{
  "video": {
    "id": "1P1hJ36rxM0",
    "title": "\"Software engineering is not about writing code\" — Benoit Schillings, Google DeepMind VP of Research",
    "duration": 1226,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "1P1hJ36rxM0",
    "title": "\"Software engineering is not about writing code\" — Benoit Schillings, Google DeepMind VP of Research",
    "one_liner": "Google DeepMind VP of Research Benoit Schillings argues that models have already won at generating syntax, so the remaining frontier — and the human's job — is multi-step codebase work, architecture, security-by-construction, and evals open-ended enough to force novel algorithms.",
    "summary": "Schillings traces three eras of software — machine-limited assembly, the cloud era where the real constraint was the 7-to-9-token human working memory, and the current AI frontier where writing code is free and specifying what you actually wanted is the bottleneck. He argues superhuman syntax generation is settled ('code is over'), while managing 35-million-line codebases, architecture with ten-year consequences, and cross-domain inductive reasoning are where frontier models are still moving. With human code on GitHub exhausted as training data, he points to AlphaZero-style self-play on self-generated, self-verified coding challenges as the next scaling axis, and calls for open-ended benchmarks, security-by-construction, multimodal 'visual' code reasoning, and possibly a new non-human-readable programming language that puts the burden of correctness on the model.",
    "key_points": [
      "He ran Pitchfork at Google X starting in 2018 to apply ML to how code is written; at the time 'nobody would give us the time of day' at Google, and he himself dismissed writing code in English because 'English is not a programming language' — he says he was wrong.",
      "Three eras: the machine-limited assembly era; the cloud era where compute was cheap and the real limit was the human brain's 7–9 token context versus ML context heading toward effectively infinite; and the AI frontier where writing code is no longer the challenge but specifying what is really wanted is.",
      "Superhuman syntax generation is done — he can no longer look at a Gemini-written function and do better. What remains hard is multi-step codebase work: 'software engineering is the first time you join a company and you realize that there are 35 million lines of PHP in the codebase', plus Jeff-Dean-level architecture spanning hardware optimization, security, and ten-year regret avoidance.",
      "Code was a uniquely good ML domain for two reasons: enormous scrapeable training data on GitHub, and cheap verification (compile it, run it, unit test it). That first advantage is ending — he claims 80% of new code added to GitHub today is machine generated, so human-knowledge mining is reaching an end.",
      "Self-play is the replacement: like AlphaZero reaching superhuman Go and chess with no human knowledge, frontier code models can now create their own challenges, judge the validity of answers, and to some extent judge architecture. Hundreds of millions of hours of self-play coding is the next layer; the horizon is set by how much compute and self-play time you have.",
      "On security: the wave of model-found vulnerabilities will be never-ending — each fix layer makes the next, subtler layer findable. His team is working on the grail of teaching models to write correct code from the start rather than detect-then-patch, which is very hard because correctness is so context dependent.",
      "Benchmarks that only check that code runs and produces the right output are 'infamous' in his book. He wants open-ended problems instead — his example: give a model 10 megabytes of text/code, ask for the best lossless compressor, and score it on compressed size plus source size. It never ends, and it forces genuinely new algorithms.",
      "Two structural bets: code reasoning shouldn't be only a chain of tokens (code writing is visual — block diagrams, data flow — hence Gemini's multimodal-from-the-onset choice and spatial/dynamic representations becoming a must-have), and it may be time to design a language for models — strongly typed, Lean-inspired, correctness pushed onto the model, and not necessarily human readable.",
      "Beyond code: because experimentation becomes nearly free, the code-meets-atoms intersection is where true novelty appears — chemistry (humans understand a sliver; past 20 atoms in a molecule we don't know what it does, and 10,000 atoms starts to look like life), biology ('nature did an incredible engineering job and terrible job at documentation'), and 'the gold we cannot see', breakthroughs invisible to brains trained by evolution for the jungle rather than quantum computing."
    ],
    "takeaways": [
      "Stop optimizing for code-writing throughput and move your effort upstream to architecture, problem decomposition and specification — writing the code is free or nearly free, so design adequacy and reliability at the microscopic level become the scarce human contribution.",
      "Plan processes for a world where generated code is not read. He predicts that within a year Gemini or similar will generate code nobody looks at, the way nobody checks compiler assembly output — so invest in active guard rails and automated correctness rather than human code review as the safety net.",
      "Treat code security as at least as much work as code generation, and prefer approaches that make code correct by construction over detect-and-patch cycles, which he expects to be never-ending.",
      "Build open-ended evals with unbounded loss functions (his text-compression example: minimize compressed size plus source size) instead of pass/fail 'does it run and produce the right output' benchmarks, which capture only a small part of code engineering.",
      "If you have verification you trust, generate your own challenges and self-play against them — the same move a brilliant engineer locked in a room for two years would make, and the one that replaces exhausted human training data."
    ],
    "topics": [
      "code generation",
      "self-play",
      "evals",
      "code security",
      "software engineering",
      "multimodal",
      "programming languages",
      "ai for science"
    ],
    "tools": [
      "Gemini",
      "Google DeepMind",
      "Google X",
      "Pitchfork",
      "Waymo",
      "Google Glass",
      "GitHub",
      "AlphaZero",
      "Python",
      "C++",
      "Lean",
      "Apple II",
      "Commodore 64"
    ],
    "quotes": [
      {
        "text": "Software engineering is the first time you join a company and you realize that there are 35 million lines of PHP in the codebase and that you need to make some changes. That that's the day you understand what software engineering is.",
        "at": "07:42",
        "url": "https://www.youtube.com/watch?v=1P1hJ36rxM0&t=462s"
      },
      {
        "text": "I think that 80% of the new code added to GitHub today is machine generated.",
        "at": "09:42",
        "url": "https://www.youtube.com/watch?v=1P1hJ36rxM0&t=582s"
      },
      {
        "text": "I would predict that in one year we'll let Gemini or other model generate the code and nobody will actually look at it.",
        "at": "12:09",
        "url": "https://www.youtube.com/watch?v=1P1hJ36rxM0&t=729s"
      },
      {
        "text": "We're getting to the point where since the pain of writing the code does not exist anymore. How about we make writing the code much harder by having you know very strongly typed languages",
        "at": "17:12",
        "url": "https://www.youtube.com/watch?v=1P1hJ36rxM0&t=1032s"
      }
    ],
    "words": 3658
  },
  "summary_url": "/#1P1hJ36rxM0",
  "transcript": {
    "html": "/transcripts/1P1hJ36rxM0.html",
    "txt": "/transcripts/1P1hJ36rxM0.txt",
    "vtt": "/transcripts/1P1hJ36rxM0.vtt"
  }
}