Monday, November 16, 2026
HomeCrypto MiningGoogle says AI video decoder is 20–30% sooner in assessments

Google says AI video decoder is 20–30% sooner in assessments



Google says AI video decoder is 20–30% sooner in assessments

Google says its Antigravity /teamwork AI system rewrote libgav1, its AV1 video decoder, in Rust and made it 20–30% sooner than the unique C++ and hand-optimized meeting code on pattern video clips. The end result, disclosed in a November 13 engineering replace, adopted a port that originally ran round 4 occasions slower than the C++ model.

The corporate studies round 170,000 strains of Rust and 100% bit-exact parity throughout a broad conformance take a look at suite, that means the rewritten decoder matched the unique’s output in these assessments.

The efficiency reversal got here after /teamwork acquired profiling instruments and a aim of maximizing pace on pattern clips, Google says. It studies the benefit on each x86_64 and ARM.

Associated Studying

AI brokers simply accomplished a four-day cryptography take a look at with solely 20 human messages wanted – DeepMind says

Google preserves habits earlier than tuning pace

The migration separates two engineering issues: protecting a longtime program’s habits and making the alternative quick sufficient to make use of. Google’s account illustrates why a profitable translation alone doesn’t settle the efficiency query.

Discover what every examine measures

Google studies full bit-exact parity throughout a broad conformance take a look at suite.

Matching a take a look at suite solutions a distinct query from measuring pace on pattern video clips.

Google’s November 13, 2026 account. Conformance parity is proscribed to its reported suite; pace is a sample-clip comparability. {Hardware} fashions, construct settings and clip identities are usually not disclosed. These outcomes don’t set up common correctness, safety or manufacturing deployment.

For code migrations, Google describes a forthcoming normal rewrite sample that schedules every module solely after its dependencies compile, then assessments leaf modules in opposition to the unique, unmodified take a look at suite. Maintaining these checks intact provides the brokers a set goal whereas they modify the implementation. Profiling then supplies a separate suggestions loop for locating the place the brand new model spends its time.

Google locations this work inside collaborations with human consultants. Researchers body questions and outline constraints, whereas brokers discover potentialities, take a look at candidate options and examine outcomes.

Associated Studying

Robots now run AI-written code with out an AI calling the photographs, succeeding 92% of the time

The decoder work additionally predates this month’s replace. In its September 30 Argon announcement, Google described brokers optimizing an present Rust libgav1 port to run 2.7 occasions sooner than that earlier Rust model, bringing it nearer to optimized C++. November’s comparability is in opposition to C++ and meeting as a substitute. The totally different baselines forestall treating these figures as one measured development.

CryptoSlate’s current AWS Emulator protection involved agent-written software program and human-designed launch checks. Google’s decoder instance provides the problem of preserving an present implementation’s outputs whereas bettering its pace.

Associated Studying

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

The libgav1 challenge repository linked by Google nonetheless paperwork a C++17 and CMake construct, together with architecture-specific optimization switches. The November report doesn’t specify processor fashions, compiler settings, clip identities, thread counts or timing distributions. It additionally doesn’t establish a copy package deal for the Rust benchmark or announce that the rewritten decoder has entered manufacturing.

That leaves a concrete subsequent take a look at for the disclosure: a specified Rust revision, comparability construct and video corpus would let different builders look at whether or not the reported benefit holds beneath their very own situations.

RELATED ARTICLES

Most Popular

Recent Comments