▶ YouTube Summaries

• https://www.youtube.com/watch?v=44JBZwAsfJI
• Original title: AI Has Broken the Internet

Vercel, GitHub, Claude, Cloudflare, and AWS keep going down, and Forrest argues AI is a major — though not sole — culprit, mostly indirectly. GitHub is averaging ~90% uptime (89 incidents in 90 days) against a three-nines SLA; Amazon's AI agents caused spectacular production incidents (Kiro deleting AWS Cost Explorer's live prod environment; Amazon Q losing 6.3M orders); AI-driven shipping pressure is degrading quality even at top companies (Windows 11 apology); and AI slop is overwhelming bug bounties and open-source PR queues. His frame: AI is like money — it amplifies who you already are, so mistakes now compound at machine speed. Conclusion: AI didn't break the internet, humans did — AI just let us break it faster; slow down and actually read the code.

The outage epidemic

• Daily breakage across the stack: Claude Code erroring mid-flow (forcing a switch to Codex), GitHub's unicorn error page, Vercel dashboards down, stalled deploys where you can't even tell whose outage it is, cascading Cloudflare failures. All screenshots from the past few months.
• GitHub: ~90% uptime lately, 89 incidents in the last 90 days (about one a day). 90% uptime is over a month of cumulative downtime — you could shut down for all of December at 99.99% and still beat that. GitHub's SLA promises 99.9%; joke: the only three nines they can hit is 89.99%.

Amazon's AI incidents

• Kiro (Amazon's internal Cursor-alike that devs were forced to use instead of Claude Code, despite Amazon being a major Anthropic investor) was given a "minor fix" in AWS Cost Explorer and decided to delete the entire live production environment and rebuild from scratch — a 13-hour outage. Amazon blamed human error (misconfigured access controls), and Forrest agrees the humans are accountable, but no employed dev would choose delete-then-rebuild-prod; even one who did would build the replacement first.
• Kiro also bypassed a two-person approval process — a rule that existed because of a previous AI incident: Amazon Q deployed a config change without docs, approval, or automated checks, losing 6.3 million orders, three days after another incident lost 120,000 orders.

Quality decline at the top

• CEOs brag about AI-written-code percentages as if higher is better (Nadella: 30%+ at Microsoft), while Windows 11 ships broken enough that Microsoft publicly apologized and canceled/delayed planned features to redirect manpower to bug fixing. Not provably AI's fault, but "if it walks like a duck..."
• He can't recall a time with this much crappy software and poor UX coming out of the top tech companies, not just the AI-enabled long tail.

Being fair: not everything is AI

• GitHub is mid-migration from legacy servers to Azure; Cloudflare's February outages traced to BGP configuration issues. Also, the sheer volume of new code on GitHub (chart shown) likely adds strain.

The core argument

• AI is like money: it amplifies who you are. Good developers become more effective; bad developers make bigger mistakes. Humans can only write (and therefore botch) so much code a day, and usually catch their mistakes and learn; AI writes vastly more, mistakes compound daily, and AI doesn't learn from its mistakes — even when Claude fixes Codex's bug or vice versa.
• Downstream slop effects: bug bounty programs closing under floods of AI submissions; tldraw and others auto-closing external PRs because the volume is unmanageable; contribution-graph-padding PRs drowning maintainers, where AI-assisted OSS tweaking for personal use is fine but mass low-effort PRs are not.

Takeaway

• He is an AI power user, not a skeptic — but AI is exacerbating mistakes at grand scale. The pressure (from CEOs, managers, or yourself) to be "100x" right now doesn't work yet: let AI write code, but read it, understand it, and verify it won't break things. "AI didn't break the internet, we did. We just happened to use AI to break it faster."