Benoit Schillings: R&D after code (Category AI4SDLC)
I saw it. keynote report Benoit Schillings, VP from Google DeepMind at AI Engineer World's Fair 2026. His position is interesting to me - he is not a product IT leader implementing an agent in SDLC, but a head of R&D. His team is building the technology that Gemini will need on a month-to-year horizon. So instead of backlog, CI/CD and SLO, he discusses what the next model should learn.
The story has a funny beginning. In 2018 In the year X, his team launched the Pitchfork project about applying ML to code, but almost no one took the idea seriously. Schillings himself rejected natural language programming: there are already programming languages for this. Now the man with 45Years of experience - from assembler to Python - admits the error and uses vibe coding.
Code generation and software engineering are different tasks. According to Schillings, models already generate syntax better than humans. But real development begins when you need to change the system. 35 With millions of lines of PHP, consider the architecture, security, and implications of a solution ten years from now. The bottleneck moves into problem setting, decomposition and design verification.
The code for DeepMind is a convenient R&D laboratory: there is a lot of data, the result is checked by the compiler and tests. The next step is self-play, where the model creates, solves and validates problems, just as AlphaZero learned through playing with itself. Compute and the quality of the learning environment become a limitation.
The research roadmap from the report looks like this:
Teach the model to write secure code right away, not just find vulnerabilities after the fact. Develop planning, decomposition and transfer of ideas between fields; Change evals: the “started and answered” test is too narrow for the architecture. Go beyond the linear chain of tokens to multimodal representations; It is possible to create strict languages for models, even if the person is uncomfortable reading them.
The last point shows R&D optics particularly well: remove the human legibility restriction and shift the proof of correctness to the model and language.
Schillings predicts that the code will become almost free, its volume will explode, and after a year people will almost stop reading the result of the agent - as today rarely check the assembler after the compiler. This is a prediction, not a fact, and there is a definite difference between an agent and a compiler - the latter works according to a strict specification, and the agent - in an ambiguous business context.
Legacy, ownership, SLO, cost of inference and rollout remain outside the framework. But the finale goes to chemistry and biology: for Schillings, coding models are a testing ground for general reasoning and scientific discoveries.
I liked this talk - it's not a guide to implementing AI in IT. (I have no problem with that.)It is a map of upstream research. Conditionally. R&D asks, “What can a model discover and build?” Product IT adds, “How do you prove that the outcome is needed, safe, and manageable?”
#AI #AI4SDLC #Research #Engineering #Architecture #Evals