Summary

OPM expects the civilian federal workforce to shrink by roughly 300,000 people, from about 2.4 million at the start of the year to around 2.1 million at year-end. Scott Kupor says much of the reduction came through voluntary deferred-resignation programs; one of his broader goals is to make operational efficiency a “first-class metric” and remember that “OPM stands for other people’s money.”
The principal brake on federal technology is a system that counts downside meticulously while rarely measuring upside. Kupor calls it a “cult of obsession” with risk, while Greg Barbaccia says, without hyperbole, “I have more auditors auditing my team than I have team members.” Kupor’s alternative is “measured risk”: protect national security and benefit delivery without treating ordinary experimentation as existential danger.
Among rated federal employees, performance ratings are so inflated that only 0.3% score below “meets expectations,” while 65%-70% receive a four or five. OPM is guiding agencies to limit four and five ratings among senior executives to 30%, aiming to stop bonuses and promotions being “peanut buttered out” and to replace tenure-based advancement with measurable merit.
Government will not solve compensation parity with Silicon Valley, so its talent strategy rests on mission, unusually early responsibility and a two-to-four-year tour of duty that preserves a path back to the private sector. Federal technologists could affect more than 300 million people, yet only 7% of government employees are under 30 versus roughly one-quarter outside government. Kupor wants young engineers to serve for two to four years without choosing between 40-year public- and private-sector careers.
A shortage of technical judgment—not simply a shortage of software—feeds both weak hiring and sprawling contractor relationships. Barbaccia’s analogy: anyone would reject a Chevy Suburban contract priced at $900,000 per vehicle, but nontechnical approvers cannot similarly judge a software project’s cost, duration or staffing. Functional assessments, technical managers and proposed private-sector secondments are intended to break that cycle.
The near-term AI opportunity is distributed workflow adoption, not a five-, ten- or fifteen-year master plan. At OPM, after gaining ChatGPT through the GSA’s $1-a-year deal, Kupor asked employees to find 5% efficiency improvements and rejected a three-page acknowledgment form in favor of a few rules on PII and hallucinations. “I don’t think you adopt new technology through white papers.”
The strategic end state is “one government”: lower-cost operations, connected data and a citizen portal that replaces repeated agency-by-agency disclosure. Barbaccia wants consent-based sharing that could surface benefits, tax-return status and passport expiration together; his companion metric is “rework percentage”—how often government performs the same task again.

Deep dive

1. Federal talent policy targets a 2.1-million-person workforce

Kupor’s reason for leaving a16z was a confluence of fiscal and technological urgency: the country is “on this precipice” from overspending, while government is “nowhere near prepared” for rapidly advancing technology. He saw an administration treating AI and merit hiring as real priorities, with presidential backing and permission from Mark Andreessen and Ben Horowitz to serve.
OPM is, in Kupor’s shorthand, the federal government’s talent organization: it sets hiring standards, performance-management policy and the conditions for attracting and retaining strong employees. Its remit excludes uniformed military personnel but began the year with roughly 2.4 million civilians; Kupor expects around 2.1 million by year-end after an approximately 300,000-person reduction.
Much of that reduction, Kupor stressed, came through voluntary mechanisms catalyzed by Elon Musk and the DOGE team, including Deferred Resignation Programs. Job loss remains “a serious issue,” but the programs let employees who did not want to join the “go-forward organization” self-select into leaving.
Barbaccia’s CIO role is broader than its private-sector analogue: he oversees technology policy and budgeting across the executive branch. His organizing principle is “one government”—replacing agency silos with a CIO community capable of moving administration priorities together.

2. Government prices downside project by project and ignores portfolio upside

After roughly 230 days, Barbaccia’s biggest surprise was the complexity of the compliance and regulatory regime. Silicon Valley’s “move fast and break things” does not translate at this level: “Often things do take an act of Congress,” and sometimes officials really must “make a federal case out of something.”
Kupor found something deeper than formal rules: employees defended processes because otherwise “we might get sued,” yet produced blank stares when asked how often suits had happened. Risk had become “this pass-fail thing”—either absent or unacceptable—with no sizing of the downside and no discussion of “how unbounded is the upside relative to that risk.” He also acknowledged that many compliance rules exist for good reasons.
Kupor’s concrete example was Solyndra: one government-backed company failed, and opponents treated it as proof that the administration’s whole position was foolish. The missing questions were how perhaps 50 other projects performed and whether some might succeed wildly. “There’s no concept, I would say, of a portfolio.”
Oversight reinforces that one-project mentality through GAO reports, inspectors general, congressional oversight and partisan “victory laps.” Barbaccia says auditors outnumber his team members; Kupor nevertheless rejects importing full Silicon Valley risk tolerance. His standard is “measured risk”: reserve the strongest caution for catastrophic mistakes involving sovereignty, national security or Social Security delivery, while recognizing most decisions are not catastrophic.

3. Grade inflation makes merit, promotion and removal almost meaningless

The updated performance numbers were worse than Katherine Boyle’s premise. For most employees, on the federal one-to-five scale, three means meeting expectations, yet 65%-70% receive a four or five and 99.7% receive at least a three. Only 0.3% of rated employees are assigned a one or two—nothing resembling a normalized distribution.
One OPM intervention targets the most senior government employees: only 30% may now receive a four or five. Kupor described this as merely “a modicum of a forced distribution”; the guidance does not prescribe how everyone else must be rated, but begins forcing real comparison.
Kupor said the annual FEVS employee survey asks whether managers manage performance effectively and hold people accountable, and that scores are extremely low. Employees generally know who is carrying their weight and who is “dialing it in,” even if formal ratings do not reflect that knowledge.
The consequence is not just grade inflation. Bonuses and promotions get “peanut buttered out,” while managers avoid ones and twos because removal can require 12, 18 or 24 months of appeals. Kupor’s description of the equilibrium: some decide “it’s better to keep a warm body in a job” than attempt removal.
Boyle preserved the central pushback: could limiting top ratings unfairly punish a genuinely excellent cohort or create counterproductive executive competition? Kupor rejected equal contribution as a “fallacy” and argued teamwork should itself be measured. Drawing on a16z’s belief that “teams beat individuals,” he said incentives can reward collaboration rather than isolation while still distinguishing performance.

4. Mission can beat compensation only if service becomes a tour of duty

Barbaccia’s recruiting equation starts with tools, freedom to innovate and sufficiently difficult problems. Government will struggle to compete with Silicon Valley compensation and equity, “but we do win on mission”: employees could affect more than 300 million people and reach consequential work that might require a decade of tenure inside a large company.
Kupor says government has sold precisely the wrong benefit—tenure and “effectively lifetime employment.” His replacement pitch is sharper: solve hard, cutting-edge problems while serving the country. “There’s no such thing as lifetime employment,” he argued, and stability is not the story likely to attract ambitious technologists.
The demographics make the reset urgent. Only 7% of federal employees are under 30, compared with close to one-quarter of the broader workforce; almost 45% are over 50, versus roughly 33% outside government. Kupor framed this as a pipeline problem, not “an ageism comment,” especially as AI adoption accelerates.
His preferred model dissolves the false choice between lifelong sectors: young engineers could spend two, three or four years in government, then return to employers that value the experience. Barbaccia extends the pitch to people after a first exit or several successful private-sector years: government teaches operation amid unusual opacity, and “everything you do after working in tech in the government will be easier.”

5. Technical gaps compound from hiring into procurement sprawl

Government’s missing engineers create a self-reinforcing loop: agencies outsource work, internal technical jobs become contractor-management roles, and strong builders decline to join because they want to make things rather than supervise vendors. Kupor and Barbaccia want early-career technical teams led by private-sector directors, senior directors or VPs serving proposed two-year placements.
Barbaccia’s procurement analogy makes the competence gap tangible. Americans know a Chevy Suburban should not cost $900,000, so that fleet contract would look “objectively insane.” Nontechnical approvers may not be able to similarly judge the appropriate software solution, delivery time, team size or required expertise—giving “Beltway bandits” and systems integrators room for indefinite-delivery, indefinite-quantity contracts and sprawl.
The problem begins before contracting. Nontechnical HR employees have historically screened résumés for numerous unrelated roles, relying substantially on self-assessment: applicants claim the skills listed in the job description, and the government does not verify them. Kupor argued a CEO in a private company might be fired for hiring engineers without a real technical assessment.
Kupor traced that practice to the civil-service exam’s removal following disparate-impact concerns and a consent decree that remained in place for 43 years. After exiting the decree, the administration’s merit-hiring plan now requires functional assessment: coding work for software engineers and role-appropriate tests elsewhere. The accumulated system must be dismantled “each brick one by one.”

6. AI adoption will happen one user and one workflow at a time

Barbaccia’s cultural target is government’s love of “process and ritual.” Completing every procedural box can be celebrated independently of mission results; AI offers a chance to automate repetitive work and redirect employees toward higher-value outcomes. His historical analogy is evolutionary: ledgers gave way to Excel and Post-it notes gave way to email, and roles must now modernize again.
A top-down AI directive from the President’s AI Action Plan gives CIOs institutional cover, but Kupor emphasizes micro-adoption. OPM, an organization of a few thousand people, did not have ChatGPT on employees’ desktops until that week, when a GSA arrangement supplied it for $1 a year. His instruction: learning it is now an obligation, and every employee should find a workflow with even a 5% efficiency gain.
Rulemaking is one candidate: AI should not independently write regulations, but it can disseminate information and help determine what is statutorily required. Kupor worries that a grand federal AI agenda will otherwise become a five-, ten- or fifteen-year plan in which “quite frankly nothing will ever happen.” His call: “I don’t think you adopt new technology through white papers.”
Boyle asked whether federal employees already use ChatGPT personally or require basic prompting instruction. Kupor’s honest non-answer was that he lacked quantitative evidence, but guessed that many used it personally; he said employees had also been warned off at work by a “parade of horrors.” He rejected a proposed three-page memo and signed acknowledgment, preferring a few rules—do not enter PII and verify possible hallucinations—plus two hours of monthly training through free resources. Before official access, some Silicon Valley hires on his team used ChatGPT on personal phones—what Boyle called “swivel-chairing.”

7. The end state is cheaper operations and one citizen-facing government

Kupor wants operational efficiency elevated to a first-class metric because federal power is currently associated with larger budgets and headcount. That incentive predictably produces growth. His alternative rewards continuously improving service at lower cost, grounded in the reminder that “every single dollar we spend is someone else’s money”—hence OPM as “other people’s money.”
His second success condition is cultural: early-career employees in technology, finance or HR should leave saying government offered hard problems, development and a mission, while private employers should recognize and reward that service. If that bridge becomes normal, Kupor said the country could have “the most amazing Pax Americana for the next 50 years.”
Barbaccia’s citizen-facing ambition is to rationalize the “insane” web of government websites. With responsible, consent-based and citizen-centric data sharing, one portal could show benefits, tax-return status and passport expiration rather than repeatedly asking agencies for the same birth date and address. “This is one government. It’s all the executive branch.”
Behind that portal, Barbaccia wants disparate datasets converted into intelligence for decision-making rather than trapped in silos. His lightning-round metric was “rework percentage”—how often government repeats the same work—while Kupor chose operational efficiency.
In the lightning round, Kupor said the myth to retire is that government cannot be a leading technology innovator; his process tweak was to treat risk as a spectrum and weigh upside as well as downside. Their final conditions for American success were equally compressed: “win the AI race” and “remain undistracted.”

Born to be wired (September 2025 Book Club)Periodic Labs: Training AI Scientists, with Liam Fedus & Ekin Dogus Cubuk (from a16z)