Atlas record: https://www.aitimeline.dev/en/events/evt-0279/
Research · June 7, 2023
AlphaDev
DeepMind applied reinforcement learning to searching machine instructions and found faster algorithms for sorting short sequences.
Why it matters
The result entered LLVM's libc++, an implementation of the C++ standard library: machine-found code became part of infrastructure everyone uses.
The problem was posed as a game: the agent adds one processor instruction at a time, and the payoff is correctness and speed. The sequences found are up to 70 per cent faster than the human-written ones for short sequences and about 1.7 per cent faster for sequences of over 250,000 elements (DeepMind’s post), and they look counterintuitive. The gain for long sequences is small, but sorting runs billions of times a day. The contribution was committed to LLVM’s libc++ on 8 April 2022 (commit 194d196, whose message says the functions were “generated using Reinforcement Learning”), a year before the paper appeared.
Event record
- Event date
- June 7, 2023
- Timeline date
- Event date
- Verification
- Sources gathered automatically · September 17, 2026
- Lines
- ID
- evt-0279
Nature paper of 7 June 2023.
Related events
Earlier
- Related AlphaFold 2 at CASP14 2020
A second problem outside AI where a model produced a result another field needed.
- Related AlphaZero 2017
The same self-play approach, carried over to search in the space of processor instructions.
Later
- Related GNoME and new materials 2023links to this record
A second work from the same laboratory that year whose result another science needs.
- Related AlphaGeometry 2024links to this record
The second system from the same laboratory in a year to search a space of formal steps.
- Extends AlphaEvolve 2025links to this record
The same task of searching for a better algorithm, now with a language model in place of learning from scratch.