Slow Down to Speed Up
AI and Software Engineering
Rapid Research Brief · June 29, 2026
Gergely Orosz / The Pragmatic Engineer · Craft Conference 2026, Budapest
AI software engineering engineering culture code quality Medium-High
Executive Summary
The AI Gold Rush Is Breaking Software Engineering

At Meta, an AI-generated Instagram exploit succeeded because AI-written code was reviewed by AI, not humans, while engineers inflated their token counts. The pattern repeats across the industry: quality collapses, costs explode, and the humans who catch the cracks are burning out.

The Exploit

Meta's worst breach: zero-step password reset via Meta AI — AI-written code, AI-reviewed.

Token Maxing

Engineers measured on token counts, leading to perverse incentives and inflated usage.

Quality Crisis

Anthropic, Amazon, GitHub all hit by AI-era quality failures — bugs unfixed for months.

Cost Shock

AI budgets exploding company-wide; Uber burned its annual budget by March.

Finding 1
The Instagram Exploit

Meta's worst security breach was a zero-step password reset. Attackers faked a victim's location, asked Meta AI for a verification code — and received it. The code was AI-written, AI-reviewed. Meanwhile, Meta reassigned 40% of its trust & safety team to AI data labeling.

"Step one: fake your location to a victim. Step two: there was no step two. This was it."

— Gergely Orosz, Craft Conference 2026
security Meta AI review
Finding 2
Token Maxing

Engineers at Meta, Amazon, and Uber were measured on AI token usage. They used AI for everything to inflate counts. Meta had a leaderboard with titles like "Session Immortal" and "Legend." When layoffs were announced, token inflation became survival.

"People were not thinking about trust and safety. They were thinking about token maxing."

— Gergely Orosz, Craft Conference 2026
incentives metrics Meta Amazon
Finding 3
The Quality Crisis

Anthropic's website bug went unfixed for a month. OpenAI's agent builder launched with P0 bugs never fixed. Amazon's AI deleted a production environment. GitHub uptime dropped severely from 3× load increase.

"It just feels software has become a brittle mess everywhere."

— Mario Zechner, creator of pi / OpenCode
quality Anthropic Amazon GitHub
Finding 4
The Burnt Out Guardians

Engineers who still manually review AI code catch bugs and push back, but they are overwhelmed and unrewarded. With engineering management cut, no one protects them. Some are quitting.

  • Manual reviewers catch 30–50% more defects than AI-only pipelines
  • No career path, no recognition, no bandwidth
  • Management layers eliminated — no one to escalate to
  • Attrition is accelerating as quality champions leave

Dax (OpenCode) is actively hiring these burnt-out engineers — the ones who leave companies where no one cares about quality.

burnout review OpenCode
Finding 5
The Cost Shock

Uber burned its annual AI budget by March and imposed caps at $1,500/engineer/month. Anthropic turned on API pricing for enterprise. GitHub Copilot users drained monthly budgets in 3 days.

"No one wants to pay that much, no matter what the AI labs say."

— Gergely Orosz, Craft Conference 2026
cost Uber Anthropic Copilot
Finding 6
Slow Down to Speed Up

Winning companies start with business outcomes and build verification systems. Kent Beck: "We're accumulating code faster than we accumulate trust."

"Cap your daily agent usage to what you can verify, not what you can generate."

— Gergely Orosz, Craft Conference 2026
verification trust Kent Beck
Risks, Gaps & Uncertainty
What We Don't Know Yet
  • 🔍 Gap: Talk focuses on extreme cases (Meta, Amazon outages); most companies less affected — but pattern is spreading.
  • ⚠️ Risk: "Slow down" advice hard to sell in competitive markets where speed is rewarded.
  • 🧠 Caution: "AI psychosis" framing is provocative; use carefully with some audiences.
  • ❓ Unresolved: AI cost shock is very fresh — adaptation path is still unclear.
risk uncertainty adaptation
Recommended Next Actions
What to Do Now
  1. Instagram Exploit slide: Step 1 (fake location) vs Step 2 (blank) — use as a warning story.
  2. "Token Maxing" slide with Meta leaderboard titles — audit your own metrics.
  3. "Quality Crisis" with 3 examples: Anthropic, Amazon, GitHub — share with your team.
  4. Kent Beck quote as thesis: "code faster than trust" — make it a team mantra.
  5. "Cost Shock" with $15,000 joke and Uber's $1,500/month cap — run the numbers for your org.
  6. Call to action: "Be the chief tech debt remover. Slow down to speed up."
actions leadership tech debt
Sources & Process Provenance
Evidence & Verification
Tier 1 — Primary source: video transcript, conference talk Process: Transcript auto-extracted via Obsidian Web Clipper. Verify key technical claims independently.
source transcript Craft 2026
1 of 11