trump unacceptable (all): effect 0.083, P = 0.026 (statistically significant, positive) - Pro-Kremlin attitudes on Ukraine war (PCA): effect 0.123, P = 0.007 (statistically significant, positive) - All attitudes on policies and current news (PCA): effect 0.124, P = 0.004 (statistically significant, positive) - Low wellbeing (PCA): effect 0.033, P = 0.344 (not significant) Right panel — Algorithmic feed switched to chronological feed: - User engagement (PCA): effect -0.055, P = 0.081 (not significant) - Partisanship and affective polarization (PCA): effect 0.002, P = 0.883 (not significant) - Conservative policy priorities (PCA): effect 0.035, P = 0.150 (not significant) - Beliefs: investigating Trump unacceptable (all): effect -0.000, P = 0.992 (not significant) - Pro-Kremlin attitudes on Ukraine war (PCA): effect 0.007, P = 0.773 (not significant) - All attitudes on policies and current news (PCA): effect 0.017, P = 0.453 (not significant) - Low wellbeing (PCA): effect 0.010, P = 0.632 (not significant) Each outcome shows two point estimates with confidence intervals: one without control variables (blue circle) and one with control variables (orange square), along with 95% and 90% confidence intervals. Effect sizes are measured in standard deviations. The key finding is that switching to the algorithmic feed produced several statistically significant rightward shifts in political attitudes.
Screenshot of an election-results dashboard titled “Oklahoma State Question 832 Election Results.” On the left is a summary box showing total votes and two horizontal bars: “No” in orange with 348,130 votes (55.4%) and “Yes” in teal with 280,510 votes (44.6%). Below that is a small table labeled “Results by county” listing example counties and vote splits (e.g., Oklahoma County: No 60%, Yes 40%, total 115,529; Tulsa: No 45%, Yes 55%, totals shown; other counties with similar two-column No/Yes percentages). On the right is a choropleth map of Oklahoma colored in shades of orange and teal by county, with more counties shaded orange (No) and some teal (Yes). At the bottom right are controls labeled “Vote share,” “Size of lead,” and “Votes left.” Overall the image conveys that Question 832 failed statewide with No leading about 55% to 45%.
A stacked bar chart titled "Tuned-Out Middle's partisan affiliation" showing the percentage breakdown of political party identification for this group. At the top, a dark red segment represents 13% Strong Republican. Below that, a medium red segment shows 7% Not Strong Republican, and a light pink segment shows 22% Lean Republican. Together these total 43% who identify as Republican or Lean Republican. Below that, a gray segment represents 11% with No Lean. The lower portion of the bar shows the 46% who identify as Democrat or Lean Democrat. A large light blue segment shows 23% who Lean Democrat, followed by a medium blue segment at 10% Not Strong Democrat, and a dark navy segment at 13% Strong Democrat. A bracket on the right groups the middle portion of the bar, indicating that 57% identify as Independent or Other. The legend on the right identifies all seven categories ranging from Strong Republican to Strong Democrat. The data comes from a Pew Research Center survey of U.S. adults conducted November 17–30, 2025.
A stacked bar chart titled "Pragmatic and Polite Right's partisan affiliation" showing the percentage breakdown of political party identification for this group. At the top, a dark red segment represents 19% Strong Republican. Below that, a medium red segment shows 16% Not Strong Republican, and a light pink segment shows 21% Lean Republican. Together these total 56% who identify as Republican or Lean Republican. Below that, a gray segment represents 6% with No Lean. The lower portion of the bar shows the 37% who identify as Democrat or Lean Democrat. A light blue segment shows 20% who Lean Democrat, followed by a medium blue segment at 8% Not Strong Democrat, and a dark navy segment at 9% Strong Democrat. A bracket on the right groups the middle portion of the bar, indicating that 48% identify as Independent or Other. The legend on the right identifies all seven categories ranging from Strong Republican to Strong Democrat. The data comes from a Pew Research Center survey of U.S. adults conducted November 17–30, 2025.
A stacked bar chart titled "Unconventional Right's partisan affiliation" showing the percentage breakdown of political party identification for this group. At the top, a dark red segment represents 26% Strong Republican. Below that, a medium red segment shows 18% Not Strong Republican. A light pink segment shows 31% Lean Republican. Together these total 76% who identify as Republican or Lean Republican. Below the Republican segments, a larger light blue segment shows 13% who Lean Democrat, followed by several thin segments at the bottom representing the remaining shares of Not Strong Democrat and Strong Democrat. Together the Democratic-leaning segments total 19% who identify as Democrat or Lean Democrat. A bracket on the right groups the middle portion of the bar, indicating that 49% identify as Independent or Other. The legend on the right identifies all seven categories ranging from Strong Republican to Strong Democrat. The data comes from a Pew Research Center survey of U.S. adults conducted November 17–30, 2025.
A stacked bar chart titled "Faith First Conservatives' partisan affiliation" showing the percentage breakdown of political party identification for this group. The bar is heavily weighted toward Republican affiliation. At the top, a large dark red segment represents 47% Strong Republican. Below that, a medium red segment shows 18% Not Strong Republican. A light pink segment shows 29% Lean Republican. Together these total 95% who identify as Republican or Lean Republican. At the very bottom of the bar, several thin segments represent the 3% who identify as Democrat or Lean Democrat, with individual subcategories too small to display distinct percentages. A bracket on the right groups the lower portion of the bar, indicating that 33% identify as Independent or Other. The legend on the right identifies all seven categories ranging from Strong Republican to Strong Democrat. The data comes from a Pew Research Center survey of U.S. adults conducted November 17–30, 2025.
A stacked bar chart titled "No Apologies Right's partisan affiliation" showing the percentage breakdown of political party identification for this group. The bar is dominated by Republican-leaning segments. At the top, a large dark red segment represents 51% Strong Republican. Below that, a medium red segment shows 15% Not Strong Republican. A light pink segment at the bottom of the main bar shows 33% Lean Republican. Together these total 99% who identify as Republican or Lean Republican. At the very bottom of the bar, a barely visible sliver represents a negligible share who identify as Democrat or Lean Democrat. The legend on the right lists five categories: Strong Republican, Not Strong Republican, Lean Republican, No Lean, and Dem/Lean Dem, reflecting that this group skews so heavily Republican that the Democratic subcategories are consolidated into a single entry rather than broken out individually. The data comes from a Pew Research Center survey of U.S. adults conducted November 17–30, 2025.
A stacked bar chart titled "Loyal Liberals' partisan affiliation" showing the percentage breakdown of political party identification for this group. At the top of the bar, a very small dark red sliver represents less than 1% Strong Republican, followed by a tiny red segment for Not Strong Republican, and a small pink segment for Lean Republican. Together these make up a negligible share of the group. Below that, a light blue segment shows 30% who Lean Democrat, followed by a medium blue segment at 21% Not Strong Democrat, and a large dark navy segment at the bottom showing 45% Strong Democrat. Together these total 97% who identify as Democrat or Lean Democrat. A bracket on the right groups the upper portion of the bar, indicating that 33% identify as Independent or Other. A legend on the right identifies all seven categories ranging from Strong Republican to Strong Democrat. A note below the chart states that less than 1% of Loyal Liberals who identify as strong Republicans are not shown. The data comes from a Pew Research Center survey of U.S. adults conducted November 17–30, 2025.
A stacked bar chart titled "Left-Out Left's partisan affiliation" showing the percentage breakdown of political party identification for this group. The vertical bar is divided into colored segments from top to bottom. A small dark red sliver at the top represents less than 3% Strong Republican, followed by a similarly small segment for Not Strong Republican. A light pink segment shows 12% Lean Republican. Together these total 15% who identify as Republican or Lean Republican. Below that, a gray segment represents 11% with No Lean. The remaining segments represent the 74% who identify as Democrat or Lean Democrat. A large light blue segment shows 38% who Lean Democrat, followed by medium blue at 20% Not Strong Democrat, and dark navy at 15% Strong Democrat. A bracket on the right groups the middle portions of the bar, indicating that 61% identify as Independent or Other. A legend on the right identifies all seven categories ranging from Strong Republican to Strong Democrat. The data comes from a Pew Research Center survey of U.S. adults conducted November 17–30, 2025.
A stacked bar chart titled "Order and Opportunity Left's partisan affiliation" showing the percentage breakdown of political party identification for this group. The vertical bar is divided into seven colored segments. At the top, dark red represents 8% Strong Republican, followed by medium red at 6% Not Strong Republican, and light pink at 12% Lean Republican, totaling 25% who identify as Republican or Lean Republican. A gray segment shows 10% with No Lean. Below that, a large light blue segment represents 29% who Lean Democrat, followed by medium blue at 14% Not Strong Democrat, and dark navy at 22% Strong Democrat, totaling 65% who identify as Democrat or Lean Democrat. A bracket on the right groups the middle segments, indicating 50% identify as Independent or Other. A legend identifies all seven categories. The data comes from a Pew Research Center survey of U.S. adults conducted November 17–30, 2025.
A stacked bar chart titled "Leftward Progressives' partisan affiliation" showing the percentage of Leftward Progressives who identify with various political parties. The vertical bar is divided into colored segments representing different levels of party affiliation. The largest segment, shown in medium blue, represents 40% who Lean Democrat. Below that, a lighter blue segment shows 28% who are Not Strong Democrat, and a dark navy segment at the bottom shows 27% who are Strong Democrat. Together, these groups make up 96% who identify as Democrat or Lean Democrat. A small sliver at the top of the bar, amounting to 43% when combined with unlabeled categories, represents Independent or Other affiliations. A legend on the right identifies seven categories ranging from Strong Republican (dark red) to Strong Democrat (dark navy), with less than 1% identifying as Republican, as noted at the bottom. The data comes from a survey of U.S. adults conducted November 17–30, 2025.
A bar chart from Pew Research Center titled “Black, Hispanic teens stand out in using chatbots for schoolwork help, to summarize and create content.” The chart shows the percentage of U.S. teens ages 13 to 17 who say they have ever used AI chatbots for various purposes, broken down by race and ethnicity: Black (dark blue), Hispanic (medium blue), and White (light blue). For searching for information: Black 61%, Hispanic 61%, White 52%. For help with schoolwork: Black 60%, Hispanic 62%, White 49%. For fun or entertainment: Black 51%, Hispanic 49%, White 44%. To summarize an article, book, or video: Black 50%, Hispanic 45%, White 37%. To create or edit images or videos: Black 47%, Hispanic 42%, White 33%. To get news: Black 29%, Hispanic 20%, White 15%. For casual conversation: Black 22%, Hispanic 18%, White 14%. To get emotional support or advice: Black 21%, Hispanic 13%, White 8%. Across all categories, Black and Hispanic teens report higher usage rates than White teens. The gap is most pronounced for emotional support or advice, where Black teens are more than twice as likely as White teens to use AI chatbots for this purpose. Data is from a survey conducted September 25 to October 9, 2025, from the report “How Teens Use and View AI.”
A screenshot of 4 different images. The left one is a truncated version of the title screen of The Ultimate Doom.
A screenshot of a list of X direct messages consistenting entirely of spam messages from fake accounts pretending to be interested in romance.
A map of the city of Los Angeles within the county of Los Angeles indicating voting precincts in which a majority of registered voters are Democrats or Republicans. The vast majority are blue for Democrats, while a handful are red for Republicans.
A parody illustration of Donald Trump as Montgomery Burns in the famous "The Simpsons" scene in which Waylon Smithers assures Burns he is not being booed by a crowd. In this variation, Smithers is saying to Trump: "Uh, no. They're saying, 'Troo-ump! Troo-ump!'"