--- schema: 1 type: video title: 'AI-Generated Code Is Already Competing With Human Code — Daksh Gupta, Greptile' description: >- More than a quarter of the pull requests Greptile reviewed in April showed signs of being written largely or entirely by AI agents, up from under 1% a year earlier. Daksh Gupta, co-founder of Greptile, analyzed more than a million pull requests a month from companies like NVIDIA, Coinbase and Scale to test a simple question: are fully vibe-coded PRs actually any good in enterprise codebases? He compares agent-written PRs (Codex, Claude Code, Devin, Cursor) with human-written ones on four measures: revert rates, revert rates by PR size, the severity of issues Greptile flags (P0/P1/P2), and how many review rounds it takes to get to a mergeable PR. On all four, agent code landed in the same range as human code, and humans were actually more likely to introduce P0 bugs. The differences show up in *how* each one fails: Claude is about 1.5x more likely than humans to introduce SQL injection, Devin is half as likely to cause auth bypasses, and N+1 queries are far more common from Cursor. Daksh also shares what this means for code review. The median Greptile user makes fewer than 50 commits a month, while the 99th percentile makes close to 1,000, which makes manual review impossible at that scale. He argues that code validation should answer three questions: Does this change violate the user contract? Does it make a future violation more likely? Does it do what the author intended? He then shows how Greptile answers them with blast-radius analysis and sandboxed browser agents that try to break the app. Speaker info: - Greptile: https://greptile.com - Daksh Gupta on X: https://x.com/dakshgup - Daksh Gupta on LinkedIn: https://www.linkedin.com/in/dakshg/ - Daksh's website: https://dakshgupta.com Timestamps: 00:00 Intro 00:25 Daksh Gupta & Greptile 00:57 From tab complete to fully autonomous agents 02:09 Are fully vibe-coded PRs any good? 02:29 How to detect vibe-coded PRs: author fields, co-author footers, branch prefixes 03:49 25%+ of PRs vibe coded, up from under 1% a year ago 04:18 What makes a pull request "good"? 04:38 Revert rates: Codex vs. Devin vs. humans 05:28 Do revert rates change with PR size? 06:08 Severity of issues found: P0s, P1s and P2s 07:02 Review rounds to reach a mergeable PR 07:47 Different failure modes: SQL injection, auth bypass, N+1 queries 09:07 Autonomous agents work in enterprise settings 10:01 What this means for code review: 50 vs. 1,000 commits a month 10:51 Rethinking code validation: three questions 11:16 How Greptile validates: blast radius + sandboxed browser agents 11:56 Closing --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/474j-n1Ltxc-youtube-0771ba030cee4a5fb3ba Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: 'https://liquidgold.cc/watch/474j-n1Ltxc' markdown: 'https://liquidgold.cc/watch/474j-n1Ltxc.md' llms: 'https://liquidgold.cc/watch/474j-n1Ltxc/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:20:19.700Z' language: null children: [] videoId: 474j-n1Ltxc durationSeconds: 761 publishedAt: '2026-09-27T18:30:07.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 25 endSeconds: 35 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=25' heldReason: rights - ordinal: 1 text: null startSeconds: 35 endSeconds: 57 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=35' heldReason: rights - ordinal: 2 text: null startSeconds: 68 endSeconds: 134 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=68' heldReason: rights - ordinal: 3 text: null startSeconds: 68 endSeconds: 134 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=68' heldReason: rights - ordinal: 4 text: null startSeconds: 134 endSeconds: 145 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=134' heldReason: rights - ordinal: 5 text: null startSeconds: 169 endSeconds: 238 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=169' heldReason: rights - ordinal: 6 text: null startSeconds: 238 endSeconds: 248 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=238' heldReason: rights - ordinal: 7 text: null startSeconds: 248 endSeconds: 259 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=248' heldReason: rights - ordinal: 8 text: null startSeconds: 270 endSeconds: 283 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=270' heldReason: rights - ordinal: 9 text: null startSeconds: 283 endSeconds: 294 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=283' heldReason: rights - ordinal: 10 text: null startSeconds: 294 endSeconds: 336 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=294' heldReason: rights - ordinal: 11 text: null startSeconds: 348 endSeconds: 372 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=348' heldReason: rights - ordinal: 12 text: null startSeconds: 383 endSeconds: 427 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=383' heldReason: rights - ordinal: 13 text: null startSeconds: 427 endSeconds: 437 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=427' heldReason: rights - ordinal: 14 text: null startSeconds: 474 endSeconds: 486 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=474' heldReason: rights - ordinal: 15 text: null startSeconds: 486 endSeconds: 552 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=486' heldReason: rights - ordinal: 16 text: null startSeconds: 552 endSeconds: 563 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=552' heldReason: rights - ordinal: 17 text: null startSeconds: 563 endSeconds: 608 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=563' heldReason: rights - ordinal: 18 text: null startSeconds: 608 endSeconds: 619 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=608' heldReason: rights - ordinal: 19 text: null startSeconds: 652 endSeconds: 662 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=652' heldReason: rights - ordinal: 20 text: null startSeconds: 662 endSeconds: 684 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=662' heldReason: rights - ordinal: 21 text: null startSeconds: 696 endSeconds: 718 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=696' heldReason: rights - ordinal: 22 text: null startSeconds: 718 endSeconds: 761 watch: 'https://liquidgold.cc/watch/474j-n1Ltxc?t=718' heldReason: rights --- # AI-Generated Code Is Already Competing With Human Code — Daksh Gupta, Greptile More than a quarter of the pull requests Greptile reviewed in April showed signs of being written largely or entirely by AI agents, up from under 1% a year earlier. Daksh Gupta, co-founder of Greptile, analyzed more than a million pull requests a month from companies like NVIDIA, Coinbase and Scale to test a simple question: are fully vibe-coded PRs actually any good in enterprise codebases? He compares agent-written PRs (Codex, Claude Code, Devin, Cursor) with human-written ones on four measures: revert rates, revert rates by PR size, the severity of issues Greptile flags (P0/P1/P2), and how many review rounds it takes to get to a mergeable PR. On all four, agent code landed in the same range as human code, and humans were actually more likely to introduce P0 bugs. The differences show up in \*how\* each one fails: Claude is about 1.5x more likely than humans to introduce SQL injection, Devin is half as likely to cause auth bypasses, and N+1 queries are far more common from Cursor. Daksh also shares what this means for code review. The median Greptile user makes fewer than 50 commits a month, while the 99th percentile makes close to 1,000, which makes manual review impossible at that scale. He argues that code validation should answer three questions: Does this change violate the user contract? Does it make a future violation more likely? Does it do what the author intended? He then shows how Greptile answers them with blast-radius analysis and sandboxed browser agents that try to break the app. Speaker info: - Greptile: https://greptile.com - Daksh Gupta on X: https://x.com/dakshgup - Daksh Gupta on LinkedIn: https://www.linkedin.com/in/dakshg/ - Daksh's website: https://dakshgupta.com Timestamps: 00:00 Intro 00:25 Daksh Gupta & Greptile 00:57 From tab complete to fully autonomous agents 02:09 Are fully vibe-coded PRs any good? 02:29 How to detect vibe-coded PRs: author fields, co-author footers, branch prefixes 03:49 25%+ of PRs vibe coded, up from under 1% a year ago 04:18 What makes a pull request "good"? 04:38 Revert rates: Codex vs. Devin vs. humans 05:28 Do revert rates change with PR size? 06:08 Severity of issues found: P0s, P1s and P2s 07:02 Review rounds to reach a mergeable PR 07:47 Different failure modes: SQL injection, auth bypass, N+1 queries 09:07 Autonomous agents work in enterprise settings 10:01 What this means for code review: 50 vs. 1,000 commits a month 10:51 Rethinking code validation: three questions 11:16 How Greptile validates: blast radius + sandboxed browser agents 11:56 Closing --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/474j-n1Ltxc-youtube-0771ba030cee4a5fb3ba Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/watch/474j-n1Ltxc · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:12 to 0:57) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (0:57 to 2:14) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (2:14 to 2:49) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (2:49 to 4:19) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (4:19 to 4:54) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (4:54 to 6:12) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (6:12 to 7:07) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (7:07 to 8:06) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (8:06 to 9:12) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (9:12 to 10:08) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (10:08 to 10:52) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (10:52 to 11:24) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (11:24 to 12:41) > Held: transcript rights do not permit this chunk's text.