Stanford's 20% Cliff: The 22-25 Cohort Meta Just Cut, Series A Should Hire

Stanford HAI's 2026 AI Index, released April 13, contains a single chart that should reshape how every Series A and Series B founder thinks about engineering hiring this quarter: employment for software developers ages 22 to 25 has fallen nearly 20% since 2024, while headcount for developers in their 30s and 40s kept growing. Meta's May 20 cuts (8,000 layoffs, 6,000 cancelled reqs, recruiting and non-AI product teams specifically named) and Newsom's May 21 AI executive order made it official this week: the junior ladder at Big Tech is not paused, it is gone. The talent is still out there. Your Boolean string just cannot see it.

What Stanford actually measured

The number that matters is in Chapter 4 of the AI Index, built from an ADP payroll cross-reference reportedly led by Erik Brynjolfsson and Tyna Eloundou. It is not a survey. It is payroll data. And it isolates a single cohort, 22 to 25 year old software developers, against the rest of the developer labor market.

20%

The mechanism is not "AI replaced software engineering." Developers in their 30s and 40s are growing. The mechanism is that AI replaced the specific tasks juniors were hired to do: boilerplate, CRUD endpoints, scripted tests, routine bug triage. Senior engineers now keep that work themselves with Copilot and Claude Code instead of handing it down. The entry rung got sawed off.

Stanford's executive surveys say it gets worse. Planned reductions in service operations, supply chain, and software engineering outpace what has already happened. The 20% is not the floor.

The aftershocks this week

Three signals landed inside seven days, and they all point the same direction.

Meta, May 20. Roughly 8,000 people cut, another 7,000 reassigned into the AI pivot, and 6,000 open requisitions cancelled. The cancelled reqs are the loudest tell. Layoffs can be cyclical; cancelling reqs means the org chart itself no longer has those seats. Recruiting, sales, middle management, and non-AI-adjacent product took the hit. Recruiting absorbing 35 to 40% of the cuts is why your inbound referrals from FAANG university programs went quiet: there is no one inside left to refer juniors out.

Newsom, May 21. Executive Order N-6-26 tells state agencies to come back in 180 days with proposed updates to California's WARN Act, explicitly to make it work as an early-warning system for AI-driven displacement. Read that as a sourcing signal, not a policy story. If the WARN update lands, CA companies will have to publicly file AI-driven layoffs earlier and in more structured detail. That is a free, public lead list of recently released early-career engineers, refreshed weekly.

Layoffs.fyi, year to date. Roughly 110,000 cuts at 137 tech companies in 2026 already, after about 125,000 across all of last year. The pace doubled.

The compounding effect: a structural Stanford finding, a flagship company confirming it on payroll, and the largest state tagging the trend as worth a legal instrument. All in five weeks.

Why this is a Series A and B opportunity, not a tragedy

The cohort being released is the most AI-fluent end-user cohort in the labor market. Generative AI hit 53% population adoption in three years, faster than the PC or the internet, and the 22-25 cohort learned to code with Copilot and Claude as peers, not threats. They do not need to be retrained on agentic workflows. Their baseline includes them.

That is exactly the engineer a 20-person startup wants. Big Tech is optimizing for senior researchers with $1.5B packages on one end (see Meta hiring Alexandr Wang as Chief AI Officer through a $14.3B Scale AI deal) and pure throughput on the other. The middle, the generalist who can ship a feature end to end with an AI agent in the loop, is the exact profile Series A needs and the exact profile Big Tech just defined out of its org chart.

Survivor comp at Meta also helps your pitch. Median total comp fell from $417,400 in 2024 to $388,200 in 2025, with another 5% stock cut to annual raises in February 2026 on top of last year's 10%. The "stay at Meta vs. join a Series B" math is closer than it was 18 months ago.

The cohort is not unemployed. It is mistitled, miscoded in payroll, and invisible to a Boolean string written for the 2019 career ladder.

Why your sourcing stack misses them

Three structural reasons, all fixable.

LinkedIn titles describe the old ladder

The 22-25s who landed somewhere did not land as "Software Engineer I." They landed as "AI Engineer," "Forward Deployed Engineer," "Solutions Engineer," "Applied AI," or as the second hire at a 12-person agency where their title is whatever the founder typed into Gusto. Nucamp's analysis of 2026 grad placement shows the volume moved to mid-sized SaaS, healthcare and government tech, agencies, and professional services. Boolean for ("Software Engineer" OR "Junior Developer") AND ("2023" OR "2024" graduate) returns the 20% who got the legacy title. The other 80% are invisible.

GitHub is the honest signal now

When the resume floor moves, the signal moves with it. Two years ago, "clean GitHub, three portfolio projects" was enough for a first-round at a Series B. Today those repos get scrolled past because the floor is shipped systems with auth, billing, observability, and a deploy pipeline, not a to-do app clone. The candidates worth talking to have commit histories that show this. Their pinned repos have CI badges, real READMEs, and traffic.

This is the part Boolean cannot do. You cannot grep "shipped a working SaaS with Stripe and Postgres while in undergrad" out of LinkedIn. You can read it off a GitHub profile in 30 seconds, but you have to find the profile first. That is the friction we built Refolk to remove: describe in plain English the kind of repo history and project shape you want, and get a ranked shortlist across GitHub, LinkedIn, and the open web in one pass.

The candidate pools are now community-maintained

The affected cohort self-organized. Two repos to watch this week:

• vanshb03/New-Grad-2027: US, Canada, and remote SWE, quant, and PM roles, maintained jointly with WeCracked and Resumes.fyi.
• jobright-ai/2026-Software-Engineer-New-Grad: the 2026 equivalent.

The job lists are useful. The contributor graphs are more useful. Every person who PRs a role into one of those repos is a candidate who is actively sourcing themselves, with a public GitHub handle attached. Most recruiters are reading the README. Read the contributors tab.

A concrete sourcing plan for this quarter

Six moves, in order. None of them require LinkedIn Recruiter seats.

1. Reframe the title net. Replace "junior software engineer" filters with intent signals: graduation year (2023-2025), university or bootcamp, plus current company size under 200. Do not filter by title at all. Use Refolk's plain-English query layer to express this: "engineers who graduated in 2024, currently at companies under 200 people, with at least one shipped LLM project on GitHub." That captures the FDEs, the AI Engineers, the Solutions hires, and the unlabeled second-hires at agencies in one pass.

2. Mine the new-grad repos as a contributor list. Pull the last 90 days of contributors to vanshb03/New-Grad-2027 and jobright-ai/2026-Software-Engineer-New-Grad. Enrich each handle with their pinned repos, languages, and most recent employer. This is a few hundred high-intent candidates per week that no one else is treating as a list.

3. Watch California WARN filings weekly. Even before the EO's 180-day window closes, current Cal-WARN filings name the company, the date, and (often) the affected functions. Cross-reference against LinkedIn alumni who joined in the last 24 months. That intersection is your warm outreach pool.

4. Score on system completeness, not project count. When you look at a repo, ask: is there auth? Is there a payment path? Is there a deploy target that is not localhost? Is there at least one observability hook? Three repos that answer yes beat ten that answer no. The 22-25 cohort that survived the cliff is the one that already learned this.

5. Skip the take-home. Read the commit history. A six-month commit graph with steady, meaningful commits on a real project is a stronger signal than a four-hour exercise. The cohort you are sourcing has been told this their whole short career; honor it and you double your reply rate.

6. Time the outreach to comp cycles. Meta's February stock cut and the general Big Tech raise compression mean retention conversations are happening internally in May and June. Outreach this quarter lands differently than it would have in Q4 2024.

This is the second mention worth making: the bottleneck in steps 1, 2, and 3 is the same bottleneck. You need to translate a fuzzy human description ("AI-native 2024 grad, shipped real systems, currently underpaid at a 30-person agency") into a list across three different surfaces (GitHub, LinkedIn, the open web). Refolk does that translation in one query, which is the entire reason it exists.

The bifurcation, in one image

Meta is paying up to $1.5B for a single elite researcher and cutting 8,000 generalists in the same quarter. The middle is gone. The 22-25 cohort is the missing middle. Stanford named it. Newsom legislated around it. Layoffs.fyi counts it weekly.

If you are hiring engineer five through engineer twenty at a Series A or B this year, the cohort that just got cut from FAANG is the most AI-fluent talent pool you will see in your career, available at compensation that finally makes sense, and structurally invisible to the sourcing stack everyone else is using. Source them by commit shape and intent, not by title. The window where this is a secret closes when LinkedIn ships an "AI-native junior" filter, which on their cadence is at least two quarters away.

FAQ

Is the Stanford 20% number specific to Big Tech, or is it the whole market?

It is a payroll-wide measurement of US software developers ages 22 to 25, built off ADP data, not a Big Tech subsample. But the drop is concentrated in roles AI most directly substitutes for (boilerplate, CRUD, routine testing) and in companies that built the tightest junior-to-senior pipelines, which skews the visible effect toward Big Tech. The mid-market 1.6% hiring increase for grads in healthcare, gov tech, and SaaS is where the displaced volume landed.

How do I actually evaluate a 2024 grad if take-homes are out?

Read six months of their most active repo's commit history. Look for incremental, meaningful commits (not one massive initial push), a real README that explains tradeoffs, at least one closed issue they wrote themselves, and evidence the project is deployed somewhere a human can reach. Then ask them to walk you through one architectural decision they would now make differently. That conversation is worth three take-homes.

What about bootcamp grads specifically?

The bootcamp pipeline as a Big Tech feeder is broken, but the talent inside good bootcamps (Bloom, Codesmith, Hack Reactor cohort grads with shipped capstones) is still strong. The filter is the same: did they ship a system with auth, billing, observability, and a deploy target, or did they ship an exercise? Treat the bootcamp credential as neutral and the GitHub as the signal.

Will Newsom's EO actually change WARN filings in time to matter?

The 180-day window closes in late November 2026, and any statutory change would come after that. But the EO's framing already pushes CA companies to disclose AI-driven workforce decisions more publicly. Treat existing Cal-WARN filings as the live signal now, and assume the structured AI-displacement data shows up in mid-2027. Sourcers who build the scraping habit this quarter will be ahead when the feed gets richer.