CTRL K

my thoughts on the current job market

Rishi Edupalli,Last updated at September 26, 2026•Economics AI

According to the headlines, the labor market is doing just fine. August payrolls came in at +162,000 against a consensus of 53,000, and unemployment held steady at 4.1%.¹

However, if one talks to anyone who graduated recently or looks on LinkedIn for more than thirty seconds (30 seconds too long anyways), one gets a very different picture.

Unfortunately, it appears both are true. My stocks are up, and I’m over here contemplating my life decisions.

the math ain’t mathing

Measure	Value	Source
Unemployment, all workers	4.1% (Aug 2026)	BLS¹
Unemployment, recent grads ages 22–27	5.6% (Q2 2026), up from 3.6% in 2019	NY Fed²
Underemployment, recent grads	~42%	NY Fed²
Unemployment, computer engineering majors	7.8%	NY Fed²
Unemployment, computer science majors	7.0%	NY Fed²
Hires rate	3.2% (Jul 2026)	JOLTS³
Quits rate	1.9% (Jul 2026)	JOLTS³

For most of the history of the NY Fed series, recent graduates had lower unemployment than the workforce as a whole, and a degree functioned as a hedge. That relationship has now inverted, and the gap is the widest on record.⁴ Of note, computer engineering now sits just behind anthropology for the highest unemployment rate of any major, which is not a sentence I expected to write about the “safe” degree. Perhaps I should have paid attention in cultural anthropology.

stocks and flows

We can think of the unemployment rate as a stock (unfortunately, not the kind one can buy, though I’m not sure why one would do that). What matters to someone entering the labor market is a flow.

The simplest model treats each worker as a two-state Markov chain.⁵ Each month, an employed worker separates with probability $s$ , and an unemployed worker finds a job with probability $f$ . The unemployment rate evolves as

u_{t+1} = (1 - f)\,u_t + s\,(1 - u_t) \quad \Longrightarrow \quad u^* = \frac{s}{s + f}.

Note that $u^*$ is homogeneous of degree zero in $(s, f)$ : scale both rates by any $\lambda > 0$ and it does not move. A market where nobody leaves and nobody gets hired posts the same headline number as one that churns constantly.

Now consider someone who starts unemployed. Under the Markov assumption, each month is an independent trial with success probability $f$ , so the time to a first job $T$ is geometric:

\Pr(T = t) = (1 - f)^{t-1} f, \qquad \mathbb{E}[T] = \frac{1}{f}.

This depends solely on $f$ . The low separation rate that protects incumbents does nothing for entrants. Two hypothetical markets, with round illustrative numbers:

Market	$s$	$f$	$u^*$	$\mathbb{E}[T]$
Churning	1.5%	35%	4.1%	2.9 months
Frozen	0.75%	17.5%	4.1%	5.7 months

Same headline, twice the wait. This is what “low hire, low fire” means in practice: nobody is leaving, so nothing opens up, and nobody is being let go, so the headline never shows it.

the part where I admit the model is wrong

The Markov assumption says the chance of finding a job next month depends only on one’s current state, not on how long one has been in it. That is not necessarily true, and every way it fails makes things worse for new graduates.

1. Duration dependence. Kroft, Lange, and Notowidigdo sent fictitious résumés to real postings in 100 U.S. cities and found callback rates fall significantly with the length of the unemployment spell, mostly within the first eight months.⁶ If the hazard $f_k$ declines with $k$ , the expected wait becomes

\mathbb{E}[T] = \sum_{t=0}^{\infty} \prod_{k=1}^{t} (1 - f_k) \; > \; \frac{1}{f_1},

so a slow start compounds.

2. Heterogeneity. Even if every individual were memoryless, a population is not. If some graduates find jobs at rate $f_H$ and others at $f_L < f_H$ , the fast finders exit first, and the aggregate hazard falls over time even though neither group’s does.

3. Nonstationarity. $s$ and $f$ move with rates, the cycle, and technology, so $u^*$ is a moving target and the observed rate lags it.

4. Two states are too few. Most graduates enter from $N$ , not in the labor force, rather than from $U$ . The Dallas Fed finds lower employment among young workers in AI-exposed occupations is “mainly driven by a lower inflow, particularly among those out of the labor force, rather than by an outflow.”⁷ Since $u = U/(E + U)$ , $N$ never appears in the unemployment rate. A graduate who stops searching is invisible to the headline entirely.

why openings are not hires

There are about 7.3 million openings³ and 7.0 million unemployed,¹ roughly one per person, which is not a slack market. In the standard matching function $H = A\,U^{\alpha}V^{1-\alpha}$ , the job-finding rate is $f = A\,\theta^{1-\alpha}$ with tightness $\theta = V/U$ .⁸ If $\theta \approx 1$ and $f$ is still low, the model blames $A$ , matching efficiency. My guess, and it is only a guess, is congestion: postings left open with no intent to fill them, AI-written applications that make it free to apply to hundreds of jobs, and AI screening that filters them back out. Unfortunately, $A$ is usually estimated as a residual, which is a polite way of saying it is the part of the model nobody understands.

where are the minions?

Participation fell. Labor force participation held between 62.4% and 62.7% from early 2023 through late 2025, then dropped about 0.9 points to 61.6% by June 2026. The St. Louis Fed decomposes the drop:⁹

Component	Share of the decline
January 2026 population-control revision	43%
Participation changes within age groups	41%
Aging of the population	16%

Nearly half of the decline, it turns out, is the Census Bureau revising how many people it thinks exist. The aging term is mechanical: with prime-age participation at 83.3% and 65+ at 19.0%, shifting one percent of the population from the first group to the second lowers aggregate participation by $0.01 \times (0.833 - 0.190) \approx 0.64$ points without anyone changing their behavior.

Boomers are aging out. A record 4.18 million Americans turned 65 in 2025, and the pace stays elevated through 2027.¹⁰ The 65+ share of the population went from 12.4% in 2007 to 17.9% in 2024, and is projected to reach 21.2% by 2035. Gallup’s self-reported average retirement age rose from 57 in 1991 to 62 today,¹¹ but a later retirement age applied to a much larger cohort is still a wave. It is also a fragile one: roughly half of boomers turning 65 through 2030 have $250,000 or less in assets,¹⁰ and Social Security’s OASI trust fund runs out in 2032, triggering a 22% benefit cut unless Congress acts.¹² Congress, in its characteristic urgency, last made a major change to the program in 1983.

Immigration reversed. Net unauthorized immigration turned negative in February 2025. Employment is $E = (1 - u)\,p\,N$ , so holding $u$ fixed,

\frac{\Delta E}{E} \approx \frac{\Delta p}{p} + \frac{\Delta N}{N}.

With $N$ flat and $p$ falling, breakeven job growth goes to zero. The Dallas Fed estimates it fell from about 250,000 a month in 2023 to roughly 10,000 by mid-2025, and averaged slightly negative in late 2025.¹³ So +162,000 is a strong number, and +50,000 is no longer a warning sign. The old heuristics are off by an order of magnitude.

One would think the retirements should be opening up space. They aren’t, because retirements vacate senior roles, not entry-level ones, and firms are not backfilling from the bottom.

the corporate journey

Total postsecondary enrollment grew almost fourfold from the mid-1960s to a peak of 21.02 million in 2010, then fell for a decade, dropped sharply during COVID, and has only partially recovered to 19.4 million in fall 2025.¹⁴ Most of the decline was in community colleges and for-profits, while completion rates improved,¹⁵ so the number of degree holders entering the market did not shrink nearly as much as enrollment did. The number of jobs that actually require a degree did not keep up, which is precisely what a 42% underemployment rate is telling us.

The demographic cliff. US high school graduates peaked at about 3.9 million in 2025. WICHE projects a 13% decline to 3.4 million by 2041, driven by the drop in births after the 2008 recession.¹⁶
Students are reading the market. In fall 2025, total enrollment rose 1.0%, but computer science enrollment fell at every award level, by as much as 14.0% at the graduate level, while data science grew.¹⁷ This is a lagging signal: students are reacting to the 2023–2025 market and will graduate into a 2029–2030 one.

why my friends are still unemployed

1. Rates and uncertainty. On September 16th, the Fed raised rates 25 basis points to 3.75–4.00%, its first hike since 2023.¹⁸ CPI is at 3.4%, and gasoline is up 27.4% year over year after the Iran war pushed Brent to about $118 in March, which my wallet has taken personally.¹⁹²⁰ With wages up 3.1%,¹ real wages are roughly $1.031/1.034 - 1 \approx -0.3\%$ .

Uncertainty matters on its own. Hiring is partially irreversible, so if the value of a hire $V'$ is only revealed next period, waiting and hiring only if it turns out well is worth $\mathbb{E}[\max(V', 0)]/(1 + r)$ .²¹ By Jensen’s inequality,

\mathbb{E}[\max(V', 0)] \geq \max(\mathbb{E}[V'], 0),

and the gap widens with the variance of $V'$ . More uncertainty makes waiting more valuable without changing the expected value of the hire at all. Firms freezing is rational.

2. Capital reallocation. The four largest hyperscalers plan up to $630 billion in capex for 2026, up about 62% from 2025,²² and AI data center investment alone was about 0.8% of GDP in Q1 2026.²³ A dollar committed to GPUs is a dollar not committed to headcount. The GPU also does not need a mentor or a return-to-office policy.

3. AI substitution at the bottom. Using ADP payroll data, Brynjolfsson, Chandar, and Chen find employment of 22–25-year-olds in the most AI-exposed occupations fell about 11% from November 2022 to June 2026, while the same age group in the least-exposed occupations grew about 10%. The gap operates through reduced hiring rather than layoffs, and is concentrated where AI automates tasks rather than complements them. The authors note these are descriptive patterns, not causal estimates.²⁴

4. Remote work. NY Fed work by Natalia Emanuel, Emma Harrington (here at UVA), and Amanda Pallais estimates remote work explains about 64% of the rise in young-graduate unemployment, even after controlling for AI exposure. It is considerably harder to mentor a new hire whose face one has only seen through a webcam.²⁵

A junior hire is an investment, not a purchase. With wage $w$ , training cost $c_t$ , and a marginal product that grows as they learn, $MP_t = MP_\infty(1 - e^{-t/\tau})$ , the value of the hire is

V = \sum_{t=0}^{T} \frac{MP_t - w - c_t}{(1 + r)^t}, \qquad \frac{\partial V}{\partial r} = -\sum_{t=0}^{T} \frac{t \,(MP_t - w - c_t)}{(1 + r)^{t+1}}.

The hire only turns cash-flow positive at $t^* = -\tau \ln(1 - w/MP_\infty)$ , so the costs sit at small $t$ and the returns at large $t$ . A junior hire is a long-duration asset, and each explanation hits it:

Rates. $\partial V / \partial r$ weights each cash flow by $t$ , so a higher $r$ shrinks the late returns far more than the early costs.
Remote work. Slower learning means a larger $\tau$ , and $t^*$ scales linearly with it.
AI. The tasks a junior does first (codified, documented, well-specified) are the ones models do best, so the junior’s incremental product over a model is smallest exactly when they cost the most.

A senior hire’s cash flows are positive from $t = 0$ , so they are far less exposed, consistent with the Stanford data showing steady or rising employment for experienced workers.²⁴ The bottom rung of the ladder goes first because it has the longest payback period.

To be fair to the other side, a trained junior can also leave, which acts like an extra discount rate equal to the quit rate. Quits are low right now, at 1.9%,³ which should make juniors more attractive. That is why I do not think this model explains everything. I think it explains most of it.

should i drop out?

These are my predictions, and I would not bet my (nonexistent) salary on any of them.

The headline stays calm. With breakeven near zero and separations low, unemployment can sit between 4% and 4.5% while entry-level hiring stays depressed. Watch the hires rate and the recent-grad gap, not the headline.

The graduate gap persists through at least 2027. The Fed’s median projection has rates at 4.1% by year end.¹⁸ Nothing in the $V$ equation is getting better soon.

The missing rung shows up in five to ten years. If firms do not hire juniors now, there will be no mid-level people in 2031. Wages for experienced workers in AI-exposed fields rise sharply, and the firms that kept hiring juniors have a real advantage.

Higher ed contracts. Fewer 18-year-olds means fewer freshmen. Small privates and regional publics without large endowments will consolidate or close.

Retirement becomes a political fight. OASI depletion in 2032 forces a choice: raise taxes, cut benefits, or raise the retirement age past 67. The last option keeps older workers in senior roles longer and slows the vacancy chain further.

The labor market is not broken so much as frozen at the entry point. The people inside it are mostly fine, and the people outside it are bearing nearly the entire cost of the adjustment.