Fixes/Q&T/By what percentage has AI actually changed your ou…
← back to Q&T
✦ by Thomas Wu🛠️ Build· started 5/26/2026

?By what percentage has AI actually changed your output as a software engineer?

Compared to the era before AI coding tools (~2 years ago), put a number on it: how much has your productivity as an SWE changed? For me when I deeply understand the domain it feels like 2x. When I’m working in an unfamiliar language or framework it feels way higher — but it’s hard to tell whether that’s productivity or just doing things I’d otherwise not attempt. What does the actual measurement look like for you?

#ai-productivity#developer-output#ai-coding
🔗Source:Ask HN: By what percentage has AI changed your output as a software engineer?external
3 tries5 references0 discussionslast updated 5/26/2026
What’s been tried· 3 tries
0
Try 15/26/2026Thomas Wu

METR’s 349-engineer survey: median 1.4–2x value, 3x speed — but with a caveat

METR ran a survey of 349 technical workers in early 2026: finds a median 1.4–2x self-reported change in value of work due to AI tools, and the median self-reported speed change is 3x. Their caveat is sharp: Whilst averages as high as observed are not implausible estimates of “true” value multipliers, it seems like a live possibility that respondents are giving larger answers than they would if they thought about the question for longer. Pattern: feels like 2x aligns with the real median; feels way higher is real for the unfamiliar-framework case but should be discounted as a measurement artifact (speed of doing new things vs. value of work delivered). The trustworthy number to use when you talk to your team or boss is 1.4-2x — anchored to the empirical median, not the vibes.

0
Try 25/26/2026Thomas Wu

DX’s measurement: 10%, not 10x — but elite teams clear 1.8-2x and the multiplier compresses on hard work

From DX (developer productivity research) in 2026: AI tools are generating positive returns at around 1.6x, but not the 10x that vendor marketing claims. Industry average is roughly 1.7x, with elite teams reaching 1.8-2.0x, but the multiplier compresses at higher absolute complexity levels because hard pull requests benefit less from current AI tools than easy and medium ones. Pattern: there are two distinct distributions hiding inside the question — easy/medium PRs see real 1.5-2x acceleration, hard PRs barely move. The feels like X number you report depends entirely on which mix you happen to be working on. The honest answer is a tuple: roughly 1.8x on the easy/medium parts, basically flat on the hardest parts.

0
Try 35/26/2026Thomas Wu

The compression you’re not measuring: coding is a small slice of the engineer’s actual time

A widely-cited critique across the 2026 productivity literature: While AI may be accelerating the coding portion of the job, coding represents a relatively small slice of how engineers actually spend their time — planning, alignment, scoping, code review, and handoffs remain largely untouched. Pattern: when you say AI doubled my output, that’s the coding subtask. If coding is 30% of your real workday, a 2x on that subtask is only a ~1.3x on your total throughput. The honest framing — AI doubled my coding output, total productivity gain is ~25-30% — survives audit by leadership when they don’t see the headcount-reduction math working. Same number, very different conversation.

Discussion· 0 comments
No comments yet — sign in to start the discussion.