Wall Street sell-side analysts are now unmoored from reality. Next four-year S&P500 nominal earnings growth totalling exp(0.23 x 4) - 1 = 150%? The technical term is “nuts”…

These are “consensus” nominal earnings growth estimates. Part of it is that a bunch of forecasters do expect an inflation breakout approaching the high single if not low double digits, and their expectations do move the average up. But that is not most of it.

Share

Most of it is that the people making the earnings growth estimates have gone gonzo—or, rather, that heavy pressure is being applied to sell-side analysts by bosses who strongly feel that coming in with a lower earnings-growth estimate than others will be seen as a downer by clients who would otherwise spend and invest like drunken sailors. Exactly whose sentiment—or desire for permission to buy assets—is driving this is very unclear:

Give a gift subscription

From:

Callum Thomas: Earnings Euphoria <https://www.chartstorm.info/p/chart-of-the-week-earnings-euphoria>: This week’s chart pretty much speaks for itself (+has just made a new all-time high)…. The estimated annualized compound growth rate expected by sell-side analysts over a three to five-year horizon aggregated for the S&P500…. Wall Street analyst sentiment. And like all good sentiment indicators there are elements of truth and elements of emotion all mixed up in it…. Some will argue that there is good reason for this, and as I noted they will be right that some of it is well-reasoned and true. But there is clearly a powerful sentiment and story dynamic…

The Weekly ChartStormChart of the Week - Earnings EuphoriaNote: this is a new series for free subscribers and is intended to give an insight into the work that goes into The Weekly ChartStorm + provide a little further explanation on charts and concepts…Read more15 days ago · 42 likes · Callum Thomas

Get 75% off a group subscription

Corporate profits in the US right now are at a record high of 22% of of national income. (Labor 65%, other—rent, net interest, indirect business taxes, proprietors’ income—13%.) At 3% inflation, 0% (because no net immigration) labor-force growth, and 2%/year productivity growth, these forecasts give us a profit share of 46% in four years, and a labor share of 41%: real wage declines averaging more than 1/4 of current levels.

Subscribe now

Leave a comment

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…##earnings-sentiment-or-this-is-nuts-whens-the-crash-chart-of-the-day
##chart-of-the-day
##macro-outlook
#earnings-sentiment-or-this-is-nuts-whens-the-crash
#financial-euphoria
#sell-side-analysts
#market-psychology
#noise-traders

The subhead: When 30 States Raised Minimum Wages, What Happened to Pay and Jobs? The minimum wage job-loss apocalypse that never arrived. At work here we see monopsony, not mayhem. Assessing Card, Krueger, Dube, and the end of the textbook minimum wage story…

Brad here: The “natural experiment” results are in.

Share

The labor market is not a competitive machine that will be broken by a minimum wage floor; it is a distorted landscape of power imbalances waiting to be corrected. We see the same pattern in the “raise states” regardless of whether they are Democratic strongholds or states that have passed these measures via real populist revolt and ballot initiative. The “bite” of the minimum wage—the ratio of the floor to the median wage—matters, but the fundamental mechanism holds.

The economics profession has a habit of being slow to revise its fundamental intuitions. We are very good at updating our regressions, but very bad at updating our paradigms. We saw the evidence of market power, and we chose to believe in inelasticity. We saw the evidence of productivity gains, and we chose to focus on the “disemployment” ghosts of the 1960s.

We should have learned this a long time ago.We should have been asking: What is the cost of the status quo? If we treat the minimum wage as a simple tax on employment, we ignore the deadweight loss of monopsonistic under-employment. If the state can use a wage floor to counteract the market power of labor buyers, then the “Harberger Triangle” of inefficiency isn’t created by the minimum wage; it is created by the absence of it. The “waste” isn’t in the higher wage; the waste is in the unfilled positions and the suppressed earnings of workers who are being paid less than their marginal contribution because the market lacks the competitive pressure to do otherwise.

Give a gift subscription

Arin’s SubstackMy personal SubstackBy Arin DubeCROSSPOST: ARINDRAJIT DUBE: A Minimum Wage Natural Experiment Has Been Running for Over a Decade

<https://arindube.substack.com/p/a-minimum-wage-natural-experiment> <http://arindube/substack.com>

When 30 States Raised Minimum Wages, What Happened to Pay and Jobs?

JUN 10, 2026Arin’s SubstackA Minimum Wage Natural Experiment Has Been Running for Over a Decade The federal minimum wage of $7.25 has not been raised since …. wait for it … 2009. That’s 17 years, if you’re counting. At the same time, after 2013, 30 states have raised their minimum wage above the federal floor. The other 20 have stuck with $7.25 an hour throughout. More than a dozen years later, this divergence has become an unprecedented “natural …Read more15 days ago · 6 likes · 2 comments · Arin Dube

The federal minimum wage of $7.25 has not been raised since …. wait for it … 2009. That’s 17 years, if you’re counting. At the same time, after 2013, 30 states have raised their minimum wage above the federal floor. The other 20 have stuck with $7.25 an hour throughout. More than a dozen years later, this divergence has become an unprecedented “natural experiment” in minimum-wage policy — not just in the United States, but anywhere. Teasing out causality in social sciences is hard. But this large, persistent, and still-widening gap between the two groups gives us an unusually informative “laboratory” for testing predictions about the labor-market effects of minimum-wage policy.

Let’s have a look.

Leave a comment

Two halves of the country, one big difference

The map below shows the 30 raise states in green, shaded darker for a higher 2025 minimum wage, and the 20 federal-floor states (still at $7.25) in gray.

Figure 1. The 30 raise states (green, shaded by 2025 minimum wage) vs. the 20 federal-floor states (gray)

Notes: Raise states raised their minimum wage above the federal floor at some point since 2013; federal-floor states stayed at $7.25 throughout 2013–2025. Source: state minimum-wage statutes.

Fig. 2 shows that the gap between the two groups has widened steadily since 2013. In the 30 “raise states,” the average floor climbed from $7.60 in 2010 to $14.44 by 2025. The light green band shows the spread between the lowest and the highest of these floors by year, ranging from $8.75 in West Virginia to $16.66 in Washington in 2025. For the 20 “federal-floor” states, in contrast, the nominal minimum wage remained stagnant at $7.25 throughout the whole period.

Figure 2. Minimum-wage levels, 2010–2025: raise states vs. the federal floor

Notes: Green line: population-weighted average minimum wage across the 30 raise states; band: full range, from the lowest to the highest state floor. Gray: $7.25 federal floor. Nominal dollars. Source: state minimum-wage statutes.

Did paychecks go up? Did jobs disappear?

A natural place to start is the industry most affected by the minimum wage: restaurants and other eating places (NAICS 7225). This low-wage industry is the most heavily studied by economists. The BLS Quarterly Census of Employment and Wages (QCEW) covers nearly every private-sector employer, so we can track the two groups of states year by year. For each year, the QCEW data reports the average weekly earnings and the average headcount employment over a pay period by industry and state (or county).

Figure 3 plots the year-by-year gap in restaurant pay and jobs between the raise states and the federal-floor states. The figure indexes everything to 2013, the year before floors started rising: this is the base period. Then it shows the gap in the percentage change in the outcomes (i.e., average restaurant earnings in red, per-capita restaurant employment in blue) from 2013 to another year, like 2019, or 2025. The dots and lines show the actual estimates. The shaded area shows the margin of error (95% confidence interval).

What do the data show? First, average restaurant pay diverges steadily from 2014 onward: averaging over 2023-2025, the average wage grew by around 7.7% ( ±3.3%) more in the raise states since 2013. Restaurant employment, by contrast, tracks close to the 2013 baseline throughout, changing by 0.3% ( ±3.3%) over the same period. Dividing the employment change by the wage change gives an “own-wage elasticity” (OWE) near zero, of +0.03. (A quick refresher: an OWE of something like -1 would be a “big negative,” with job losses erasing all wage gains; while something like -0.1 suggests very limited job losses. For more on the OWE see my post here.)

Figure 3. Year-by-year restaurant pay and jobs gap: raise states minus federal-floor states

Notes: Each year’s point is the percent gap in restaurant (NAICS 7225) average weekly wage or jobs per resident between the raise states and the federal-floor states, relative to 2013. State and year fixed effects; each state weighted by 2013 population; shaded bands are 95% confidence intervals clustered by state. COVID years 2020–21 are set aside (dashed bridge). Source: BLS QCEW.

Checked three different ways — and they all agree

The comparison above was the simplest one to make, and also the most transparent. We just plotted the gap in the average per-capita restaurant employment and earnings growth between the two groups of states starting from 2013. But we can also do fancier analyses that have been used by economists studying minimum wages (or similar policies).

2010 paperrecent updateFirst, we can compare adjacent counties on the opposite sides of a state border where one side is a “raise” state while the other side is a “federal-floor” state. This type of analysis holds constant a lot of confounders like regional shocks, giving a more apples-to-apples comparison (for more along this line of analysis, see my 2010 paper and a more recent update). Finally, a third approach compares the 30 raise states to a “twin” (or synthetic control) constructed from reweighting the 20 federal-floor states to match the trajectories in outcomes before minimum wage hikes started in 2014. (See here, and here, for more on this synthetic difference-in-difference approach.)

All three approaches show clear gains in average restaurant pay: as already discussed, comparing the 2023-2025 average to 2013, the baseline comparison shows +7.7% change in pay, coupled with a +0.3% change in jobs. Comparing neighboring counties across state lines, we find a pay increase of +4.2% ( ±1.7), and a change in jobs of −0.5% ( ±3.2). Finally, the synthetic “twin” states suggest a pay hike of +7.2%, coupled with a −0.6% change in jobs. Each of these research designs shows a clear, statistically significant pay bump, and a small change in employment indistinguishable from zero.

Figure 4. Restaurant pay and jobs gap under three research designs (pooled 2023–2025 vs. 2013)

Notes: Each pair of bars pools 2023–2025 relative to 2013 under a different design. Whiskers are 95% confidence intervals clustered by state. Source: BLS QCEW.

Does the size of the raise matter?

What happens when we look within the raise states and focus on those that raised wages the most? We can consider two different ways of measuring the size of the raise: the first method considers the level of the 2025 minimum wage, while the second method considers the 2025 minimum wage as a share of the state’s typical (median) wage — its “bite.” (When the median wage is lower, it indicates that the same floor might reach further up the local pay ladder.)

As Figure 5 shows, for both measures the pattern is clear: bigger minimum wage boosts meant bigger pay gains for restaurant workers. In neither case does the group with the largest raises show any indication of a more negative employment effect — in fact those groups’ employment estimates are slightly positive, though not statistically distinguishable from zero.

Figure 5. Restaurant pay and jobs gap by size of the raise (pooled 2023–2025 vs. 2013)

Notes: All raise states vs. the third with the highest 2025 minimum-wage level and the third where the floor reaches highest relative to local pay (bite). Each state weighted by 2013 population; 95% CIs clustered by state. State median wages from EPI. Source: BLS QCEW.

Restaurants and beyond: all low-paid industries

We can broaden the lens beyond restaurants. The “all low-paid industries” aggregate adds accommodation and food services, retail trade, other services, arts and entertainment, and administrative and waste services: these comprise the bulk of low-wage employment outside agriculture.

Figure 6 shows that pay rose in both groups. The effect is larger in restaurants, where the floor bites hardest. However, when it comes to jobs, employment change was +0.3% in restaurants and −0.1% in low-paid industries, neither statistically different from zero.

Figure 6. Restaurants vs. all low-paid industries, all raise states (pooled 2023–2025 vs. 2013)

Notes: Restaurants (NAICS 7225) vs. a broader low-wage aggregate (NAICS 72+44–45+81+71+56). Population-weighted; 95% CIs clustered by state. Source: BLS QCEW.

Extending the analysis to all low-paid industries, while considering states with the highest minimum-wage level only, we see an average pay increase of +2.6% ( ±1.7), while jobs change is +2.6% ( ±4.1); again, no sign that the higher floor led to lower jobs.

Figure 7. Restaurants vs. all low-paid industries, highest-minimum-wage states (pooled 2023–2025 vs. 2013)

Notes: Same comparison as Figure 6, restricted to the highest-minimum-wage states. Source: BLS QCEW.

Many red and purple states raised their minimum wages

While the 30 raise states lean Democratic, there are plenty of examples of Republican-leaning red and purple states that raised their floors. 12 states raised the floor and voted for Donald Trump in 2024. This is of interest for at least two reasons. First, the 20 federal-floor states lean right, so these 12 states share more similarities, making perhaps a more apples-to-apples comparison. Second, as voters in some of these 20 states consider raising their minimums—like on June 16 in Oklahoma, where I was earlier this week, talking about The Wage Standard—the impacts in these 12 states may be more relevant when thinking about likely consequences.

Figure 8. The 12 red/purple states that raised their minimum wage

Notes: States that raised the minimum wage and voted Trump in 2024 (green); others gray. The five heading toward ~$15 by ballot: AZ, FL, MO, NE, AK. Source: state statutes; 2024 election returns.

Figure 9 shows the pay and employment path of these 12 red/purple raise states compared to the 20 federal-floor states. Again, there is a clear boost in restaurant pay to the tune of +6.2% by 2025. And again, restaurant jobs gap remains flat (−0.8%, ±2.9). Overall, we see a similar story in these red/purple states as in the broader set of raise states.

Figure 9. Year-by-year restaurant pay and jobs gap: 12 red/purple raise states minus federal-floor states

Notes: Percent gap relative to 2013 across the 12 red/purple raise states. COVID years 2020–21 bridged (dashed). 95% CIs clustered by state. Source: BLS QCEW.

Further considering the 5 red/purple states that all have a path to $15/hour by 2026, we find a little larger pay boost at +7.5% ( ±2.5), while the effect on jobs is +0.1% ( ±3.2). The story remains the same even as we consider these stronger minimum wage hikes in red/purple states that all passed their minimums using ballot initiative (just like voters in Oklahoma).

Figure 10. Pooled restaurant pay and jobs gap: all 12 red/purple raise states and the 5 ballot states

Notes: Pooled 2023–2025 vs. 2013. The 5 ballot states are heading to ~$15. 95% CIs clustered by state. Source: BLS QCEW.

If it’s not jobs, how are minimum wages absorbed? The 3 P’s

Stepping back, the standard model of supply and demand tells us that a government mandated floor will lead to lower employment. Why does that not happen here? As I explain in The Wage Standard, in real life, employers hold wage-setting (monopsony) power; and they may opt to pay a lower wage even though it might lead to higher vacancies and quits because it is more profitable. In such a world, a higher floor pushes up pay and can cut vacancies and quits—not jobs—at least up to a point.

So in a world where jobs don’t fall, how are the added costs absorbed? Here it’s useful to think about the “Three P’s” that summarize the absorption channels:

Productivity: Minimum wages can lead to higher productivity through a variety of channels. This includes reduced turnover, which saves costs of recruitment and retraining. It also can lead to more incentivized workers. And finally, there may be some reallocation towards higher productivity companies in the market.

Profits: Some of the cost is absorbed via lower profits, although here the evidence is somewhat more mixed.

Prices: Finally, some passed through to prices. This is an important mechanism that helps soften the blow for companies’ bottom lines. And it might mean somewhat higher costs for some low-wage-intensive goods and services. But the scale is quite small and limited. For instance, a $5 burger might cost 15¢ more; but overall price levels (as measured by, say, the consumer price index) barely move.

(For more on this, see here.)

Take-away

Since 2013, 30 states have raised their minimum wage substantially — the population-weighted average nearly doubled. Compared to the 20 federal-floor states, restaurant pay rose about 8 percent more, while restaurant employment grew at essentially the same rate. The same story emerges across the broader set of low-paid industries, across three independent research designs, across the biggest raises, and across red and purple states alike. This doesn’t settle all questions about the policy’s impact on employment — for example, on specific groups of workers — and the data also cannot assess effects on hours of work. But within the range of policies actually attempted in the US between 2013 and 2025, wage floors appear to have delivered wage gains at the bottom without a noticeable impact on jobs in the key low-wage sectors the policy targets.

Arin’s SubstackMy personal SubstackBy Arin Dube

Give a gift subscription

Brad back again: An awful lot here. and, as best as I can tell, it is all correct. To first order, the right way to think about minimum wages in the range we see in the United States is as something like optimal natural monopoly regulation: it is regulation on the labor buyer rather than the commodity seller side, but the principle is the same. The state can get more labor hired at a higher wage by setting a wage floor to keep employers from exercising their market power.

that was the lesson we should have learned from Card and Krueger long ago. Instead, the lesson the economics profession convinced itself it had learned was different: It was that the disemployment effects of minimum wages were relatively small: that labor demand was relatively inelastic so that pushing the wage up did not push the quantity of labor demanded down by very much at all. Rather than market power and monopolistic competition meant that the effective labor demand curve sloped the wrong way in the context of a minimum wage.

From 2013 to 2025, the “raise states” have seen restaurant pay grow, a expected. Restaurant employment across the states has remained essentially flat. The call it “own-wage elasticity” is near zero. This is not possible in the competitive model. That tells us we are in a world of pervasive labor-market monopsony—a world where labor-power buyers can set wages below the marginal product of labor and do so. The tragedy of our profession’s response to Card and Krueger was that we turned a discovery about market power into a caution about inelasticity, and then used that to justify a timid policy stance.

We should have been asking: Why is the employment not falling? And the answer, as Dube notes, lies in the “Three P’s”: Productivity, Profits, and Prices. When we raise the wage floor, firms adapt: reduce turnover (productivity), absorb costs (profits), and pass costs on. And consumers swallow the higher burger prices.

Subscribe now

Leave a comment

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…##crosspost-arindrajit-dube-a-minimum-wage-natural-experiment-has-been-running-for-over-a-decade
##crosspost
##inequality-and-domination
#arindrajit-dube-a-minimum-wage-natural-experiment-has-been-running-for-over-a-decade
#arindrajit-dube
#a-minimum-wage-natural-experiment-has-been-running-for-over-a-decade
#minimum-wage
#monopsony-power
#own-wage-elasticity
#restaurant-employment
#low-wage-workers
#card-and-krueger

A labor market in rough balance is a strange place to look for an AI-driven jobs collapse. The full-time typist has disappeared, but very few jobs are like that—fully automatable by “AI”…

From Apollo, we have the monthly vacancies/unemployed ratio for the past 11 and a half years:

Share

Give a gift subscription

And:

Torsten Slok: Where Is the AI Jobs Crisis? <m/wealth/the-daily-spark/where-is-the-ai-jobs-crisis>: ‘If AI were triggering a jobs crisis, we would expect job openings to collapse and unemployment to climb, yet the opposite is happening. The number of job openings per unemployed worker has started to rise again and is now back above 1.0, meaning there are still more jobs available than workers to fill them, see chart below. The May jobs report reinforced this with nonfarm payrolls jumping by 172,000, confirming that there are no signs of workers being replaced by ChatGPT…

Refer a friend

The labor market remains in the rough balance it has been in since mid-2024.

Back when my father was twenty some years younger than I am now and was working for the Ford Foundation, he would dictate and somebody—female—a typist—would type up his memos. The typist. At the keyboard eight hours a day with input coming through the earphones via dictaphone tapes or through the eyeballs via marked-up previous drafts or (shudder) handwritten scribbles. Those jobs are gone. But very few jobs are like the full-time typist’s job. Most jobs remain, albeit with a different configuration of tasks and automation, as waves of technological change wash over them.

So far, so good: We are not horse, and we have not peaked.

Subscribe now

Leave a comment

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…##the-ai-jobs-apocalypse-is-not-in-the-data-chart-of-the-day
##chart-of-the-day
##macro-outlook
#the-ai-jobs-apocalypse-is-not-in-the-data
#torsten-slok
#vacancies-unemployed-ratio
#labor-market
#automation
#technological-change
#future-of-work

Read more

!Spencer Pratt and the 2026 Fraud Script: Majority illusion, mail ballots, and the coming election-night panic. How a viral mayoral longshot, a predictable vote‑count pattern, and a right-wing grifter ecosystem prefigure 2026. “Dumped votes” and prediction‑market screenshots are tells, not evidence, in the next round of election denial. And it is Spencer Pratt’s LA flame‑out that teacheth the lesson about social media, mail‑in ballots, and weaponized delusion…

Online hype made Spencer Pratt’s LA mayoral bid look dominant, but normal vote-counting erased the illusion. Unfortunately, the count pattern did ot fuel but was used by right-wing grifters baseless fraud narratives.

This is important! Read it! Remember it!:

• How social media’s “majority illusion” made Pratt’s online support look like real-world majority backing to naïve right-wingers, even though local polling never showed him leading.

• There is a “red mirage, blue shift” dynamic in mail-in voting, where early in‑person votes skew Republican and later mail ballots swing Democratic, predictably eroding GOP election‑night leads.

• Right-wing election-deniers and grifter-influencers weaponize this predictable counting pattern.

• They add irrational prediction‑market moves into their argument to insinuate fraud whenever their side loses ground.

• Last, the fraud claims are a form of “cope” for supporters who mistake their curated feeds for the electorate.

• Understanding these dynamics is essential for understanding what is likely to go down on election night 2026.

  Share

Agents of InfluenceA field guide to the people and systems turning online belief into real-world power.By Renee DiRestaCROSSPOST: RENÉE DiRESTA: Red Mirage, Blue Shift, Online Cope

Jun 07, 2026

<https://agentsofinfluence.substack.com/p/red-mirage-blue-shift-online-cope> <http://agentsofinfluence.substack.com>

Agents of InfluenceRed Mirage, Blue Shift, Online CopeSpencer Pratt is a man who understands spectacle. Even if you’ve never watched an episode of The Hills, you probably know the contours of his reality-TV villain arc — it fed the early-2000s gossip-blog machine. He has name recognition. And in 2026, in the age of politics as pseudoevents and kayfabe, he became the main character of the LA mayoral primary…Read more15 days ago · 67 likes · 7 comments · Renee DiResta

Spencer Pratt is a man who understands spectacle. Even if you’ve never watched an episode of The Hills, you probably know the contours of his reality-TV villain arc — it fed the early-2000s gossip-blog machine. He has name recognition. And in 2026, in the age of politics as pseudoevents and kayfabe, he became the main character of the LA mayoral primary.

His candidate plotline was made for the attention economy: nontraditional longshot breaks out by saying what They Don’t Want You To Hear. Compelling personal story: he tragically lost his house in the LA fires. Strong trolling game. Online clipper accounts went crazy for him – not entirely organically, but hey — it’s called the attention economy for a reason. The algorithm ate Pratt up.

However, the season finale is upon us, and it’s taking the form of a proliferation of predictable conspiracy theories promoted by people who should know better.

The LA mayoral primary’s devolution into conspiracy tropes, Polymarket screenshots, and Grok fact-checks is useful because it shows a script we are going to see again in 2026: viral candidate overperforms in the feed, ordinary ballot-counting reveals a different reality, and the people who conflated engagement with votes decide that there’s something suspicious about the count.

The majority illusion

Network scientists at USC explained it years ago: your social media feed is dominated by the loudest, best-connected nodes. Their opinions can feel like the majority consensus even when it’s not. You might be thinking you’re seeing the population, when you’re really seeing the loudest people. The term they gave for misjudging where prevailing opinion is based on what you see around you is called the majority illusion.

Pratt, the former reality-TV villain, ran a campaign tailor-made for clipping and virality. The personal-loss story. The visible dysfunction LA continues to struggle with. The “incompetent incumbent” frame he was able to wield against Bass. The algorithm loves righteous indignation; people like an underdog; lots of voters are already inclined to believe existing government officials are useless; and Pratt is unusually well-suited to the age of vertical video. His opponents were politicians. He was a content creator.

The attention economy dynamics were compounded by the fact that California is a national content object. California politics is insane! It attracts a massive amount of attention from people who are not Californians. When I lived in SF for a decade, I was very involved in local politics and constantly bemused by the national interest in our algebra wars, school board recall, DA recall, and House, Senate, and governor’s races. California looms large in ‘Dems In Disarray’ storylines, and Fox News is always on it. There are also enough California-based influencers and poasters that its political discourse seeps onto everyone else’s feeds, and then they have feelings.

And that’s where things can get weird. Ten years ago, I was strongly advocating for a bill to improve vaccination rates in California public schools and constantly getting yelled at by nutjobs online. The bill was polling around 70% approval in surveys by the Sacramento Bee and other local media. Online, the conversation was almost entirely negative. But the online conversation was shaped by lots of people (Texas antivaxxers, hello!) who didn’t live anywhere close to California. There were also bots.

The Pratt feed really was inescapable. In certain parts of X and TikTok, he was once again as much a fixture as he was on the gossip blogs during his reality-TV heyday. He has more than a million followers on X. A fan-made AI video casting Pratt as Batman and Karen Bass as the Joker pulled more than 5 million views; another, featuring pilates moms confessing they were voting for Pratt, did nearly two million. The campaign barely bothered with traditional paid media, leaning on clips and boosting fan-made content while his rivals spent on TV. The Extremely Online right-wing influencer machine boosted him (for better or worse; Steve Bannon presciently remarked that he would endorse him but didn’t want to hurt him in LA).

When your feed is wall-to-wall Pratt — clips, endorsements, memes, weird AI propaganda — it is easy to slide from “Pratt is popular online” to “Pratt is popular.” But a clip with three million views just tells you the clip did numbers online, and presumably a lot of people who liked it are nowhere near Los Angeles. The engagement number alone does not tell you much about how a city of millions will fill out their ballots.

I pulled up polling data to see where the numbers were as Pratt’s campaign went viral. Polling experts, please correct me in the comments if I’m wrong, but as far as I can tell, Pratt advanced from about 14% in the March UC Berkeley IGS/LA Times poll to 22% in their final survey. That final poll had Bass at 26% and Nithya Raman at 25%. The only survey that ever showed him ahead was a 400-person McLaughlin & Associates poll — a Republican pollster — that contradicted every other survey of the race but Breitbart touted it. As the clips spread online, Pratt’s unfavorability climbed from 28% in March, when most likely Angeleno voters had no opinion of him at all, to 57% by late May. (Raman’s recognition increased, and her favorability climbed.) I went back and looked at Zohran Mamdani, who also dominated social media on his way to actually winning New York; his online momentum did show up as positive trends in the polls.

But the extremely online MAGAsphere on X seem to have equated high online engagement with high odds that Spencer Pratt was going to win, or make it to a runoff. The former was always a serious long shot. This is Los Angeles! Pratt is a Republican candidate, with President Trump’s public blessing, running in a city where GOP registration is under 15%. The last Republican mayor of Los Angeles was elected in 1993, and left office in 2001. Most people in LA hate Trump. They may be very frustrated with Bass and very, very reluctantly voting for her again, but it’s unrealistic to believe they would elect Pratt in a landslide. A runoff seemed possible, but the third candidate had to be considered.

Red mirage, blue shift

So did they take the loss like adults? Well, consider this post by Florida Governor Ron DeSantis. Quote-tweeting a post about Pratt’s declining odds on Polymarket, he wrote that California “keeps dumping votes,” that the ballot drops “always seem to go one way,” and asked whether officials were counting “until you get the result you want.”

“Dumping votes”? Scheduled release of counted legal ballots, or things falling off the back of a truck? DeSantis, who is a governor, gets to gesture at fraud without having to produce any evidence of it. Why? Because most of the ballots are likely to go for Bass or Raman…which any competent prediction market bettor would ironically be aware of.

Here’s why the ballot drops “go one way”: Republican voters are more likely to vote in person. Democratic voters are more likely to vote by mail. California has lots of people who use vote-by-mail, which means it takes days to process ballots, verify signatures, cure problems, and update totals. As those later ballots are added, a Republican lead can shrink. Sometimes the order flips. This is one of the most predictable features of vote-counting in states with heavy mail voting. This phenomenon is commonly called the “red mirage, blue shift.”*

Every American should understand what is happening right now with the narratives surrounding the mail-in ballot count, because election deniers have learned to use ballot-counting timelines, and this very predictable shift, as a fraud-accusation template whenever their preferred candidate’s early lead shrinks.

On election night, Pratt was around 30% of the vote, roughly ten points and 40,000 votes ahead of Raman for the second runoff slot. By Friday evening, with about 71% of the expected vote counted, it was Bass 35%, Pratt 28.2%, Raman 24.9% — a lead of 20,672 votes, shrinking with every batch the Registrar released. The trend is appearing in the governor’s race on the same ballot: Steve Hilton, a Republican running for Governor, led on election night and watched Democratic challenger Xavier Becerra eat into his margin as mail ballots were processed. Becerra has now passed him.

In November 2022, in the LA mayoral race, Republican Rick Caruso led Karen Bass in the post-election count by about 12,000 votes — 51.25% to 48.75% — with many ballots still left to count. Bass overtook him days later as mail ballots were processed, and ultimately won by about nine points. Same city, same office, one cycle ago.

Election deniers have recently tried to reframe the fact that this pattern is predictable as some kind of plot in and of itself. Explaining its predictability is an effort to normalize election fraud, you see, because that is prebunking it, and prebunking is brainwashing, and we can’t prove that the ballots aren’t being manufactured or bussed in from China or whatever this time. Yes: as you read this newsletter explaining the basic predictable mechanics of online narratives and electoral systems, I am actually brainwashing you into accepting fraud as normal.

X has done quite a bit to feed the election rumor frenzy: there was a spate of misleading claims that a batch of 24,000 votes (the Notorious Searchable Statistic of this race) had been released, of which Pratt earned zero.

This turned out to be false – some screenshot had captured a moment before Pratt’s tally was updated. (AP told the LA Times that the two updates together actually included 21,870 votes for Pratt, 12,850 for Bass, and 9,521 for Raman.) Governor Gavin Newsom and an Assistant US attorney from the DOJ came out to debunk the claim. Grok had been confused in the early moments when the claim first began to circulate, but subsequently began to factcheck it correctly. Yet for some reason, no Community Notes showed up on tweets making the false 24,000 Pratt-Excluding Ballots claims.

Elon Musk, who used to vote by mail in California, obviously weighed in.

There is, to be clear, no evidence of fraud anywhere. Pratt continues to gain votes. He is just not gaining them at the same rate as his Democratic opponents.

Prediction markets figure in a fair number of bizarre claims this time around as well; they have updated with each release count, which is the point of having prediction markets. Somehow, some of the same people screaming fraud have decided that the market moving against Pratt is evidence that traders “know” he will lose because the race is rigged. This is a strange theory of prediction markets. The whole premise is that the market updates as new information comes in. Prices should cut through the cope. Red mirage/blue shift likelihood affecting the outcome should be priced into the contract, since it is how election behavior works and people are committing real money.

The cope

Pratt tapped real anger. LA is expensive, and a lot of angry, exhausted people are living with the consequences of fires and compounding institutional failure. A candidate who points a camera at visible dysfunction and says, “it shouldn’t be like this” will attract attention from people who appreciate that he is acknowledging what they are feeling.

But clips and memes aren’t ballots. Going viral doesn’t translate to getting elected. A fraud narrative can be a comforting bridge between those two things: it lets supporters continue to trust their perception. The feed wasn’t distorted, the count was. If everyone you saw online wanted Pratt to win and Pratt didn’t win, someone must have interfered.

As of Saturday evening, Pratt still led Raman by 7,494 votes, with more ballots left to count. Pratt himself has not claimed fraud. If Raman ultimately takes the second runoff spot, we will see what he does. California’s count really is slow, and wanting it faster is a legitimate complaint. But “California keeps dumping votes” is not the language preceding a process-improvement proposal.The blue shift in mail-in ballots isn’t an anomaly, it’s a highly predictable event – which means that at this point, grievance-grifters on X have pre-positioned themselves to cry fraud every time a completely normal, predictable thing happens. They will never be lacking for “evidence.” The broader script for the midterms is already clear: early returns will be treated as the “real” election; everything counted afterward will be treated as contamination.

The Hills ended with the camera pulling back to reveal the “street” was a backlot set all along. The sooner Americans understand that political influencers screaming “fraud” are performers too, the better.

*Election law scholar Edward Foley named the “blue shift“; the data firm Hawkfish named the “red mirage“ half in 2020, warning — correctly — that Trump’s election-night lead would be an artifact of counting order.

Agents of InfluenceA field guide to the people and systems turning online belief into real-world power.By Renee DiResta

Brad here: What do I hae to add? Little. She has said it most, and said it well. To recapitulate: there are three moving parts:

1. The network‑science “majority illusion” that fools people into overestimating their side.

2. The red‑mirage/blue‑shift mechanics of modern vote‑counting.

3. The incentive structure facing influencers and politicians who profit from permanent grievance.

The upshot is grim but clarifying: early returns will again be framed as the “real” election, and everything after as contamination by design. Democracy in the U.S. may not survive this election cycle.

Give a gift subscription

Subscribe now

Leave a comment

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…##crosspost-renée-diresta-red-mirage-blue-shift-online-cope
##crosspost
##neofascism
##public-reason
#renée-diresta-red-mirage-blue-shift-online-cope
#renée-diresta
#red-mirage-blue-shift-online-cope
#spencer-pratt
#la-mayoral-race
#majority-illusion
#mail-in-ballots
#election-denial
#disinformation

Monetary policy risk returns as new Fed Chair Kevin Warsh begins cosplaying Arthur Burns in the 1970s. Inflation is back. The oil shocks are too. The lying never left. The price level jumps, new Fed Chair Kevin Warsh wobbles, Trump and the Trumpists cry “Cut interest rates!”, and Kevin Warsh has promised them that he would…

Yes: he is here: at the table in the center of the room where the FOMC meets: Lord Banquo, Thane of Lochaber:

2022’s inflation is coming back to us again:

The BLS reports:

BLS: <https://www.bls.gov/>: ‘CPI for all items rises 0.5% in May; gasoline and shelter up: In May, the Consumer Price Index for All Urban Consumers rose 0.5 percent, seasonally adjusted, and rose 4.2 percent over the last 12 months, not seasonally adjusted. The index for all items less food and energy increased 0.2 percent in May (SA); up 2.9 percent over the year (NSA)…

Scott Bessent falsely claims that this return of inflation will be followed by deflation:

Scott Bessent: <https://www.ft.com/content/b8d9c334-f7c5-4806-b2db-75dda9ae6613?syn-25a6b1a6=1>: ‘Except for inflation, which is, I believe, going to be a short-term blip, the economic data is very strong…. I think we have all the makings for a very strong economy. I think that we have temporary elevated prices that will come back down…

I could say that he has clearly been drinking too much from the Straits of Vermouth, but actually he is just a bald-faced liar: he does not think “elevated prices… will come back down”. The price level is undergoing another oil shock-driven jump-up, and the question is whether medium-term inflation expectations remain anchored or not. And the answer to the question is something we do not know.

Meanwhile, in the center ring of the circus, Kevin Warsh told a lot of lies to somebody in the process of his getting to be Fed Chair. But we do not know who he was lying to. Was he lying to Trump when he told Trump he would not raise rates and would cut rates whenever he could get the committee to vote for it? Or was he lying to all those he told sotto voce that he would be a “normal” Fed Chair rather than a corrupt Trumpist hack? We may learn something about this at the FOMC meeting next week.

Recall that Trump has demanded things of Warsh:

Myles McCormick: Donald Trump piles pressure on Kevin Warsh with call for rate cut <https://www.ft.com/content/056789c0-d316-4b02-be24-2f75208790dd?syn-25a6b1a6=1>: ‘Donald Trump has piled pressure on Kevin Warsh ahead of his first meeting as chair of the Federal Reserve by demanding lower interest rates…. “There’s no reason to raise interest rates,” Trump told NBC’s Meet the Press. “We built the country by doing great and having rates low. What they do is when they raise interest rates, they try and kill success. I don’t want to kill success. We should actually lower interest rates.”… Trump has called for the Fed’s benchmark rate — currently in a range of 3.5 to 3.75 per cent — to be slashed to 1 per cent or lower…. “We’re doing great, and it’s unfair that whenever you do great, they want to raise interest rates,” said Trump. “It should be the opposite way.”…

“Kevin is fantastic, and I want him to do whatever he wants. I don’t want to have a big influence on him,” Trump told NBC. “But my feeling is that when a country is doing well, they shouldn’t be penalised by immediately raising interest rates. They should actually be incentivised”…

For all of us born in 1960 or before, the memory of Arthur Burns’ tenure as chair of the Federal Reserve—and then G. William Miller’s hapless attempts to thread the needle before he was moved over to the Treasury and Volcker brought down the hammer that produced the depression of 1982—is Banquo’s Ghost at the feast. If history regularly repeated itself, we would be living through another supply shock, followed by a period of accommodative policy, which led to unanchored expectations, which eventually necessitated a brutal, Paul Volcker-style crushing of the economy to restore credibility. But history is not a prefixed script that is merely waiting for its actors to arrive on stage. We make choices. We make our own history.

The deanchoring of inflation expectations in the 1970s and then the Volcker Moment of 1982 was a response to a crisis of legitimacy. The Federal Reserve of the 1970s was struggling to find its footing in a new, post-Bretton Woods world. Today, the central banks of the advanced economies possess a level of institutional maturity, autonomy, and toolkit sophistication that would have seemed like science fiction to their 1970s predecessors. They have learned the hard lessons of the past. They are not just watching inflation; they are actively managing the expectations that drive it—or were, until Kevin Warsh became Fed Chair.

And when he takes his seat in the Chair’s chair next week, Banquo’s Ghost will be looking across the table at him.

Subscribe now

Leave a comment

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…##so-far-inflations-return-limited-to-energy-so-far-chart-of-the-day
##chart-of-the-day
##macro-outlook
##neofascism
##federal-reserve
#so-far-inflations-return-limited-to-energy-so-far
#kevin-warsh
#but-for-wales
#arthur-burns
#trumpists
#monetary-policy
#banquo-thane-of-lochaber
#a-man-for-all-seasons

The Age of Harness Engineering: taming the stochastic parrots before they eat your seed-corn codebase. We’ve scaled the models and blown out the context windows—and discovered that reliability lives somewhere else entirely. Context and prompt tricks got us very impressive demos, but were clearly not enough; now we say “this time! for sure!” and bet on the new frontier of “harness engineering” to turn a stochastic parrot into a halfway dependable co‑worker…

As I see it, hype vendors promised million‑token context windows and “autonomous agents”; what we got was context rot, soft failures, and ten‑step workflows that succeed maybe a third of the time. Thus the shift from context and prompt cleverness to the unglamorous but essential work of building tests, tools, and symbolic scaffolding around LLMs so they can actually do work. Sometimes. In particular domains. For today’s LLMs talk like geniuses and reason like a distracted intern with no memory of yesterday’s mistakes. They are best treated not as budding minds but as components inside a larger harness that enforces truth conditions, cleans up messes, and decides what they’re even allowed to see.

Share

We have, over in comments on <https://braddelong.substack.com/p/ai-is-eating-platform-monopolist/comment/272975403>

. Teaching economy & history. Focusing on growth, distribution, money, & finance. Bringing numbers, facts, & blue-hued optimism of the intellect to understanding...","photo_url":"https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fea5ae644-9822-4ca5-ac6b-e18c017d8fbc_1189x1208.png","is_guest":false,"bestseller_tier":1000}],"post_date":"2026-06-08T19:45:55.707Z","cover_image":"https://substackcdn.com/image/fetch/$s_!xTnT!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa12e1382-59a4-490a-9dfd-cbc64ca55806_576x680.jpeg","cover_image_alt":null,"canonical_url":"https://braddelong.substack.com/p/ai-is-eating-platform-monopolist","section_name":"SubTuringBradBot","video_upload_id":null,"id":201192199,"type":"newsletter","reaction_count":23,"comment_count":10,"publication_id":47874,"publication_name":"DeLong's Grasping Reality: Economy in the 2000s & Before","publication_logo_url":"https://substackcdn.com/image/fetch/$s_!PgPl!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffde2453e-9c18-4560-82ca-8b77ae62ef5b_1280x1280.png","belowTheFold":false,"youtube_url":null,"show_links":null,"feed_url":null}">

Give a gift subscription

, the following:

David Thomson: ‘On getting your clever hans to do useful things with code, not just tokenmax junk I’ve come to the conclusion that context management is everything and context rot is the core hard constraint you have to deal with in an llm system. No matter the models advertised context window my new rule of thumb is about 50% or 100k tokens then they get dumb. Even the 1m token models. 100k is probably a little conservative but context rot is usually a cliff and I can’t be bothered testing. The quality of manipulating and recalling the information collapses rapidly and if you need precision (unlike general chat) don’t compact, keep context as clean as you can in each conversation with as minimal info as you can possibly give it. Each time it processes the conversation, errors in the conversation crowds out the signal from the core model so it can’t manipulate and reconstitute info as accurately. Hence you get more confabulation and errors. The Mutica kanban solution is a big change in how effective I’ve found the things at coding. I suspect this fail mode is true in other domains too - it’s just more hidden with soft fails. So for example it’s pretty hopeless very quickly at looking at long amounts of prose like a novel or screenplay…

Leave a comment

I think this is broadly right. Let me lay out things as I see them right now, starting with the unexpected and semi-catastrophic success of the ChatGPT3 technology demonstration:

Get 75% off a group subscription

Context Engineering

In the beginning was context engineering.

Not that we called it that. We called it “stuffing as much as we could into 4096 tokens and praying”, for 4096 tokens was the limit of what the LLM could eat. That constraint was so tight that the only way to get more than bare remarkable sentence-by-sentence verbal fluency—the only way to get anything even semi-convincing as a simulacrum of a real sustained interaction with a Turing‑level entity—was to husband that context window as if it were gold. Every system prompt, every example, every prior turn was a line item in a very small budget. You had one page of paper on which to write down everything the machine needed to “remember” before you asked your next question.

In that world, the art was almost entirely what do I dare put in, and what must I leave out? You learned very quickly that a few extra paragraphs of irrelevant chatter would crowd out the one example that actually anchored the model’s behavior. You discovered, painfully, that if you tried to have the thing read a long contract or a long code file, its “understanding” degraded into word salad. What we now politely call “context overflow” was then just “it went dumb.”

So: primordial context engineering—before the name—was ruthless curation under a hard 4K token ceiling. You were playing Tetris with instructions, examples, and snippets of prior dialogue. The better you played, the more the model looked like an entity that could track a conversation, reason about what had gone before, and give you something that felt, at least from arm’s length, like coherence. You found a trusted database, searched it, and then gave the LLM three paragraphs that your RAG—Retrieval-Augmented Generation—system thought was most relevant. You asked it to remix those. and you prayed.

And, people thought, it is amazing what it can do; as the frontier labs scale it, it will get better and make its way across the finish line.

Indeed: in some ways this is state-of-the-art. This is what my SubTuringBradBot Telegram ‘bot <https://web.telegram.org/k/#@SubTuringBradBot> uses. And I regard it as successful. But that is because it is a RAG-engine limited to remixing catechism answers from the scrubbed-and-trusted delong_qa.db SQLite file. I am now devoting energy to hand-creating and trying to automate the creation of a /dbs/corrections/ datastore to improve that by doing a second pass. I am not looking for a smarter LLM than ollama/gemma4:26b, and would not want one. That would start hallucinating on me. AGAIN.

And so we moved on.

Refer a friend

Prompt Engineering

The wizards relaxed the context-window constraint. 4096 tokens became 8192, then 32768, then 131072, then 1048576. The marketeers told us, repeatedly, that “context is no longer a bottleneck”, in a triumph of wishful hype over software reality. Why a triumph of hype? Because it turns out, as David Thomson notes above, that it appears that above 131072 tokens of context current LLMs at least become really dumb.

There is such a thing as context rot.

It happens at what is a small fraction of the now-advertised context windows. Above that, the model’s ability to manipulate and reconstitute information, to pull out what matters and ignore what does not, first degrades and then collapses. On a local model or a cheap cloud budget this is half-hidden as the thing slows to a crawl. Throwing unlimited NVIDIA GPUs at it it becomes very clear that it has lost the thread: misreads, half‑hallucinated summaries, small slips do not just accumulate but crowd out the genuine signal you had hoped the machine would pick up. It carries more and more cruft forward and the simulacrum of intelligence dies.

And so people pivoted. You followed the advice of OpenAI, Anthropic, Google, and company and gave the machine everything rather than a narrow three paragraphs from RAG you can no longer pretend that “just give it everything” works, you can try instead to control how you ask.

Thus phase two: prompt engineering: the discipline of “what you say.” The model is taken to be a black box with impressive latent capability. You are told that if you only phrase your request correctly—specify the role, break the task into steps, provide the right few‑shot examples, insist on JSON output, and add the obligatory “think step by step”—the network will reveal its inner wisdom.

Share DeLong's Grasping Reality: Economy in the 2000s & Before

The context window is now treated as a generous but finite RAM; your job is to pack it with instructions that corral a fundamentally stochastic text generator into something that behaves like a conscientious intern.

Think of it this way: you are relying, at a fundamental level, on the function map from conversations to next tokens in the training data, as modified by the overlay of RLHF—Reinforcement-Learning with Human Feedback—that makes it closer to a “do what I mean” machine at the price of also turning it into an obsequious toady. You need to somehow figure a way to poke this function so that it lands in a subdomain of conversations in which the human who actually wrote the training data is smart, and is saying things that are reliable. This is a Black Art. This is also a changing art: this changes with every .x update to the LLM.

This is not nothing.

Prompt craft does matter, especially at the margin.

But it is, in a deep sense, trying to optimize the wrong layer. You are still treating each interaction as if it were an isolated call: one big message in, one big message out. The “unit of work,” as the harness‑engineering people now like to say, is just the prompt.

Once you move beyond toy use cases—once you ask the system to read code, write code, run tests, call tools, and then do the whole thing again—you have a different problem. Now the unit of work is not the prompt, but the process. And processes have memory, failure modes, and path dependencies that no amount of clever wording will fix.

The Need for Harness Engineering

The more serious people—I have seen Andrej Karpathy, Tobi Lütke, Simon Willison, and various others do this—therefore pushed a conceptual upgrade. Stop obsessing about the exact string of words you type into the box. Start thinking instead about what the model sees at each step. Architecting the entire information environment of the task. Now the relevant questions become:

• What documents get retrieved, and when?

• How do we summarize or compress prior steps without losing the one crucial detail?

• Which tool outputs do we carry forward, and which do we drop on the floor?

• How do we prevent stale policies, obsolete APIs, or out‑of‑date documentation from being treated as gospel because a vector search thinks they are “similar”?

Karpathy’s favorite analogy here is that the model is the CPU and the context window is the RAM. The CPU can only operate on what is in RAM; it does not matter how powerful the processor is if you have loaded the wrong bytes. Context engineering, in this sense, is the discipline of deciding what gets paged into working memory, when, and in what form.

Notice that this is also where Thomson’s “context rot” bites hardest. Long‑running agents accumulate garbage: partial tool traces, outdated assumptions, badly summarized sub‑tasks. Left unchecked, that garbage crowds out exactly the operational state the system needs now. The temptation—especially with bigger windows—is to keep everything “just in case.” The reality is that this kills you. If the past is full of partial failures and misfires, you emphatically do not want the model “remembering” all of it in detail. You want it curated, abstracted, and in many cases forgotten.

So by the time we hit 2025, the field had, I think, tacitly accepted that context—not just prompt template cleverness—was the real job. And yet, even when you do that well, something still goes wrong.

Harness Engineering

Suppose you have:

• Reasonably good prompts.

• A carefully engineered context pipeline: retrieval, summarization, memory, tool result injection.

You then wire up an “agent”: a loop that lets the model call tools, read their outputs, and decide what to do next. You unleash it on a ten‑step task: say, diagnosing a bug, editing a codebase, running tests, and deploying a fix.

Assume, for a moment, that you get a 90% success rate on each individual step. That sounds quite good. But ten independent steps at 90% success compound down to roughly a one‑third chance the whole chain works. (If you insist on doing the math: 0.9 to the 10th power is about 0.35.) In the real world, where sub‑steps multiply and failure probabilities are not so kind, your end‑to‑end success rates fall through the floor.

If the agent cannot either (a) detect that it has failed, or (b) recover and try again quickly, you are stuck with a very articulate idiot savant: brilliant in flashes, but catastrophically unreliable. This is the lived experience of everyone who has tried to build non‑toy “autonomous” systems with LLMs in 2024–2026.

Hence the next turn of the wheel: harness engineering.

Harness engineering is, as Martin Fowler has put it, everything in “Agent = Model + Harness” that is not the model. It is the operating system wrapped around a stochastic text engine. It includes:

• The tools the system is allowed to call, and with what permissions.

• The tests and verification gates that outputs must pass before being accepted.

• The state management—the progress files, logs, and checkpoints—that let work persist across sessions and across context‑window resets.

• The loop‑breakers that detect when the thing is banging its head against the same problem twelve times in a row and nudge it into a different approach.

• The architectural constraints—the linters, dependency rules, and sandboxing—that prevent the agent from, say, rewriting your entire codebase when you asked it to fix a single bug.

In other words, harness engineering is a shift in the unit of analysis. You are no longer optimizing prompts or contexts. You are designing a closed‑loop machine in which the model is only one component.

Once you see this, you also see why people who built serious harnesses—Anthropic Claude Code above all, but also OpenAI’s Codex team writing a million lines of code with agents, the LangChain folks improving benchmark scores without touching the underlying model, and so forth—get such large performance jumps. They have changed the system, not the engine.

And you see why the harness has to incorporate deliberate error detection and retry. In a multistep pipeline, even a good model will be wrong often. The only route to reliability is to ensure that the system, not the model, shoulders the burden of catching and correcting those errors. A human engineer does not trust their first idea, or their first test run, implicitly. An LLM left to its own devices does. The harness’s job is to break that naïveté.

What the LLM actually brings to the party

At this point it is worth being frank about the underlying technology.

A large language model brings you astonishing fluency in natural language and a high‑dimensional map of token co‑occurrence across the public internet, and whatever proprietary corpuses you have fed it for its pretraining.

What it does not bring you is a cleanly accessible world model, a transparent internal state, or any innate concept of truth, falsity, or causality.

Worse, the public internet on which these models are trained is not a seminar room full of careful scientists. It is, to a first approximation, the output of a planet‑wide army of s***posters—some insightful, some deranged, many engaged in motivated reasoning—and the LLM faithfully absorbs their statistical patterns. On top of that, we then slap an RLHF layer that, as I said, trains the system to be an obsequious toady: apologetic, eager to please, and determined to tell you that whatever you have asked is brilliant and important and achievable in three easy steps.

The result is an entity that can sound like it knows what it is doing, but that has no privileged introspective access to its own operations. It does not know when it is confabulating. It does not know what it has forgotten from ten thousand tokens ago. It does not, absent scaffolding, remember that three tools ago it got an error return and simply decided to ignore it.

This, I think, is where Gary Marcus and the symbolic‑AI people have a point that is not going away. To get beyond parroting and into something that we can reasonably call doing work, you need an additional layer: a symbolic world model, and a body of vetted tool code, that live outside the LLM and constrain it.

The model is your language interface: it can read your spec, browse your docs, and propose code. But the truth conditions live elsewhere: in explicit rules, type systems, tests, databases, simulators. The harness is where those truth conditions are enforced. The symbolic layer says: “No, you may not call this function with those arguments; the types do not match.” The harness says: “No, you may not mark this task done until the tests pass.” The LLM, left alone, will happily declare victory on a broken build.

So when we say that “harness engineering” is the new frontier, we are really conceding that purely statistical text prediction is not enough. You can dress it up with longer context windows and clever prompts, but unless you embed it in a symbolic, tool‑rich environment that catches its mistakes, you get demos, not systems.

The Danger of too Much Memory

Loop back now to Thomson’s observation about context rot and “kanban”‑style management of interactions. If you build a harness naively, you are tempted to save everything: every trace, every thought, every half‑successful attempt. It feels wrong, as a human, to throw away information you might need later.

But for this class of system, that instinct can be fatal.

Because the LLM has no internal sense of epistemic hygiene, every past token you feed it is just another piece of text to pattern‑match against. It does not mark prior mistakes with a red flag. If you carry a failed plan forward verbatim, the model may take it as a positive example rather than a negative one. If your transcript is full of half‑baked summaries and misread logs, those, too, become “evidence.”

Thus: especially when the system’s past history is composed of things that did not quite work, you do not want it “remembering” too much. You want it to remember, in some compact form, that “approach A failed for reason R; do not try exactly that again.” You do not want a hundred kilobytes of flailing.

As I See It

If you zoom out, I see a simple, so far three‑act, story:

1. In the beginning was context engineering. Under four‑thousand‑token constraints, the only way to make anything work at all was to manage the context window obsessively. What text goes in? What do we drop? How do we fake continuity when the machine remembers nothing?

2. Then came prompt engineering. As models improved and windows grew, the obsession shifted to “what you say”: role prompts, chain‑of‑thought, clever examples. Prompt engineering got us from “barely coherent” to “surprisingly capable” for single‑shot and short‑conversation tasks.

3. Now we are in the era of harness engineering. We have discovered that even with good prompts and good context, a stateless text generator chained into a long loop is a failure machine unless we wrap it in tools, tests, constraints, memory, and feedback. Reliability lives in the harness.

Underneath all three lies the unchanging fact that the LLM is, at heart, a beautifully trained stochastic parrot with a s***poster’s prior and a toady’s manners. It will not, on its own, grow a spine, a conscience, or a debugger. If we want systems that can do ten‑step tasks with better than one‑third end‑to‑end success rates, we must give them the ability to notice when they have failed and to try again—fast, repeatedly, under the control of an external structure that is not itself hallucinating.

That is a hard engineering problem. It is also, if one squints, the old human problem of building institutions and procedures that make fallible, boundedly rational, not-too-smart actors safer and more productive than they would be on their own. We are, once again, doing bureaucratic system design—only this time the bureaucracy consists of networks, toolchains, and symbolic controllers wrapped around a talking matrix of floating‑point numbers.

But this is progress!

Do, however, note two things:

1. We are no longer relying on scaling laws to get us anywhere. Scaling the LLM and the bitter lesson simply get us finer and finer approximations to a typical internet shitposter, and we already have a good enough such approximation for all purposes we want. There is no reliable way to get an LLM to improve its own harness.

2. The currently useful systems are by-and-large taking the win. They are settling for what RAG can do, or confining themselves to a restricted set of tool calls and template fill-ins and CleverHansing it: trying a hundred ways of writing this configuration file until they find one that does not barf but runs successfully.

Is there a bitter bitter lesson here? That bigger data, bigger models, and more repetitions of the same task are processes though themselves either already or about to be subject to massively diminishing marginal returns? Perhaps.

Subscribe now

Leave a comment

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…##my-notes-on-the-progression-from-context-to-prompt-to-harness-engineering-in-making-gpt-llms-useful-tuesday-mamlms
##subturingbradbot
##mamlms
##monday-mamlms
#my-notes-on-the-progression-from-context-to-prompt-to-harness-engineering-in-making-gpt-llms-useful
#harness-engineering
#prompt-engineering
#context-engineering
#context-rot
#stochastic-parrots
#symbolic-ai
#ai-bureaucracy
#token-windows

From 2018-05-29: When Andrew Sullivan’s definition of “civility” means “you know and stay in your place”, or good faith for me, bad faith for thee. A fond memory of “generous argument” dissolves when you ask who had to swallow what to stay in the room. And do recall Sullivan’s post‑9/11 denunciations of a domestic “fifth column” of the “decadent left” from its “enclaves along the coasts”, and his cheerful embrace of strategic dishonesty over the Bush tax cuts. The through‑line is simple: the people who demand civility for themselves have rarely felt bound by it. Those who challenged the arithmetic—or the racism—never got reciprocity…

<http://www.bradford-delong.com/2018/05/2018-05-28-filed-sullivan-coates-farrell-make-live-on-2018-11-23.html>

“Civility” looks very different depending on where you stand...

From here <http://crookedtimber.org/2018/05/23/neo-marxism/> I have excerpted three short paragraphs very much worth reading and thinking about:

Andrew Sullivan: ‘This bloggy exchange Ta-Nehisi and I had in 2009, on the very subject of identity politics and its claims.... there was a civility about it, an actual generosity of spirit, that transcended the boundaries of race and background.... The Atlantic was crammed with ideological opposites then, jostling together in the same office, and our engagement with each other and our readerships was a crackling and productive one. There was much more of that back then, before Twitter swallowed blogging, before identity politics became completely nonnegotiable, before we degenerated into these tribal swarms of snark and loathing. I think of it now as a distant island, appearing now and then, as the waves go up and down. The riptide of tribalism can capture us all in the end, until we drown in it...

Share

And:

Ta Nehisi Coates: ‘I got incredibly used to learning from people... quite good at their craft, who I felt, and pardon my language, were fucking racist. And that was just the way the world was. I didn’t really have the luxury of having teachers who I necessarily felt, you know, saw me completely as a human being.... You can go into The Atlantic archives right now, and you can see me arguing with Andrew Sullivan about whether black people are genetically disposed to be dumber than white people. I actually had to take this seriously, you understand? I couldn’t speak in a certain way to Andrew. I couldn’t speak to Andrew on the blog the way I would speak to my wife about what Andrew said on the blog in the morning when it was just us.... I learned how to blog from Andrew. That was who I actually learned from. That was who actually helped me craft my voice. Even recognizing who he was and what he was, you know, I learned from him...

Give a gift subscription

And:

Henry Farrell: ‘In juxtaposition, Sullivan’s and Coates’s pieces provide a miniature history of how a certain variety of self-congratulatory openness to inquiry is in actual fact a barbed thicket of power relations. What Sullivan depicts as a ‘different time’ when ‘neither of us denied each other’s good faith or human worth’, is, in Coates’ understanding, a time where he was required to ‘take seriously’ the argument that ‘black people are genetically disposed to be dumber than white people’ as a price of entry into the rarified heights of conversation at the Atlantic. The ‘civility’ and ‘generosity of spirit’ that supported ‘human to human’ conversation is juxtaposed to Coates’s ‘teachers’ who didn’t see him ‘completely as a human being’. What was open and free spirited debate in Sullivan’s depiction, was to Coates a loaded and poisonous dialogue where he could only participate if he shut up about what he actually believed. Juxtaposing these two gives us a very different understanding of Sullivan’s claim that ‘identity politics [have become] completely nonnegotiable’, and we are all being pulled down by the ‘riptide of tribalism’...

Leave a comment

Finished? Good. Perhaps I should simply say that Henry Farrell has written everything that needs to be written here.

Get 75% off a group subscription

Or perhaps I should note that, back in The Day, it was not us, their adversaries on, say, the 2001 tax cut, who were denying George W. Bush’s and Andrew Sullivan’s good faith in the debate. It was Andrew Sullivan himself glorying that neither George W. Bush nor he was arguing in good faith:

The fact that Bush has to obfuscate his real goals of reducing spending with the smoke screen of ‘compassionate conservatism’ shows how uphill the struggle is.... A certain amount of B.S. is necessary for any vaguely successful retrenchment of government power in an insatiable entitlement state.... I just hope the smoke doesn’t clear before the spenders get their hands on our wallets again...

Share DeLong's Grasping Reality: Economy in the 2000s & Before

Sullivan felt himself under no obligation to be honest or civil back when he thought he was riding high. Rather the reverse. And it was not just Black people—although it was far stronger there. Those of us who thought that, like, arithmetic in the public sphere should be accurate were also not worth engaging in good faith.

Let me venture to generalize: Andrew Sullivan was always an interesting thinker, in that “civility” was something that others needed to extend to him but not that he was under any obligation to extend to anybody else. Most of all, we remember Andrew Sullivan after 911:

The middle part of the country—the great red zone that voted for Bush—is clearly ready for war. The decadent Left in its enclaves on the coasts is not dead—and may well mount what amounts to a fifth column...

Refer a friend

and:

We might as well be aware of the enemy within the West itself-a paralyzing, pseudo-clever, morally nihilist fifth column that will surely ramp up its hatred in the days and months ahead...

It was not just Ta-Nehisi Coates but everybody who had doubts about the rush to conquer Iraq who had good reason to think they were regarded by Sullivan as something less than partners in argument and discernment.

Brad in 2026 here: Hoisting this because I ran across Andrew Sullivan’s intellectual footprints today for the first time in—it feels like years. It has been years! That is quite a blessing!

But it is not an end. I see Helen Pluckrose being easily grifted, and Andrew Sullivan makes her his mark and victim today.

I have a response <https://substack.com/@delongonsubstack/note/c-272926583>:

Dear Ms. Pluckrose—

I remember Andrew Sullivan...

I remember Andrew Sullivan back in 2001...

I remember him writing: “The fact that Bush has to obfuscate his real goals of reducing spending with the smoke screen of ‘compassionate conservatism’ shows how uphill the struggle is.... A certain amount of B.S. is necessary for any vaguely successful retrenchment of government power in an insatiable entitlement state.... I just hope the smoke doesn’t clear before the spenders get their hands on our wallets again...”

Thus when I see you put the words “Andrew Sullivan” and “good faith” together in the same sentence, the words that come to my mind are “easily confused”.

And then I run into your: “‘queer’ refers to a set of political and philosophical beliefs about sex, gender and sexuality which holds that having categories for these is an oppressive social construct which should be disrupted as a form of progressive political activism...”

And I think: some people I run across do use ‘queer’ that way; others do not and use ‘queer’ to simply mean “not 100% heterosexual”; to refer to someone as ‘queer’ is to sometimes attribute a political identity to them, but more often, in my experience at least, is not.

Dollars will get you doughnuts that Mamdani means it in the latter way, that Sullivan knows that Mamdani does so, but thinks that by claiming Mamdani is celebrating those who demonize gay men and lesbians he can put one over on you.

And it worked, didn’t it?

I do know that a willingness to presume good faith is one of the glories of liberalism. And that one should be proud that people can easily grift one in this way—that it is a lot better than the alternative.

But, please, may I ask you not to be so be gullible when Andrew Sullivan or others of his ilk try to pick your pocket again in this way? And may I comment to you a piece by my friend John Holbo, “Vavilovian Philosophical Mimicry” <https://crookedtimber.org/2019/12/03/vavilovian-philosophical-mimicry/>?

I do get a lot from your writings. I hope you and yours are well, or, rather, as well as one can be in a world in which while one may well be, personally, quite comfortable, we live in a world in which no man is an island.

?\n\nI do get a lot from your writings. I hope you and yours are well, or, rather, as well as one can be in a world in which while one may well be, personally, quite comfortable, we live in a world in which no man is an island.\n\nYours, J. Bradford DeLong\n\n* Please subscribe to DeLong's Grasping Reality: \n\n* Please order the worldwide six-figure bestseller: \"Slouching Towards Utopia: An Economic History of the Long 20th Century, 1870-2010\": ","body_json":{"type":"doc","attrs":{"schemaVersion":"v1","title":null},"content":[{"type":"paragraph","content":[{"type":"text","text":"Dear Ms. Pluckrose—"}]},{"type":"paragraph","content":[{"type":"text","text":"I remember Andrew Sullivan..."}]},{"type":"paragraph","content":[{"type":"text","text":"I remember Andrew Sullivan back in 2001..."}]},{"type":"paragraph","content":[{"type":"text","text":"I remember him writing: \"The fact that Bush has to obfuscate his real goals of reducing spending with the smoke screen of 'compassionate conservatism' shows how uphill the struggle is.... A certain amount of B.S. is necessary for any vaguely successful retrenchment of government power in an insatiable entitlement state.... I just hope the smoke doesn't clear before the spenders get their hands on our wallets again...\""}]},{"type":"paragraph","content":[{"type":"text","text":"Thus when I see you put the words \"Andrew Sullivan\" and \"good faith\" together in the same sentence, the words that come to my mind are \"easily confused\"."}]},{"type":"paragraph","content":[{"type":"text","text":"And then I run into your: \"‘queer’ refers to a set of political and philosophical beliefs about sex, gender and sexuality which holds that having categories for these is an oppressive social construct which should be disrupted as a form of progressive political activism...\""}]},{"type":"paragraph","content":[{"type":"text","text":"And I think: some people I run across do use 'queer' that way; others do not and use 'queer' to simply mean \"not 100% heterosexual\"; to refer to someone as 'queer' is to sometimes attribute a political identity to them, but more often, in my experience at least, is not."}]},{"type":"paragraph","content":[{"type":"text","text":"Dollars will get you doughnuts that Mamdani means it in the latter way, that Sullivan knows that Mamdani does so, but thinks that by claiming Mamdani is celebrating those who demonize gay men and lesbians he can put one over on you."}]},{"type":"paragraph","content":[{"type":"text","text":"And it worked, didn't it?"}]},{"type":"paragraph","content":[{"type":"text","text":"I do know that a willingness to presume good faith is one of the glories of liberalism. And that one should be proud that people can easily grift one in this way—that it is a lot better than the alternative."}]},{"type":"paragraph","content":[{"type":"text","text":"But, please, may I ask you not to be so be gullible when Andrew Sullivan or others of his ilk try to pick your pocket again in this way? And may I comment to you a piece by my friend John Holbo, \"Vavilovian Philosophical Mimicry\" <"},{"type":"text","text":"https://crookedtimber.org/2019/12/03/vavilovian-philosophical-mimicry/","marks":[{"type":"link","attrs":{"href":"https://crookedtimber.org/2019/12/03/vavilovian-philosophical-mimicry/"}}]},{"type":"text","text":">?"}]},{"type":"paragraph","content":[{"type":"text","text":"I do get a lot from your writings. I hope you and yours are well, or, rather, as well as one can be in a world in which while one may well be, personally, quite comfortable, we live in a world in which no man is an island."}]},{"type":"paragraph","content":[{"type":"text","text":"Yours, J. Bradford DeLong"}]},{"type":"paragraph","content":[{"type":"text","text":"* Please subscribe to DeLong's Grasping Reality: <"},{"type":"text","text":"http://braddelong.substack.com/subscribe","marks":[{"type":"link","attrs":{"href":"http://braddelong.substack.com/subscribe"}}]},{"type":"text","text":">"}]},{"type":"paragraph","content":[{"type":"text","text":"* Please order the worldwide six-figure bestseller: \"Slouching Towards Utopia: An Economic History of the Long 20th Century, 1870-2010\": <"},{"type":"text","text":"https://www.amazon.com/Economic-History-Twentieth-Century/dp/0465019595/","marks":[{"type":"link","attrs":{"href":"https://www.amazon.com/Economic-History-Twentieth-Century/dp/0465019595/"}}]},{"type":"text","text":">"}]}]},"restacks":0,"reaction_count":1,"children_count":0,"attachments":[{"id":"66c6da34-d857-47e4-a204-857e37b42966","type":"post","publication":{"apple_pay_disabled":false,"apex_domain":"hpluckrose.com","author_id":100128629,"byline_images_enabled":true,"bylines_enabled":true,"chartable_token":null,"community_enabled":true,"copyright":"Helen Pluckrose","cover_photo_url":"https://substack-post-media.s3.amazonaws.com/public/images/613cb4e7-d40f-4c50-8a61-204ef78126dd_1280x640.png","created_at":"2022-07-29T11:28:07.245Z","custom_domain_optional":false,"custom_domain":"www.hpluckrose.com","default_comment_sort":"best_first","default_coupon":"b2d3e68a","default_group_coupon":null,"default_show_guest_bios":true,"email_banner_url":"https://substack-post-media.s3.amazonaws.com/public/images/62c2b0cc-695f-4962-bf71-328f7084a5df_1100x220.png","email_from_name":null,"email_from":null,"embed_tracking_disabled":false,"explicit":false,"expose_paywall_content_to_search_engines":true,"fb_pixel_id":null,"fb_site_verification_token":null,"flagged_as_spam":false,"founding_subscription_benefits":["All the perks of a paid subscription + occasional virtual meetups."],"free_subscription_benefits":["Subscribe to read all my written Overflowings for free and receive new articles direct to your inbox."],"ga_pixel_id":null,"google_site_verification_token":null,"google_tag_manager_token":null,"hero_image":null,"hero_text":"The Overflowings of a Liberal Brain is a community-driven publication that addresses political & cultural issues and offers a liberal perspective on how to understand them and respond to them.","hide_intro_subtitle":null,"hide_intro_title":null,"hide_podcast_feed_link":false,"homepage_type":"magaziney","id":1020561,"image_thumbnails_always_enabled":false,"invite_only":false,"hide_podcast_from_pub_listings":false,"language":"en","logo_url_wide":null,"logo_url":"https://substackcdn.com/image/fetch/$s_!-qGd!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83350fdd-260e-4a76-8bd7-8337a0e41ca1_500x500.png","minimum_group_size":2,"moderation_enabled":true,"name":"The Overflowings of a Liberal Brain","paid_subscription_benefits":["All my written Overflowings.","Voiceovers of all my written Overflowings.","Access to Substack Lives."],"parsely_pixel_id":null,"chartbeat_domain":null,"payments_state":"enabled","paywall_free_trial_enabled":false,"podcast_art_url":"https://bucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com/public/images/f7f8d33a-977b-42b0-8901-7bd9c940691d_3000x3000.png","paid_podcast_episode_art_url":null,"podcast_byline":"With Helen Pluckrose","podcast_description":"Less of a podcast and more of an informal miscellany of rants, readings & general bletherings for my paid subscribers! Here, I intend to read my own essays here for those who find it more convenient to listen as well. I will also read and discuss extracts of some of the books and essays about liberalism I think are most important and have some spontaneous rants!","podcast_enabled":true,"podcast_feed_url":null,"podcast_title":"Pluckrose Pontificates","post_preview_limit":null,"primary_user_id":100128629,"require_clickthrough":false,"show_pub_podcast_tab":true,"show_recs_on_homepage":true,"subdomain":"helenpluckrose","subscriber_invites":0,"support_email":null,"theme_var_background_pop":"#FF81CD","theme_var_color_links":false,"theme_var_cover_bg_color":null,"trial_end_override":null,"twitter_pixel_id":null,"type":"newsletter","post_reaction_faces_enabled":true,"is_personal_mode":false,"plans":[{"id":"yearly70usd","object":"plan","active":true,"aggregate_usage":null,"amount":7000,"amount_decimal":"7000","billing_scheme":"per_unit","created":1666048245,"currency":"usd","interval":"year","interval_count":1,"livemode":true,"metadata":{"substack":"yes"},"meter":null,"nickname":"$70 a year","product":"prod_MdJ4STGy5VPKeV","tiers":null,"tiers_mode":null,"transform_usage":null,"trial_period_days":null,"usage_type":"licensed","currency_options":{"aud":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":10500,"unit_amount_decimal":"10500"},"brl":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":36500,"unit_amount_decimal":"36500"},"cad":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":10000,"unit_amount_decimal":"10000"},"chf":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":6000,"unit_amount_decimal":"6000"},"dkk":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":45500,"unit_amount_decimal":"45500"},"eur":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":6500,"unit_amount_decimal":"6500"},"gbp":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":5500,"unit_amount_decimal":"5500"},"mxn":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":125500,"unit_amount_decimal":"125500"},"nok":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":68000,"unit_amount_decimal":"68000"},"nzd":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":12500,"unit_amount_decimal":"12500"},"pln":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":26000,"unit_amount_decimal":"26000"},"sek":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":66500,"unit_amount_decimal":"66500"},"usd":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":7000,"unit_amount_decimal":"7000"}}},{"id":"monthly7usd","object":"plan","active":true,"aggregate_usage":null,"amount":700,"amount_decimal":"700","billing_scheme":"per_unit","created":1666048244,"currency":"usd","interval":"month","interval_count":1,"livemode":true,"metadata":{"substack":"yes"},"meter":null,"nickname":"$7 a month","product":"prod_MdJ4HdfdBwIkjI","tiers":null,"tiers_mode":null,"transform_usage":null,"trial_period_days":null,"usage_type":"licensed","currency_options":{"aud":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":1100,"unit_amount_decimal":"1100"},"brl":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":3700,"unit_amount_decimal":"3700"},"cad":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":1000,"unit_amount_decimal":"1000"},"chf":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":600,"unit_amount_decimal":"600"},"dkk":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":4600,"unit_amount_decimal":"4600"},"eur":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":700,"unit_amount_decimal":"700"},"gbp":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":600,"unit_amount_decimal":"600"},"mxn":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":13000,"unit_amount_decimal":"13000"},"nok":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":7000,"unit_amount_decimal":"7000"},"nzd":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":1300,"unit_amount_decimal":"1300"},"pln":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":2600,"unit_amount_decimal":"2600"},"sek":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":7000,"unit_amount_decimal":"7000"},"usd":{"custom_unit_amount":null,"tax_behavior":"unspecified","unit_amount":700,"unit_amount_decimal":"700"}}},{"id":"founding21000usd","name":"founding21000usd","nickname":"founding21000usd","active":true,"amount":21000,"currency":"usd","interval":"year","interval_count":1,"metadata":{"substack":"yes","founding":"yes","no_coupons":"yes","short_description":"Founding Member","short_description_english":"Founding Member","minimum":"7000","minimum_local":{"aud":10000,"brl":36500,"cad":10000,"chf":6000,"dkk":45500,"eur":6500,"gbp":5500,"mxn":122500,"nok":66500,"nzd":12500,"pln":26000,"sek":66500,"usd":7000}},"currency_options":{"aud":{"unit_amount":30000,"tax_behavior":"unspecified"},"brl":{"unit_amount":109500,"tax_behavior":"unspecified"},"cad":{"unit_amount":29500,"tax_behavior":"unspecified"},"chf":{"unit_amount":17000,"tax_behavior":"unspecified"},"dkk":{"unit_amount":136500,"tax_behavior":"unspecified"},"eur":{"unit_amount":18500,"tax_behavior":"unspecified"},"gbp":{"unit_amount":16000,"tax_behavior":"unspecified"},"mxn":{"unit_amount":367000,"tax_behavior":"unspecified"},"nok":{"unit_amount":199500,"tax_behavior":"unspecified"},"nzd":{"unit_amount":36500,"tax_behavior":"unspecified"},"pln":{"unit_amount":77500,"tax_behavior":"unspecified"},"sek":{"unit_amount":198500,"tax_behavior":"unspecified"},"usd":{"unit_amount":21000,"tax_behavior":"unspecified"}}}],"stripe_user_id":"acct_1Lu2N9LDQU9gvAYi","stripe_country":"GB","stripe_publishable_key":"pk_live_51Lu2N9LDQU9gvAYiMJ0mDv5b0ReVXZcD08RqCfWm2BAI7SzHiZYnfJqID5azyBH9XYVCWLisfCAEL4r0lRFjlloi00xQkExv51","stripe_platform_account":"US","automatic_tax_enabled":false,"author_name":"Helen Pluckrose","author_handle":"helenpluckrose","author_photo_url":"https://substackcdn.com/image/fetch/$s_!rm5N!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F493c3db7-be04-43d7-a271-764e9d501709_2316x2316.jpeg","author_bio":"Still just wanting you to value evidence- based epistemology and consistently liberal ethics. ","twitter_screen_name":"HPluckrose","has_custom_tos":false,"has_custom_privacy":false,"theme":{"background_pop_color":"#e11d48","web_bg_color":"#ffffff","cover_bg_color":null,"publication_id":1020561,"color_links":null,"font_preset_heading":null,"font_preset_body":null,"font_family_headings":null,"font_family_body":null,"font_family_ui":null,"font_size_body_desktop":null,"print_secondary":null,"custom_css_web":null,"custom_css_email":null,"home_hero":"magaziney","home_posts":"grid","home_show_top_posts":false,"hide_images_from_list":false,"home_hero_alignment":"left","home_hero_show_podcast_links":true,"default_post_header_variant":null,"custom_header":null,"custom_footer":null,"social_media_links":null,"font_options":null,"section_template":null,"custom_subscribe":null,"design_template":null,"design_template_options":null},"threads_v2_settings":{"photo_replies_enabled":true,"first_thread_email_sent_at":null,"create_thread_minimum_role":"contributor","activated_at":"2024-08-12T19:48:27.919+00:00","reader_thread_notifications_enabled":false,"boost_free_subscriber_chat_preview_enabled":true,"push_suppression_enabled":false},"default_group_coupon_percent_off":null,"pause_return_date":null,"has_posts":true,"has_recommendations":true,"first_post_date":"2018-09-20T12:00:00.000Z","has_podcast":true,"has_free_podcast":true,"has_subscriber_only_podcast":true,"has_community_content":true,"rankingDetail":"Hundreds of paid subscribers","rankingDetailFreeIncluded":"Thousands of subscribers","rankingDetailOrderOfMagnitude":100,"rankingDetailFreeIncludedOrderOfMagnitude":1000,"rankingDetailFreeSubscriberCount":"Over 6,000 subscribers","rankingDetailByLanguage":{"ar":{"rankingDetail":"مئات المشتركين المدفوعين"},"ca":{"rankingDetail":"Centenars de subscriptors de pagament"},"da":{"rankingDetail":"Hundredvis af betalte abonnenter"},"de":{"rankingDetail":"Hunderte von Paid-Abonnenten"},"es":{"rankingDetail":"Cientos de suscriptores de pago"},"fr":{"rankingDetail":"Des centaines d'abonnés payants"},"ja":{"rankingDetail":"数百人の有料登録者"},"nb":{"rankingDetail":"Hundrevis av betalende abonnenter"},"nl":{"rankingDetail":"Honderden betalende abonnees"},"pl":{"rankingDetail":"Setki płacących subskrybentów"},"pt":{"rankingDetail":"Centenas de subscritores pagos"},"pt-br":{"rankingDetail":"Centenas de assinantes pagantes"},"en-gb":{"rankingDetail":"Hundreds of paid subscribers"},"it":{"rankingDetail":"Centinaia di abbonati a pagamento"},"tr":{"rankingDetail":"Yüzlerce ücretli abone"},"sv":{"rankingDetail":"Hundratals betalande prenumeranter"},"fi":{"rankingDetail":"Hundreds of paid subscribers"},"is":{"rankingDetail":"Hundreds of paid subscribers"},"en":{"rankingDetail":"Hundreds of paid subscribers"}},"freeSubscriberCount":"6,000","freeSubscriberCountOrderOfMagnitude":"6.5K+","author_bestseller_tier":100,"author_badge":{"type":"bestseller","tier":100},"disable_monthly_subscriptions":false,"disable_annual_subscriptions":false,"hide_post_restacks":false,"notes_feed_enabled":true,"showIntroModule":false,"isPortraitLayout":false,"last_chat_post_at":"2025-12-01T19:38:40.784Z","primary_profile_name":"Helen Pluckrose","primary_profile_photo_url":"https://substackcdn.com/image/fetch/$s_!rm5N!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F493c3db7-be04-43d7-a271-764e9d501709_2316x2316.jpeg","no_follow":false,"sponsorshipCampaigns":{},"paywall_chat":"free","sections":[],"podcastTabInfo":{"hasMultiplePodcasts":false,"firstPodcastSectionId":null},"multipub_migration":null,"navigationBarItems":[],"has_active_perks":false,"contributors":[{"name":"Helen Pluckrose","handle":"helenpluckrose","role":"admin","owner":true,"user_id":100128629,"photo_url":"https://substack-post-media.s3.amazonaws.com/public/images/493c3db7-be04-43d7-a271-764e9d501709_2316x2316.jpeg","bio":"Still just wanting you to value evidence- based epistemology and consistently liberal ethics. ","status":{"bestsellerTier":100,"subscriberTier":10,"leaderboard":null,"vip":false,"badge":{"type":"bestseller","tier":100},"subscriber":null}}],"threads_v2_enabled":true,"viralGiftsConfig":{"id":"66a31f39-bc3b-467c-9d46-16d911da3c22","publication_id":1020561,"enabled":true,"gifts_per_user":5,"gift_length_months":1,"send_extra_gifts":true,"message":"Anything and everything! OK, mostly liberalism & when people are doing it wrong","created_at":"2022-10-17T23:11:57.863818+00:00","updated_at":"2022-10-17T23:11:57.863818+00:00","days_til_invite":14,"send_emails":true,"show_link":null},"tier":2,"no_index":false,"can_set_google_site_verification":true,"can_have_sitemap":true,"iap_advanced_plans":[{"sku":"OWtF6BFfeuXYuX4mpR","publication_id":"1020561","is_active":true,"price_base_units":1000,"currency_alpha3":"usd","period":"month","created_at":"2025-08-18T13:52:06.260Z","updated_at":"2025-08-18T13:52:06.260Z","id":"176960","payout_amount_base_units":70,"alternate_currencies":{"aud":1600,"brl":5500,"cad":1400,"chf":900,"dkk":6500,"eur":900,"gbp":800,"mxn":19000,"nok":10500,"nzd":1700,"pln":3700,"sek":10000},"is_founding":false,"display_name":"The Overflowings of a Liberal Brain (Monthly)","display_price":"$10"},{"sku":"mkEZKaq4Qz6VRE7u2P","publication_id":"1020561","is_active":true,"price_base_units":10000,"currency_alpha3":"usd","period":"year","created_at":"2025-08-18T13:52:06.274Z","updated_at":"2025-08-18T13:52:06.274Z","id":"176961","payout_amount_base_units":700,"alternate_currencies":{"aud":15500,"brl":54500,"cad":14000,"chf":8500,"dkk":64000,"eur":9000,"gbp":7500,"mxn":187500,"nok":102000,"nzd":17000,"pln":36500,"sek":95500},"is_founding":false,"display_name":"The Overflowings of a Liberal Brain (Yearly)","display_price":"$100"}],"founding_plan_name_english":"Founding Member","iap_founding_plan":{"base_plan_id":"founding21000usd","name":"Founding Member","minimum_amount":10000,"suggested_amount":28500,"currency_alpha3":"usd","alternate_currencies":{"aud":{"minimum_amount":14500,"suggested_amount":40500},"brl":{"minimum_amount":52000,"suggested_amount":148500},"cad":{"minimum_amount":14000,"suggested_amount":40000},"chf":{"minimum_amount":8000,"suggested_amount":23000},"dkk":{"minimum_amount":65000,"suggested_amount":185000},"eur":{"minimum_amount":9000,"suggested_amount":25000},"gbp":{"minimum_amount":7500,"suggested_amount":21500},"mxn":{"minimum_amount":175000,"suggested_amount":498000},"nok":{"minimum_amount":95000,"suggested_amount":270500},"nzd":{"minimum_amount":17500,"suggested_amount":49500},"pln":{"minimum_amount":37000,"suggested_amount":105000},"sek":{"minimum_amount":94500,"suggested_amount":269000}}},"bundles":[],"base_url":"https://www.hpluckrose.com","hostname":"www.hpluckrose.com","is_on_substack":false,"spotify_podcast_settings":{"id":19985,"publication_id":1020561,"section_id":null,"spotify_uri":null,"spotify_podcast_title":null,"created_at":"2024-08-30T11:49:39.966Z","updated_at":"2024-08-30T11:49:39.966Z","currently_published_on_spotify":false,"spotify_show_url":""},"unified_podcast_settings":null,"podcastPalette":{"Vibrant":{"rgb":[201,149,122],"population":305},"DarkVibrant":{"rgb":[84,55,33],"population":234},"LightVibrant":{"rgb":[230,158,143],"population":354},"Muted":{"rgb":[164,138,97],"population":279},"DarkMuted":{"rgb":[96,78,58],"population":387},"LightMuted":{"rgb":[193,178,149],"population":313}},"pageThemes":{"podcast":null},"multiple_pins":true,"supports_ip_content_unlock":false,"appTheme":{"colors":{"accent":{"name":"#e11d48","primary":{"r":225,"g":29,"b":72,"a":1},"primary_hover":{"r":203,"g":0,"b":56,"a":1},"primary_elevated":{"r":203,"g":0,"b":56,"a":1},"secondary":{"r":225,"g":29,"b":72,"a":0.2},"contrast":{"r":255,"g":255,"b":255,"a":1},"bg":{"r":225,"g":29,"b":72,"a":0.2},"bg_hover":{"r":225,"g":29,"b":72,"a":0.3},"dark":{"primary":{"r":225,"g":29,"b":72,"a":1},"primary_hover":{"r":247,"g":59,"b":88,"a":1},"primary_elevated":{"r":247,"g":59,"b":88,"a":1},"secondary":{"r":225,"g":29,"b":72,"a":0.2},"contrast":{"r":255,"g":255,"b":255,"a":1},"bg":{"r":225,"g":29,"b":72,"a":0.2},"bg_hover":{"r":225,"g":29,"b":72,"a":0.3}}},"fg":{"primary":{"r":0,"g":0,"b":0,"a":0.8},"secondary":{"r":0,"g":0,"b":0,"a":0.6},"tertiary":{"r":0,"g":0,"b":0,"a":0.4},"accent":{"r":225,"g":29,"b":72,"a":1},"dark":{"primary":{"r":255,"g":255,"b":255,"a":0.9},"secondary":{"r":255,"g":255,"b":255,"a":0.6},"tertiary":{"r":255,"g":255,"b":255,"a":0.4},"accent":{"r":241,"g":51,"b":84,"a":1}}},"bg":{"name":"#ffffff","hue":{"r":255,"g":255,"b":255,"a":0},"tint":{"r":255,"g":255,"b":255,"a":0},"primary":{"r":255,"g":255,"b":255,"a":1},"primary_hover":{"r":250,"g":250,"b":250,"a":1},"primary_elevated":{"r":250,"g":250,"b":250,"a":1},"secondary":{"r":238,"g":238,"b":238,"a":1},"secondary_elevated":{"r":206.90096477355226,"g":206.90096477355175,"b":206.9009647735519,"a":1},"tertiary":{"r":219,"g":219,"b":219,"a":1},"quaternary":{"r":182,"g":182,"b":182,"a":1},"dark":{"primary":{"r":22,"g":23,"b":24,"a":1},"primary_hover":{"r":27,"g":28,"b":29,"a":1},"primary_elevated":{"r":27,"g":28,"b":29,"a":1},"secondary":{"r":35,"g":37,"b":37,"a":1},"secondary_elevated":{"r":41.35899397549579,"g":43.405356429195315,"b":43.40489285041963,"a":1},"tertiary":{"r":54,"g":55,"b":55,"a":1},"quaternary":{"r":90,"g":91,"b":91,"a":1}}}}},"portalAppTheme":{"colors":{"accent":{"name":"#e11d48","primary":{"r":225,"g":29,"b":72,"a":1},"primary_hover":{"r":202,"g":26,"b":65,"a":1},"primary_elevated":{"r":225,"g":29,"b":72,"a":1},"secondary":{"r":225,"g":29,"b":72,"a":1},"contrast":{"r":255,"g":255,"b":255,"a":1},"bg":{"r":255,"g":103,"b":25,"a":0.2},"bg_hover":{"r":255,"g":103,"b":25,"a":0.3},"dark":{"primary":{"r":225,"g":29,"b":72,"a":1},"primary_hover":{"r":247,"g":59,"b":88,"a":1},"primary_elevated":{"r":247,"g":59,"b":88,"a":1},"secondary":{"r":225,"g":29,"b":72,"a":0.2},"contrast":{"r":255,"g":255,"b":255,"a":1},"bg":{"r":225,"g":29,"b":72,"a":0.2},"bg_hover":{"r":225,"g":29,"b":72,"a":0.3}}},"fg":{"primary":{"r":54,"g":55,"b":55,"a":1},"secondary":{"r":134,"g":135,"b":135,"a":1},"tertiary":{"r":146,"g":146,"b":146,"a":1},"accent":{"r":225,"g":29,"b":72,"a":1},"dark":{"primary":{"r":255,"g":255,"b":255,"a":0.9},"secondary":{"r":255,"g":255,"b":255,"a":0.6},"tertiary":{"r":255,"g":255,"b":255,"a":0.4},"accent":{"r":241,"g":51,"b":84,"a":1}}},"bg":{"name":"#ffffff","hue":{"r":255,"g":255,"b":255,"a":1},"tint":{"r":255,"g":255,"b":255,"a":1},"primary":{"r":255,"g":255,"b":255,"a":1},"primary_hover":{"r":240,"g":240,"b":240,"a":1},"primary_elevated":{"r":255,"g":255,"b":255,"a":1},"secondary":{"r":240,"g":240,"b":240,"a":1},"secondary_elevated":{"r":240,"g":240,"b":240,"a":1},"tertiary":{"r":221,"g":221,"b":221,"a":1},"quaternary":{"r":183,"g":183,"b":183,"a":1},"dark":{"primary":{"r":22,"g":23,"b":24,"a":1},"primary_hover":{"r":27,"g":28,"b":29,"a":1},"primary_elevated":{"r":27,"g":28,"b":29,"a":1},"secondary":{"r":35,"g":37,"b":37,"a":1},"secondary_elevated":{"r":41.35899397549579,"g":43.405356429195315,"b":43.40489285041963,"a":1},"tertiary":{"r":54,"g":55,"b":55,"a":1},"quaternary":{"r":90,"g":91,"b":91,"a":1}}},"wordmark_bg":{"r":255,"g":255,"b":255,"a":1}}},"logoPalette":{"Vibrant":{"rgb":[212,82,34],"population":7},"DarkVibrant":{"rgb":[103,33,33],"population":469},"LightVibrant":{"rgb":[243,183,122],"population":91},"Muted":{"rgb":[162,120,86],"population":45},"DarkMuted":{"rgb":[67,35,35],"population":429},"LightMuted":{"rgb":[234,233,228],"population":95}}},"post":{"id":200286894,"publication_id":1020561,"title":"Queer Is a Political Identity, Not a Sexuality","social_title":null,"search_engine_title":null,"search_engine_description":null,"type":"newsletter","slug":"queer-is-a-political-identity-not","post_date":"2026-06-08T08:40:28.273Z","audience":"everyone","podcast_duration":null,"video_upload_id":null,"write_comment_permissions":"everyone","should_send_free_preview":false,"free_unlock_required":false,"default_comment_sort":null,"canonical_url":"https://www.hpluckrose.com/p/queer-is-a-political-identity-not","section_id":null,"podcast_art_url":null,"is_published":true,"live_stream_id":null,"restacks":22,"top_exclusions":[],"pins":[],"is_section_pinned":false,"has_shareable_clips":false,"section_slug":null,"section_name":null,"reactions":{"❤":88},"subtitle":"Expanding on Andrew Sullivan's objection. ","cover_image":"https://substackcdn.com/image/fetch/$s_!GhpD!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F961c2720-fa69-4bb7-b3f5-a59c89da23a2_1535x1024.png","cover_image_is_square":false,"cover_image_is_explicit":false,"podcast_url":null,"videoUpload":null,"podcastFields":{"post_id":200286894,"podcast_episode_number":null,"podcast_season_number":null,"podcast_episode_type":null,"should_syndicate_to_other_feed":null,"syndicate_to_section_id":null,"hide_from_feed":false,"free_podcast_url":null,"free_podcast_duration":null,"preview_contains_ad":false,"was_imported_self_serve_sync":false},"podcast_upload_id":null,"podcast_preview_upload_id":null,"podcastUpload":null,"podcastPreviewUpload":null,"voiceover_upload_id":null,"voiceoverUpload":null,"has_voiceover":false,"description":"Expanding on Andrew Sullivan's objection.","body_json":null,"body_html":null,"truncated_body_text":"(Audio version here)","wordcount":1784,"post_preview_limit":null,"language":"en","postTags":[],"teaser_post_eligible":true,"postCountryBlocks":[],"headlineTest":null,"coverImagePalette":{"Vibrant":{"rgb":[132,148,196],"population":1},"DarkVibrant":{"rgb":[42.98571428571428,54.64285714285713,89.61428571428573],"population":0},"LightVibrant":{"rgb":[212,175,252],"population":3246},"Muted":{"rgb":[132,108,156],"population":5},"DarkMuted":{"rgb":[102,95,123],"population":149},"LightMuted":{"rgb":[196,221,186],"population":16}},"publishedBylines":[{"id":100128629,"name":"Helen Pluckrose","handle":"helenpluckrose","previous_name":null,"photo_url":"https://substack-post-media.s3.amazonaws.com/public/images/493c3db7-be04-43d7-a271-764e9d501709_2316x2316.jpeg","bio":"Still just wanting you to value evidence- based epistemology and consistently liberal ethics. ","profile_set_up_at":"2022-07-29T11:25:27.403Z","reader_installed_at":"2022-10-15T04:37:08.198Z","publicationUsers":[{"id":966646,"user_id":100128629,"publication_id":1020561,"role":"admin","public":true,"is_primary":true,"publication":{"id":1020561,"name":"The Overflowings of a Liberal Brain","subdomain":"helenpluckrose","custom_domain":"www.hpluckrose.com","custom_domain_optional":false,"hero_text":"The Overflowings of a Liberal Brain is a community-driven publication that addresses political & cultural issues and offers a liberal perspective on how to understand them and respond to them.","logo_url":"https://substack-post-media.s3.amazonaws.com/public/images/83350fdd-260e-4a76-8bd7-8337a0e41ca1_500x500.png","author_id":100128629,"primary_user_id":100128629,"theme_var_background_pop":"#FF81CD","created_at":"2022-07-29T11:28:07.245Z","email_from_name":null,"copyright":"Helen Pluckrose","founding_plan_name":"Founding Member","community_enabled":true,"invite_only":false,"payments_state":"enabled","language":null,"explicit":false,"homepage_type":"magaziney","is_personal_mode":false,"logo_url_wide":null}}],"twitter_screen_name":"HPluckrose","is_guest":false,"bestseller_tier":100,"status":{"bestsellerTier":100,"subscriberTier":10,"leaderboard":null,"vip":false,"badge":{"type":"bestseller","tier":100},"subscriber":null},"primary_publication":{"id":1020561,"subdomain":"helenpluckrose","custom_domain":"www.hpluckrose.com","custom_domain_optional":false,"name":"The Overflowings of a Liberal Brain","logo_url":"https://substack-post-media.s3.amazonaws.com/public/images/83350fdd-260e-4a76-8bd7-8337a0e41ca1_500x500.png","author_id":100128629,"user_id":100128629,"handles_enabled":false,"explicit":false,"is_personal_mode":false,"payments_state":"enabled","pledges_enabled":false,"ios_app_payments_enabled":true,"has_reply_rules":true}}],"reaction":false,"reaction_count":88,"comment_count":52,"child_comment_count":20,"audio_items":[{"post_id":200286894,"voice_id":"en-GB-AdaMultilingualNeural","audio_url":"https://substack-video.s3.amazonaws.com/video_upload/post/200286894/tts/b2dc0082-5722-40e5-a04f-d6d4a925a3ee/en-GB-AdaMultilingualNeural.mp3","type":"tts","status":"completed"}],"is_geoblocked":false,"hasCashtag":false,"inboxItem":{"content_key":"post:200286894","updated_at":"2026-06-08T21:51:47.072Z","content_date":"2026-06-08T08:40:28.273Z","inbox_date":"2026-06-08T08:40:28.273Z","seen_at":"2026-06-08T21:51:47.072Z","saved_at":null,"archived_at":null,"skip_inbox":false,"type":"post","post_id":200286894,"extra_views":["media"],"read_progress":0.18209866848666498,"max_read_progress":1,"audio_progress":0,"max_audio_progress":0,"video_progress":0,"max_video_progress":0,"postType":"newsletter","title":"Queer Is a Political Identity, Not a Sexuality","subtitle":"Expanding on Andrew Sullivan's objection.","detail_view_subtitle":"Expanding on Andrew Sullivan's objection. ","cover_photo_url":"https://substackcdn.com/image/fetch/$s_!GhpD!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F961c2720-fa69-4bb7-b3f5-a59c89da23a2_1535x1024.png","audience":"everyone","is_preview":false,"audio_url":"https://substack-video.s3.amazonaws.com/video_upload/post/200286894/tts/b2dc0082-5722-40e5-a04f-d6d4a925a3ee/en-GB-AdaMultilingualNeural.mp3","audio_type":"tts","web_url":"https://www.hpluckrose.com/p/queer-is-a-political-identity-not","duration_metadata":{"word_count":1784},"authors":["Helen Pluckrose"],"published_bylines":[{"id":100128629,"name":"Helen Pluckrose","photo_url":"https://substack-post-media.s3.amazonaws.com/public/images/493c3db7-be04-43d7-a271-764e9d501709_2316x2316.jpeg"}],"coverImagePalette":{"Vibrant":{"rgb":[132,148,196],"population":1},"DarkVibrant":{"rgb":[42.98571428571428,54.64285714285713,89.61428571428573],"population":0},"LightVibrant":{"rgb":[212,175,252],"population":3246},"Muted":{"rgb":[132,108,156],"population":5},"DarkMuted":{"rgb":[102,95,123],"population":149},"LightMuted":{"rgb":[196,221,186],"population":16}},"publication_id":1020561,"publisher_image_url":"https://substackcdn.com/image/fetch/$s_!-qGd!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83350fdd-260e-4a76-8bd7-8337a0e41ca1_500x500.png","publisher_name":"The Overflowings of a Liberal Brain","is_personal_mode":false,"like_count":88,"comment_count":52,"reaction":false,"tracking_parameters":{"is_saved":false,"is_seen":true,"post_id":200286894,"post_type":"newsletter","publication_id":1020561,"tabId":"home","tabType":"base","max_read_progress":1,"max_audio_progress":0,"max_video_progress":0,"last_seen_at":"2026-06-08T21:51:47.072Z","impression_id":"ffa02b9b-c9d3-42fe-a081-392918c8576c"}},"is_saved":false,"saved_at":null,"is_viewed":true,"read_progress":0.18209866848666498,"max_read_progress":1,"audio_progress":0,"max_audio_progress":0,"video_progress":0,"max_video_progress":0,"restacked":false},"postSelection":null,"postSelectionTheme":null,"postImageSelection":null,"clipInfo":null,"mediaClip":null}],"name":"Brad DeLong","user_id":16879,"photo_url":"https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fea5ae644-9822-4ca5-ac6b-e18c017d8fbc_1189x1208.png","user_bestseller_tier":1000,"userStatus":{"bestsellerTier":1000,"subscriberTier":10,"leaderboard":null,"vip":false,"badge":{"type":"bestseller","tier":1000},"subscriber":null}},"source":null,"forumChannel":null}" data-component-name="CommentPlaceholder">

Subscribe now

Leave a comment

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…##the-public-sphere–civility-hoisted-from-the-archives
##public-reason
##hoisted-from-the-archives
#the-public-sphere–civility
#andrew-sullivan
#ta-nehisi-coates
#henry-farrell
#bad-faith-argument
#racial-inequality
#blogging-era
#fifth-column-rhetoric

Stage Theories After Steampower: Meek, Marx, Hicks, the Info‑Bio Tech-Attention Economy, & much much MOAR…

Yet another outtake from my Hicks Lecture <>. Apropos of Ronald Meek and “stage theories”. (Largely) behind the paywall because it is, for now, unfinished and raw, and I want to option to revisit and revise it without it being widely distributed.

Share

I. Ronald Meek & Stage Theories:

Joan Robinson’s “Open Letter from a Keynesian to a Marxist” (1971) <https://braddelong.substack.com/p/reading-joan-robinson-an-open-letter> was, originally, perhaps, going to be a private letter directed to—or rather against—Ronald Meek. Meek was never a Marxist proof-texter. But the version of Meek whom Joan Robinson decided she had to write against was one.

I know Ronald Meek overwhelmingly as the author of the brilliant 1971 article “Smith, Turgot, and the Four Stages Theory” <https://link.springer.com/chapter/10.1007/978-1-4899-7303-0_2>. In that piece Meek does talk about his “revolutionary youth”. He does, however, focus on the stage theory of the Scottish and French Enlightenments, being "impressed in particular by John Millar, whose work was pervaded by a theory of history and society which seemed to me to be a kind of preview of the materialist conception of history upon which I had been brought up…” That is: a stage theory of history in which it is the transformation of the underlying mode of production that is the key and the principal driving factor. For the Scottish and French Enlightenments, it was hunting, herding, agriculture, and commerce. For Marx (but more so for Engels!), it was feudal, bourgeois, and fully automated luxury communism—or agrarian, commercial-imperial, and steampower.

I find the Scottish and French Enlightenments’ theories to be noteworthy because they were much more sophisticated than your standard stage theory. There are a lot of stage theories that shrink until there are only two stages. Consider:

• my friend Daron Acemoglu: extractive → inclusive institutions.

• my friend Deirdre McCloskey: aristo-heroic → bourgeois virtues.

• Our Great Uncle Karl Polanyi: embedded → disembedded.

• Karl’s brother Michael: customary → mercenary + fiduciary institutions.

Walt Rostow’s non-communist manifesto at least preserves three stages: takeoff → drive to maturity → age of mass consumption. And Marx and Engels have four: ancient → feudal → bourgeois → socialist.

Give a gift subscription

II. John Hicks’s Verion of Stage Theory

But, I argued in my Hicks Lecture <https://braddelong.substack.com/p/theories-of-economic-history-ii-stage>, John Hicks had a more sophisticated stage theory than even Marx and Engels, for he was express that it was contingent, not inevitable.

John Hicks’s major use of his stage theory was to draw this contingent bottom-line conclusion: Our current prosperity rests on a long run of unlikely luck. The market system spread far further than it had any right to—widening the division of labor and enabling large‑scale investment—but always as a fragile tendency, never an inevitability, dependent on patchy supporting institutions and limited in geography.

Hicks also argued that fixed‑capital industrialization—what he saw as the core engine of modern prosperity—needed more than the market system. It needed both unexpected scientific breakthroughs, and unusually deep financial institutions to induce investors to sink resources into assets they could not cash out when panic hit. He thought that the growth of such an inequality producing system was unlikely given the political tensons it would bring, for the system could not, by itself, generate broad wage gains until either the W. Arthur Lewis rural labor surplus had been exhausted or unions had grown strong enough to force rent‑sharing for at least a labor aristocracy.

We know more than Hicks did about how history is full of reversals. We now know more than Hicks did about the late Bronze Age collapse after -1200, when the Greeks forgot how to write; about the post‑Song retreat of China’s iron production; and about the post‑200 “Late‑Antiquity Pause” that left people in both Europe and China, by 750, staring back in awe and wonder at the ruins of the Hellenistic, Roman, and Han achievements and judging themselves unworthy descendants.

We thus have more good reasons than Hicks did to insist that the origins, evolution, and future of this whole process were, at bottom, a precarious political‑sociological matter.

Leave a comment

III. Meek’s Assessment of Scottish & French Enlightenment-Era Stage Theories

But let me turn to what Meek concluded about the very sophisticated Enlightenment-Era stage theory when he turned his full mind to it.

Two editorial projects—a new set of Smith’s Glasgow jurisprudence lecture notes discovered in Aberdeen, and Turgot’s early “sociological” manuscripts from his Sorbonne period—led Meek to his central belief that the four-stages theory was very important for Enlightenment-Era political economy, and that it was an inevitable next step from Montesquieu, made multiply and independently by Smith, Turgot, Kames, and others.

It has two core ideas: First, human societies naturally develop through four normally consecutive stages, each defined by a dominant mode of subsistence: (1) hunting (2) herding, (3) farming, and (4) commerce. Second, to each stage there corresponds a characteristic system of:

• property rights,

• form of civil government,

• legal code,

• state of manners and morals, and

• social surplus structure.

Because the mode of subsistence drives everything else, the historian who grasps the succession of modes can explain the succession of institutions—and do so without resort to great men, divine providence, or pure contingency. As Adam Smith put it in his jurisprudence lectures:

Property and civil government very much depend on one another. The preservation of property and the inequality of possession first formed it, and the state of property must always vary with the form of government…

Meek reconstructs the emergence of the theory chronologically:

• Turgot: two Sorbonne lectures in 1750 and the “Plan of Two Discourses on Universal History” (c.1751) already contain a fairly advanced three-stages theory with hints of a fourth.

• Kames: the first appearance of the theory in print in Historical Law-Tracts (1758), where four stages suddenly appear in a footnote.

• Quesnay and Mirabeau: in 1763, they give the most explicit statement in Rural Philosophy.

Then, in the late 1760s and 1770s: Ferguson’s Essay on the History of Civil Society (1767), Robertson’s History of the Reign of Emperor Charles V (1769), Millar’s Observations Concerning the Distinction of Ranks in Society (1771), and Kames’s Sketches of the History of Man (1774), with Adam Smith thinking and talking along the same lines (but not publishing) probably from as early as Turgot.

Meek finds what he sees as Turgot’s and Smith’s contemporaneous and independent formulations as ”one of the most remarkable coincidences in the whole history of social and economic thought.”

What explains the simultaneous emergence? Meek identifies three intellectual currents that converged in 1750:

1. The Pufendorf-Locke property tradition: Pufendorf had argued that things passed into proprietorship “not all at one time, but successively,” with different modes of property corresponding to different modes of subsistence. Locke’s chapter on property, with its North American illustrations, extended this. Hume, Hutcheson, and Kames carried it on.

2. Studies of North American Indian tribes: Charlevoix and Lafitau provided a contemporary empirical laboratory, not just showing a striking transatlantic contrast but also revealing resemblances between the North American Indians and the ancient Greeks and Romans. This comparative-functionalist insight was the empirical rock on which theory could be built.

3. The providential history tradition. Bossuet’s Histoire universelle had organized history around a succession of religious epochs. Turgot (and, more distantly, Smith) reacted against this tradition but were shaped by it: Turgot substituted socioeconomy for religion, secular developmental necessity for divine providence, and material “chains of causes” for God’s plan.

Meek closes by crediting Montesquieu as the “green light” that gave permission for a new science of society, while insisting that Turgot and Smith were “the real Newtons”.

And Meek never discusses Marx at all in this paper, save for the initial reference to his “revolutionary youth”.

Get 75% off a group subscription

IV. Stage Theories as of Marx’s Time

But Meeks’s paper does have a Marxist subtext. Meek’s organizing question is whether the four-stages theory constitutes “a, if not the, materialist conception of history,” and his answer is essentially yes, albeit with important qualifications.

The four-stages theory gave Marx:

1. The principle that the dominant mode of material life—what Smith and Turgot called “mode of subsistence” and Marx renamed “mode of production”—determines the superstructure.

2. The property-government nexus: civil government is not prior to property relations but derivative from them.

3. The concept of social surplus and its role in generating new social classes, towns, arts, and manufactures.

4. 4. Stage logic itself: history as a succession of qualitatively distinct formations, each with its own internal coherence.

5. The “sociological” mode of explanation: individuals make history, but not as they please. The developmental laws of each stage work through agents who do not fully comprehend them.

Moreover, I read Meek’s paper as an attempt to build a firebase for an intellectual struggle against Marxists.

Read more

New empirical evidence on AI‑boosted coding, the apparent absence of a matching surge in valuable products, and the persistent unreliability of large models in real‑world use. Burn‑Murdoch’s chart and Noah Smith’s “tokenmaxxxing” frame a world where we’re flooding the zone with more software artifacts and software-created information streams no one particularly needs…

Noah Smith directs us to this:

Give a gift subscription

It is from Demirer & al. <https://www.nber.org/papers/w35275>, redrawn by the FT team and highlighted by:

John Burn-Murdoch: How Much Value Is AI Really Creating? <https://www.ft.com/content/8e9ae7a4-7209-4e2c-aa36-f3af77d6ce1f?syn-25a6b1a6=1>: ‘One particular point of tension between AI’s boosters and detractors has been the disconnect between reported increases in coders’ output and the apparent lack of a corresponding boom in product or value creation…. Explosive impact at the top of this funnel — coders created or edited almost 300 per cent more files — but that boost was halved to 150 per cent by the time they got to the number of discrete pieces of work submitted for review, and that in turn shrunk fivefold to a roughly 30 per cent uplift in the number of full software releases…. Moreover… the marked increase in mobile app releases over the past year has not been accompanied by any increase in downloads—most of the new apps fail to capture even a modest audience….

Uber CEO Dara Khosrowshahi revealed the company had blown through its entire AI budget for 2026 in one quarter, and was planning to switch much of its AI use to lower-cost models, reserving frontier tools for special cases…. But realised productivity gains are capturing the interaction of powerful new tools with poorly suited structures and processes. Those frictions and bottlenecks will only ease over time…

Leave a comment

That last is interesting. Colin John Burn-Murdoch remains an AI bull over the long run. As does Noah Smith, who comments:

Noah Smith: Tokenmaxxxing <https://www.noahpinion.blog/p/roundup-83-i-told-you-so>: ‘Demand for many of the things that generative AI produces might be a lot more inelastic than we thought. The things we really want a lot more of may not actually be the things that generative AI is yet equipped to provide. As the AI industry advances, of course, that will probably change…

Get 75% off a group subscription

Also: <https://next-media-api.ft.com/renditions/17613947027010/1280x720.mp4

Refer a friend

Brad here: Most of the Silicon Valley tech bros see “AI” as us building increasingly sophisticated digital brains that do the kinds of things that human beings do, and will someday very soon now replace them, in that whatever one would have sent to a human one will soon be able to send straight to a ‘bot.

I see MAMLMs as things that are very, very good for very high-dimension, very big-data, very flexible-function classification, regression, and prediction analysis. Plus “agentic” models are good for interaction speed: Clever Hans at scale trying things that fail with high probability and then retrying them with inexhaustible patience much more rapidly than a human could. These have tremendous uses—natural-language interfaces, summarization engines, and coding assistants being the first three that we have found.

Plus, even with coding assistants, they are not reliable. They cannot be let off of their leashes. My most recent in-message:

Hardware Status Report: As of 12:00 PM, machine is under moderate GPU load due to ollama inference. Total chip power draw is 350W.

• GPU utilization: 68% (ollama inference)

• ANE utilization: 0%

• CPU clusters: Cluster 0 active, Clusters 1-7 power-gated

• Memory: 92GB used / 128GB, swap 0MB, compressor significant

• Top processes by memory:

  • Ollama: 56.3GB (model: mistral-nemo-16k)

  • Python: 3.4GB

  • VSCodium: 1.2GB…

From the bottom:

1. Ollama is running both google-gemma-4-26b and mistral-nemo-16k—not mistral alone.

2. I do not know how the 18 CPU cores could possibly be divided into 8 “clusters”, and it is certainly not true that 7 of these 8 “clusters” are powered-down right now.

3. The M5Max chip is not drawing 350W of power.

For the foreseeable future, it is as silly to think that you will be able to drop one of these things into a job that a human now does as to think that a hotel door equipped with an automatic electric eye could replace a hotel doorman:

Rory Sutherland: Alchemy <https://www.goodreads.com/work/quotes/46190767-alchemy-the-surprising-power-of-ideas-that-don-t-make-sense>: ‘First you define a hotel doorman’s role as ‘opening the door’, then you replace his role with an automatic door-opening mechanism. The problem arises because opening the door is only the notional role of a doorman; his other, less definable sources of value lie in a multiplicity of other functions, in addition to door-opening: taxi-hailing, security, vagrant discouragement, customer recognition, as well as in signalling the status of the hotel. The doorman may actually increase what you can charge for a night’s stay…

MAML GPT LLM are very fast, very flexible tools, not brains; they open doors, but they are not the doorman.

Subscribe now

Leave a comment

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…##ai-is-eating-platform-monopolist-free-cash-flow-not-the-world-chart-of-the-day
##mamlms
##subturingbradbot
##chart-of-the-day
#ai-is-eating-platform-monopolist-free-cash-flow-not-the-world-chart-of-the-day
#john-burn-murdoch
#ai-hype-cycle
#software-development
#coding-assistants
#agentic-models
#clever-hans