--- title: Research to Reality with Google DeepMind — Benoit... description: Benoit Schillings, VP of Technology at Google DeepMind, shares his perspective on the evolution of AI-assisted software development. He traces the history from... canonical: https://daily.dev/posts/research-to-reality-with-google-deepmind-benoit-schillings-google-deepmind-vp-of-technology-oayqkkb5f twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Research to Reality with Google DeepMind — Benoit Schillings, Google DeepMind, VP of Technology | daily.dev og:description: Benoit Schillings, VP of Technology at Google DeepMind, shares his perspective on the evolution of AI-assisted software development. He traces the history from... og:url: https://daily.dev/posts/research-to-reality-with-google-deepmind-benoit-schillings-google-deepmind-vp-of-technology-oayqkkb5f og:image: https://api.daily.dev/og/posts/OaYqkKb5F.png og:image:alt: Research to Reality with Google DeepMind — Benoit Schillings, Google DeepMind, VP of Technology og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Research to Reality with Google DeepMind — Benoit Schillings, Google DeepMind, VP of Technology **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 20 min read · 1 upvotes · 0 comments ## Summary Benoit Schillings, VP of Technology at Google DeepMind, shares his perspective on the evolution of AI-assisted software development. He traces the history from Project Pitchfork (2018) — an early effort to apply ML to code — through today's era of superhuman syntax generation. Key themes include: the exhaustion of human-generated training data on GitHub (with ~80% of new code now machine-generated), the shift to self-play for model improvement (analogous to AlphaZero), and the changing economics of code where writing is nearly free. He argues the new bottlenecks are design adequacy, security, and architectural reasoning rather than code generation itself. He also discusses the need for better benchmarks beyond HumanEval, the potential for new programming languages designed for models rather than humans, and how cheap experimentation via AI will accelerate breakthroughs in chemistry and biology. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What percentage of new code added to GitHub is machine-generated according to Google DeepMind's VP of Technology? Roughly 80% of new code added to GitHub is machine-generated, according to Google DeepMind VP of Technology Benoit Schillings. This shift means human-written code is becoming scarce as a source of fresh training data, pushing labs toward self-play techniques, similar to AlphaZero's approach in chess and go, where models generate and verify their own coding challenges to keep improving. _Developers tracking how AI is reshaping coding workflows can follow this shift on daily.dev._ ### What was the Pitchfork project at Google X and when did it start? Pitchfork was a project started at Google X in 2018 aimed at using machine learning to speed up how code is written, targeting slow processes like small edits requiring multi-day reviews. At the time, the idea was dismissed internally, and even vibe coding (writing code in English) was considered impractical, though both ideas later proved prescient as AI coding tools advanced. _Anyone weighing where AI coding tools are headed next can follow projects like this on daily.dev._ ### Why does SWE-bench fall short as an evaluation benchmark for AI coding models? SWE-bench only verifies whether a piece of code runs and produces the correct output, which captures a small part of real software engineering. A more complete evaluation would include open-ended problems, such as building the best lossless text compressor measured by combined compressed-file and source-code size, which forces models toward novel algorithmic thinking rather than simple correctness checks. _Teams evaluating coding models beyond pass/fail benchmarks can track these debates on daily.dev._ ## Similar posts on daily.dev - [Google CEO Says 75% of New Code is AI-Generated](https://daily.dev/posts/google-ceo-says-75-of-new-code-is-ai-generated-mmqxtfyja) · DevOps.com · 6 upvotes · 0 comments - [Google I/O showed how the path for AI-driven science is shifting](https://daily.dev/posts/google-i-o-showed-how-the-path-for-ai-driven-science-is-shifting-p08sxf2im) · MIT Technology Review · 0 upvotes · 0 comments - [Reflections on AI at the end of 2025](https://daily.dev/posts/reflections-on-ai-at-the-end-of-2025-ktddrfxcj) · Hacker News · 3 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-coding](https://daily.dev/tags/ai-coding), [#google-gemini](https://daily.dev/tags/google-gemini), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/research-to-reality-with-google-deepmind-benoit-schillings-google-deepmind-vp-of-technology-oayqkkb5f) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Research to Reality with Google DeepMind — Benoit Schillings, Google DeepMind, VP of Technology","url":"https://daily.dev/posts/research-to-reality-with-google-deepmind-benoit-schillings-google-deepmind-vp-of-technology-oayqkkb5f","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/research-to-reality-with-google-deepmind-benoit-schillings-google-deepmind-vp-of-technology-oayqkkb5f"},"datePublished":"2026-07-17T15:27:59.391Z","dateModified":"2026-09-13T21:41:49.237Z","description":"Benoit Schillings, VP of Technology at Google DeepMind, shares his perspective on the evolution of AI-assisted software development. He traces the history from...","image":"https://i.ytimg.com/vi/1P1hJ36rxM0/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/1P1hJ36rxM0/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/research-to-reality-with-google-deepmind-benoit-schillings-google-deepmind-vp-of-technology-oayqkkb5f","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-coding,google-gemini,reinforcement-learning","timeRequired":"PT20M","video":{"@type":"VideoObject","name":"Research to Reality with Google DeepMind — Benoit Schillings, Google DeepMind, VP of Technology","description":"Benoit Schillings, VP of Technology at Google DeepMind, shares his perspective on the evolution of AI-assisted software development. He traces the history from...","thumbnailUrl":"https://i.ytimg.com/vi/1P1hJ36rxM0/sddefault.jpg","uploadDate":"2026-07-17T15:27:59.391Z","duration":"PT20M","url":"https://api.daily.dev/r/OaYqkKb5F","embedUrl":"https://www.youtube.com/embed/1P1hJ36rxM0"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Research to Reality with Google DeepMind — Benoit Schillings, Google DeepMind, VP of Technology"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/research-to-reality-with-google-deepmind-benoit-schillings-google-deepmind-vp-of-technology-oayqkkb5f#faq","mainEntity":[{"@type":"Question","name":"What percentage of new code added to GitHub is machine-generated according to Google DeepMind's VP of Technology?","acceptedAnswer":{"@type":"Answer","text":"Roughly 80% of new code added to GitHub is machine-generated, according to Google DeepMind VP of Technology Benoit Schillings. This shift means human-written code is becoming scarce as a source of fresh training data, pushing labs toward self-play techniques, similar to AlphaZero's approach in chess and go, where models generate and verify their own coding challenges to keep improving. Developers tracking how AI is reshaping coding workflows can follow this shift on daily.dev."}},{"@type":"Question","name":"What was the Pitchfork project at Google X and when did it start?","acceptedAnswer":{"@type":"Answer","text":"Pitchfork was a project started at Google X in 2018 aimed at using machine learning to speed up how code is written, targeting slow processes like small edits requiring multi-day reviews. At the time, the idea was dismissed internally, and even vibe coding (writing code in English) was considered impractical, though both ideas later proved prescient as AI coding tools advanced. Anyone weighing where AI coding tools are headed next can follow projects like this on daily.dev."}},{"@type":"Question","name":"Why does SWE-bench fall short as an evaluation benchmark for AI coding models?","acceptedAnswer":{"@type":"Answer","text":"SWE-bench only verifies whether a piece of code runs and produces the correct output, which captures a small part of real software engineering. A more complete evaluation would include open-ended problems, such as building the best lossless text compressor measured by combined compressed-file and source-code size, which forces models toward novel algorithmic thinking rather than simple correctness checks. Teams evaluating coding models beyond pass/fail benchmarks can track these debates on daily.dev."}}]} ```