Orange Google-branded projector with a Rust label beside matching coastline film frames.
Image by CryptoSlate

Google says AI rewrote a video decoder and made it 20–30% faster in tests

Google reports matching conformance-test output and a 20–30% speed edge on sample clips, without announcing production deployment.

Quick Take

  1. Google reports a Rust libgav1 rewrite with matching conformance-test output and a 20–30% speed advantage on sample clips.
  2. Profiling followed an initially slower port, illustrating why preserving software behavior and improving performance require separate checks.
  3. The disclosure omits key benchmark settings and does not announce production deployment of the rewritten decoder.

Google says its Antigravity /teamwork AI system rewrote libgav1, its AV1 video decoder, in Rust and made it 20–30% faster than the original C++ and hand-optimized assembly code on sample video clips. The result, disclosed in a November 13 engineering update, followed a port that initially ran around four times slower than the C++ version.

The company reports around 170,000 lines of Rust and 100% bit-exact parity across a broad conformance test suite, meaning the rewritten decoder matched the original’s output in those tests.

The performance reversal came after /teamwork received profiling tools and a goal of maximizing speed on sample clips, Google says. It reports the advantage on both x86_64 and ARM.

Related Reading

AI agents just completed a four-day cryptography test with only 20 human messages needed – DeepMind says

Google preserves behavior before tuning speed

The migration separates two engineering problems: keeping an established program's behavior and making the replacement fast enough to use. Google's account illustrates why a successful translation alone does not settle the performance question.

Matching output and improving speed are separate checks

Google reports a Rust rewrite of libgav1. Explore the output comparison and the profiling loop behind its reported performance result.

Explore what each check measures

Google reports complete bit-exact parity across a broad conformance test suite.

Matching a test suite answers a different question from measuring speed on sample video clips.

Google's November 13, 2026 account. Conformance parity is limited to its reported suite; speed is a sample-clip comparison. Hardware models, build settings and clip identities are not disclosed. These results do not establish universal correctness, security or production deployment.

For code migrations, Google describes a forthcoming general rewrite pattern that schedules each module only after its dependencies compile, then tests leaf modules against the original, unmodified test suite. Keeping those checks intact gives the agents a fixed target while they change the implementation. Profiling then provides a separate feedback loop for finding where the new version spends its time.

Google places this work within collaborations with human experts. Researchers frame questions and define constraints, while agents explore possibilities, test candidate solutions and check results.

Related Reading

Robots now run AI-written code without an AI calling the shots, succeeding 92% of the time

The decoder work also predates this month's update. In its September 30 Argon announcement, Google described agents optimizing an existing Rust libgav1 port to run 2.7 times faster than that earlier Rust version, bringing it closer to optimized C++. November's comparison is against C++ and assembly instead. The different baselines prevent treating those figures as one measured progression.

CryptoSlate's recent AWS Emulator coverage concerned agent-written software and human-designed release checks. Google's decoder example adds the challenge of preserving an existing implementation's outputs while improving its speed.

Related Reading

Amazon says no human wrote or reviewed the code for its new cloud testing tool

The libgav1 project repository linked by Google still documents a C++17 and CMake build, including architecture-specific optimization switches. The November report does not specify processor models, compiler settings, clip identities, thread counts or timing distributions. It also does not identify a reproduction package for the Rust benchmark or announce that the rewritten decoder has entered production.

That leaves a concrete next test for the disclosure: a specified Rust revision, comparison build and video corpus would let other developers examine whether the reported advantage holds under their own conditions.

Article context

Mentioned in this article

Related Company Google An American multinational technology company