The week in AI & systems

Top of the week

1. ArchitectureCase studyWed, Oct 14

   Anthropic replaced Git with Scion as code merges rose to 30,000 per day

   Anthropic says its monorepo grew from about 2,500 merged changes per day in May to over 30,000 in September, prompting a five-month migration from Git to Scion and new review and source-control systems. The account shows how agent-driven code volume can force a redesign of developer infrastructure.

2. AgentsCase studySat, Oct 17

   Claude Code doubled merged pull requests while cutting revert share by over 40%

   From June to September, Claude Code's merged pull requests more than doubled while the share reverted fell by over 40%; the team shifted much review to machines and emphasized fast reversals. These are practical production lessons for reviewing code at agent-generated volume.

3. AgentsDeep diveWed, Oct 14

   XBOW chained three Chrome bugs to escape the renderer sandbox in six days

   XBOW combined V8, animated-WebP and Crashpad bugs into a renderer-sandbox escape, building the chain in six days through parallel campaigns. The case shows how agent-assisted security work can compress exploit development and stresses the need to evaluate browser defenses against chained attacks.

4. InferenceDeep diveTue, Oct 13

   ReplaySSM speeds hybrid-model decoding with checkpointed state and input buffers

   ReplaySSM stores a checkpoint plus recent inputs, cutting state-memory traffic roughly in half and speeding decoding kernels 1.46–1.80x; speculative throughput rises 1.85–2.10x. The design targets memory traffic, rollback, and precision trade-offs in hybrid-model serving.

5. InferencePaperWed, Oct 14

   DoSA skips unnecessary sparse-attention scores with exact selection guarantees

   DoSA uses score bounds to avoid fully scoring candidates that cannot enter Top-K, reporting up to 2.74× Indexer speedup and up to 1.94× time-to-first-token speedup. Its exact-selection approach offers a concrete path to faster sparse-attention serving.

6. ArchitectureCase studyWed, Oct 14

   Argon separated Chrome webpage handling from a six-million-line legacy folder

   Argon helped the Chrome team separate core webpage handling from a legacy folder of browser features across hundreds of files, finishing months of refactoring work in weeks. It illustrates how coding agents can change the pace and scale of architecture work in a large codebase.

7. SecurityDeep diveFri, Oct 16

   Citizen Lab documents targeted ads used to deliver zero-click spyware

   Citizen Lab reports the first public technical confirmation of a zero-click mercenary spyware attack delivered through targeted ads, including infections on Al Jazeera devices and a billion-bid-scale targeting operation. The findings expose risks at the intersection of ad systems, browser delivery and surveillance.

8. InferenceDeep diveSat, Oct 17

   Muse Spark Open reaches 36 tokens per second across four Mac Studios with MLX

   Apple reports over 36 tokens per second for the nearly trillion-parameter Muse Spark Open on four Mac Studios, using MLX and RDMA over Thunderbolt at an 8,192-token context. The setup offers a concrete reference for distributed inference on unified-memory hardware.

9. ArchitectureBenchmarkTue, Oct 13

   Postgres storage bottlenecks appear in a 3.3-billion-row workload

   A benchmark compares the same 3.3-billion-row PostgreSQL workload on baseline gp3 EBS and local NVMe, connecting storage performance to ingestion, latency, vacuum, checkpoint and replication problems. The proposed production patterns can inform how teams scale large database-backed systems.

10. Models & trainingPaperTue, Oct 13

    Researchers hide encrypted neural networks inside host model weights

    The authors describe backdoors that decrypt and run hidden neural networks only after a secret trigger, and report testing elicitation methods including fine-tuning and gradient-based trigger search. The result challenges the assumption that capability evaluations can reliably expose hidden model behaviors.

Also worth your time