François Chollet@fchollet
Code volume does not represent productivity.
Jen Zhu@jenzhuscott
Massive output uptick due to agentic AI. Complete flat adoption.
8:56 AM · Jun 6, 2026 · 61.2K Views
MIT study finds developer AI tools boost file creation by 300% but yield only a 30% increase in completed software releases
Flat App Store engagement shows AI-generated apps lack traction.
2101K119282213.6K
Sentiment
Positive users endorse the MIT study showing AI boosts code volume far more than actual releases or adoption, while negative users dismiss the outputs as worthless slop and reject volume metrics.
Pos
52.4%
Neg
47.6%
11 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS72.9KBOOKMARKS111RETWEETS31
Rohan Paul@rohanpaul_ai
New MIT study. Code volume surges by 300%, but output increases by only 30%: The AI dividend meets an awkward reality
Autonomous AI coding agents raised commits by 180%, but releases rose only 30%.
The paper’s main idea is that software production has weak links, so faster code writing does not help as much when humans still need to review, connect, test, package, and ship the work.
The authors also check app marketplaces and find more new apps, but no increase in total usage, which means more software appeared without clear evidence that users adopted more software.
The marketplace evidence points the same way: more new apps appeared, but total usage did not rise.
The authors compare more than 100,000 GitHub developers before and after they start using 3 generations of AI coding tools, from autocomplete to more independent coding agents.
Autocomplete raised commits by 40%, interactive coding agents raised them by 140%, and autonomous coding agents raised them by 180%.
The 180% commit gain shrank to 50% for the number of projects and 30% for actual releases.
The estimated "elasticity of substitution" is 0.25 i.e. for every big improvement in AI’s usefulness, only a small amount of human work can be replaced.
Because AI can write code faster, but humans are still needed to decide what to build, check if the code works, connect it with the rest of the product, fix messy edge cases, and actually ship it.
---
papers .ssrn.com/sol3/papers.cfm?abstract_id=6859839
Rohan Paul@rohanpaul_ai
FT publisehd a piece. AI is raising software supply faster than demand.
AI is producing far more work inside companies, but the new evidence says much of that extra motion is getting lost before it becomes shipped product or customer demand.
Last week's MIT study tracked software teams across the full production funnel, from files edited to reviewed work to software releases, rather than treating code volume as value.
AI helped developers create or edit nearly 300% more files, but the gain fell to 150% at review and only about 30% at release.
The gap means AI is strongest at speeding local tasks, while human review, coordination, product judgment, testing, and launch processes still decide how much value survives.
---
ft .com/content/8e9ae7a4-7209-4e2c-aa36-f3af77d6ce1f?syn-25a6b1a6=1
10hViews 72.9KLikes 134Bookmarks 111
LIKES191REPLIES56
Chamath Palihapitiya@chamath
Using AI to code, without a clear intent upfront, is just AI slop waiting to happen.
Rohan Paul@rohanpaul_ai
New MIT study. Code volume surges by 300%, but output increases by only 30%: The AI dividend meets an awkward reality
Autonomous AI coding agents raised commits by 180%, but releases rose only 30%.
The paper’s main idea is that software production has weak links, so faster code writing does not help as much when humans still need to review, connect, test, package, and ship the work.
The authors also check app marketplaces and find more new apps, but no increase in total usage, which means more software appeared without clear evidence that users adopted more software.
The marketplace evidence points the same way: more new apps appeared, but total usage did not rise.
The authors compare more than 100,000 GitHub developers before and after they start using 3 generations of AI coding tools, from autocomplete to more independent coding agents.
Autocomplete raised commits by 40%, interactive coding agents raised them by 140%, and autonomous coding agents raised them by 180%.
The 180% commit gain shrank to 50% for the number of projects and 30% for actual releases.
The estimated "elasticity of substitution" is 0.25 i.e. for every big improvement in AI’s usefulness, only a small amount of human work can be replaced.
Because AI can write code faster, but humans are still needed to decide what to build, check if the code works, connect it with the rest of the product, fix messy edge cases, and actually ship it.
---
papers .ssrn.com/sol3/papers.cfm?abstract_id=6859839
4hViews 57.4KLikes 191Bookmarks 70
Rohan Paul@rohanpaul_ai
FT publisehd a piece. AI is raising software supply faster than demand.
AI is producing far more work inside companies, but the new evidence says much of that extra motion is getting lost before it becomes shipped product or customer demand.
Last week's MIT study tracked software teams across the full production funnel, from files edited to reviewed work to software releases, rather than treating code volume as value.
AI helped developers create or edit nearly 300% more files, but the gain fell to 150% at review and only about 30% at release.
The gap means AI is strongest at speeding local tasks, while human review, coordination, product judgment, testing, and launch processes still decide how much value survives.
---
ft .com/content/8e9ae7a4-7209-4e2c-aa36-f3af77d6ce1f?syn-25a6b1a6=1
11hViews 17.9KLikes 24Bookmarks 19
Rohan Paul@rohanpaul_ai
Rohan Paul@rohanpaul_ai
New MIT study. Code volume surges by 300%, but output increases by only 30%: The AI dividend meets an awkward reality
Autonomous AI coding agents raised commits by 180%, but releases rose only 30%.
The paper’s main idea is that software production has weak links, so faster code writing does not help as much when humans still need to review, connect, test, package, and ship the work.
The authors also check app marketplaces and find more new apps, but no increase in total usage, which means more software appeared without clear evidence that users adopted more software.
The marketplace evidence points the same way: more new apps appeared, but total usage did not rise.
The authors compare more than 100,000 GitHub developers before and after they start using 3 generations of AI coding tools, from autocomplete to more independent coding agents.
Autocomplete raised commits by 40%, interactive coding agents raised them by 140%, and autonomous coding agents raised them by 180%.
The 180% commit gain shrank to 50% for the number of projects and 30% for actual releases.
The estimated "elasticity of substitution" is 0.25 i.e. for every big improvement in AI’s usefulness, only a small amount of human work can be replaced.
Because AI can write code faster, but humans are still needed to decide what to build, check if the code works, connect it with the rest of the product, fix messy edge cases, and actually ship it.
---
papers .ssrn.com/sol3/papers.cfm?abstract_id=6859839
9hViews 1.9KLikes 2Bookmarks 2
Dror Poleg@drorpoleg
@fchollet It does, but not in the way we normally expect. TikTok produces the next hit by producing 99% that no one watches. The same dynamic now comes to everything.
1dViews 702Likes 1Bookmarks 1
JJ@JosephJacks_
@fchollet Oh man, that this even has to be said. 🥴
1dViews 1.1KLikes 3Bookmarks 1
eran shir@eranshir
@chamath Intent is an important first step. Next one needs a comprehensive planning system and a deterministic harness.
4hViews 47Bookmarks 1
Sneha@Sneha_TechAI
@rohanpaul_ai The 'weak-link' hypothesis here is key. We’ve effectively replaced the 'blank page problem' with a 'code maintenance/validation problem.' AI is brilliant at generating syntax, but software engineering is still 80% decision-making and 20% typing. Efficiency \neq Effective output.
3hViews 39Bookmarks 1
KhalAI@ai_khal
@fchollet Doesn't this graph show the opposite? The fact that there are now fewer apps with enough reviews means people are splitting their time across more apps?
1dViews 157Likes 1
Rohit@rohitrango
@fchollet because people just built their own apps? I dont even check the app store any more. it takes one weekend afternoon to give a product spec and have the app ready in 30m
1dViews 132Likes 1
Lunari@0x_lun
@fchollet app releases up 80% and reviews flatlined lmao
someone shipped a lot of apps nobody opened
1dViews 134
tsunami_crypto@ls_brd
@fchollet that graph is basically the programmer equivalent of typing louder
quality still goes to zero just faster now
1dViews 134
James Blackwell@jwblackwell
@fchollet This seems more about demand than developer productivity
1dViews 98
haro@harobuilds
@fchollet releases up 80 points. reviews flat. usage flat. that chart is a graveyard dressed up as a growth story
1dViews 232Likes 3
Winston B.@DoDataThings
@drorpoleg @fchollet Volume as selection pressure works when you can actually evaluate the output cheaply. TikTok gets real engagement signals within 48 hours. Code review doesn't scale the same way, so the 99% doesn't get culled, it ships.
1dViews 51
God, King & Country@rwpopulist
@drorpoleg @fchollet Great, we're just reducing everything to "a thousand monkeys at a thousand typewriters".
What a mindless and nihilistic society to reduce intelligence to this.
1dViews 35
zoidz00@zoidz00
@fchollet @MLStreetTalk Old pre-AI productivity metric: how many LOC you can output
New post-AI productivity metric: how many LOC you can refactor and reduce in your codebase
1dViews 26
God, King & Country@rwpopulist
@fchollet It's almost always inverse to productivity.
Naive novice coders receive a task and furiously churn out hundreds of lines of code to solve it (massive maintenance burden, doesn't integrate with existing code).
Seniors meditate on it until they find a 5 line solution.
1dViews 21
Morgan@morganlinton
@fchollet 🎯
1dViews 362Likes 1
Guilherme O'Tina@guilhermeotina
@fchollet yeah the chart makes it concrete. agentic apps are being released but nobody opens them. generating 17x more code just shifts the bottleneck from writing to distribution. the taste/persona fit problem is still wide open
1dViews 360Likes 1
Load more posts