Disrupt Yourself
Somewhat scary news
Anthropic just announced Project Glasswing, a coalition of AWS, Apple, Google, Microsoft, and others built around an unreleased AI model called Claude Mythos Preview. The model found thousands of previously unknown security vulnerabilities across every major operating system and web browser. One bug in OpenBSD had gone unnoticed for 27 years. Another in FFmpeg survived five million automated tests.
The cybersecurity angle matters, but what caught my attention were the software engineering benchmarks. Mythos scores 93.9% on SWE-bench Verified, which tests whether a model can take a real GitHub issue, navigate a codebase written by other people, and submit a working fix. The jump from the previous best model (80.8%) is significant in practical terms. The core work of a junior to mid-level software engineer, taking a ticket and shipping a fix, is now something a model can do reliably and fast.
The closest historical parallel we have for what comes next is what happened to American factory workers. What happened to them is not comforting.
The numbers
In 1950, manufacturing was 30% of American jobs. Today it's about 8%. Knowledge workers are roughly 44% of the American workforce today and represent about 40% of GDP. The exposed population is larger this time and higher-paid.
That manufacturing decline played out over decades, and the people who lived through it left behind a lot of data about what actually happens when technology makes your job obsolete.
The Federal Reserve Bank of Chicago studied job displacements between 1989 and 2019 among workers ages 25 to 55 who had worked at least two years full time with their employer before being displaced. These were people with real attachment to their jobs, not casual workers drifting between gigs. In the first year after displacement, they experienced about a 60% drop in earnings. Ten years later, they still earned roughly 25% less than peers who weren't displaced. And that lingering gap came from displaced workers landing in lower-paying jobs, not from working less.
A 2025 study from the Centre for Economic Policy Research (a European economics research network) looked closer at who ended up where. The distribution of outcomes was sharply skewed. A subset of displaced workers experienced catastrophic, persistent earnings declines. These were the people the researchers called "casualties." They experienced prolonged unemployment and unstable, low-paying jobs. Their transitions were delayed and erratic. Years later, they were still worse off than where they started, and they dragged the entire average down. The casualties tended to be lower-wage workers who got pushed out of manufacturing and into low-knowledge service work where they had no leverage and no premium for what they knew.
Goldman Sachs published a report this week that extended this picture to technology-driven displacement specifically, drawing on 40 years of individual-level data from the National Longitudinal Surveys. Workers whose jobs were eliminated by technology saw their real earnings grow nearly 10 percentage points less than non-displaced workers over the following decade. They accumulated less wealth. They delayed buying homes. They were less likely to be married.
But across all of these studies, the people who came through displacement best had a few things in common.
The machinist who understood metallurgy and tolerances could become a CNC programmer or quality engineer. The one who just knew which lever to pull at which time got replaced by the machine that pulled the lever. This pattern held across industries and decades. Workers with transferable understanding of why a system worked could adapt when the system changed. Workers whose value was in executing a specific process couldn't.
They had relationships and trust. The foreman who knew every customer's quirks, who could negotiate with suppliers, who understood the politics of the shop, that person found a role even when the technical work shifted underneath them. Workers who were purely individual technical contributors with no relational capital were more exposed.
Identity
Anne Case and Angus Deaton's research on "deaths of despair" found that life expectancy for white Americans without college degrees started declining around 2000. Suicide, drug overdoses, and alcoholic liver disease all rose sharply. Case and Deaton connected this directly to the collapse of stable manufacturing work.
Their argument is that the loss of a job is also the loss of a way of being. In steelworking towns, in auto plants, the work defined who people were. It structured families, communities, friendships. Case and Deaton wrote that what brings on despair is the loss of meaning, of dignity, of pride, of self-respect that comes with the loss of marriage and community. The money mattered, but the identity mattered more.
Researchers at Youngstown State University's Center for Working-Class Studies found the same dynamic at the community level. Cities like Pittsburgh, Detroit, and Lowell had built their entire identity around a single industry. When that industry contracted, the community lost its story about itself. Population declined, institutions weakened, and people started to see their own town as a place that had failed.
This is the part of the factory worker story that should concern knowledge workers most. If your identity is "I'm a developer" or "I'm a writer" or "I'm an analyst," you're in the same position as the steelworker whose entire self-concept was fused to the mill.
My reflections
I've been thinking about this through the lens of something companies figured out a long time ago. Netflix ate Blockbuster by disrupting the model. But the companies that survive long term are the ones that learn to disrupt their own model before someone else does it for them. Apple killed the iPod with the iPhone. Amazon built AWS while everyone still thought of them as a bookstore.
People need to learn to do the same thing to themselves. Disrupt your own model of what you do.
What I mean is: don't become what you do. Don't let your identity harden around a specific set of tasks or tools. The moment you define yourself by the thing you produce rather than the problems you care about, you've made yourself easy to replace. The factory workers who thought "I am a steelworker" broke when steel left. The ones who thought "I'm someone who builds things and solves problems, and right now I do that with steel" had room to move.
I think about my own career and I've done a lot of different things already. My identity keeps growing and encompassing more things, and that feels right for this moment.
I think this might be the era of the generalist. The era where the person who can operate across domains, who can connect ideas from different fields, actually has an advantage over the specialist who went deep on one thing that a model can now do. AI tools reward you for having range, because the bottleneck is no longer execution, it's knowing what to execute. It's taste, direction, and context.
I also think this era will reward less hierarchy. The people who will thrive are the ones who can get themselves firing on all cylinders across many domains, using AI to amplify their creativity rather than waiting for someone above them to hand down instructions. Some people are going to use these tools to become wildly effective in ways that would have required entire teams before. Those people won't necessarily be the most senior or the most credentialed. They'll be the ones who are curious enough to keep learning and comfortable reinventing what they do every few months.
This is really about a learning mentality. Being consistently open to becoming better and more creative at what you're doing. "Better" means different things for different people, whether that's picking up new tools, going deeper into a domain, or getting better at working with people. And the key isn't learning about AI. It's using AI to learn about everything. The people who figure that out are going to be very hard to displace.
Sources:
Project Glasswing announcement — Anthropic
Goldman Sachs: 40 years of 'scarring' effects of technological displacement — Fortune, April 2026
The High and Lasting Costs of Job Displacement — Federal Reserve Bank of Chicago, 2024
Revisiting the Consequences of Job Displacement — CEPR / VoxEU, 2025
Deaths of Despair and the Future of Capitalism — Anne Case & Angus Deaton, Princeton University Press, 2020
The Social Costs of Deindustrialization — YSU Center for Working-Class Studies
The Professional and Technical Workforce: By the Numbers — Department for Professional Employees, AFL-CIO
Forty Years of Falling Manufacturing Employment — Bureau of Labor Statistic
- https://www.anthropic.com/glasswing
- https://fortune.com/2026/04/06/goldman-job-displacement-40-years-of-data-scarring-gen-z/
- https://cepr.org/voxeu/columns/revisiting-consequences-job-displacement
- https://ysu.edu/center-working-class-studies/social-costs-deindustrialization
- https://www.dpeaflcio.org/factsheets/the-professional-and-technical-workforce-by-the-numbers
- https://www.bls.gov/opub/btn/volume-9/forty-years-of-falling-manufacturing-employment.htm