All news

An 11-hour unattended Gemini 4 Argon run produced two accepted optimizations for Google ADK 200K-line Python codebase, cutting memory and latency.

Shard keys should come from your access patterns, not your tables: co-locating orders with customers by customer_id avoids cross-shard joins in sharded Postgres.

Evolving the interfaces that computer-use agents see makes them up to 3.8x better.

Cloudflare rolled out WAF rules for the php2shell PHP RCE flaws and notes a related session upload bug, CVE-2026-115560, that affects fully patched PHP.

Meta builds a stable token vocabulary for ads so its generative recommenders can reason over ad content without ID churn.

StackGres benchmarks open-source Citus at 10M QPS on half the data nodes PlanetScale used for its sharded Postgres Neki, arguing Citus is the more efficient option at that scale.

Analysis of why php2shell is not yet mass-exploited despite the panic.

Pinterest releases ModuleLens, a tool to inspect PyTorch module structure and cost in large recommendation models, for engineers and agents.

depesz benchmarks audit-log triggers and finds hstore can beat jsonb when logging only changed columns, with numbers on log size and trigger cost.

Autonomous AI research teams discover linear attention architectures.

A dermatologist found a fine-tuned 2015 Inception v3 CNN beat two frontier LLMs at diagnosing 200 biopsy-proven skin lesions, with the LLMs missing a melanoma.

The node2shell Brotli bug CVE-2026-111497 also affects Python and other Brotli consumers, not just Node.

Private keys for eight widely trusted eMudhra intermediate certificates were compromised on October 19, and Cloudflare explains its response and what users should do.

Tests which LLM inference optimizations designed for NVIDIA GPUs actually help on Apple Silicon.

PHP shares mitigations for a new unpatched RCE in non-default configurations.

Part 1 of a guide to system-level profiling of PyTorch inference on AWS Trainium.

Netflix built Quasar, a tool that maps a proposed Java library API change to real call sites across consumer repos, so library owners can see the blast radius before releasing.

FlashMap uses flow maps to improve the speed and acceptance trade-off in speculative decoding.

Specializing the tokenizer vocabulary of a fine-tuned small model cut token counts and inference cost by 77% in a production essay grader.

CSIRO Serverless Beacon queries 100x more genomic variants on AWS without loading VCF files into a central database.

Amazon says PHP2Shell exploitation began four days before patches.

Re-selects KV cache visual tokens during long multimodal generation based on drift, under a budget.

Shows how to use cheap decision models as eval classifiers: tune the decision threshold instead of the default 0.50, compare cost and accuracy across models on 100 agent traces, and calibrate scores.

Characterizes how LLMs keep repeating instructions while ceasing to obey them in long dialogues.

A FAST feature detector using the 128-bit SIMD unit of the ESP32-S3 runs about 2x faster, about 6 ms per QVGA frame.

Interactive animations that show how metastable failures and lost writes happen in distributed systems.

Spherical lattice vector quantization-aware training for low-bit LLMs.

Netflix added VM-level single-tenant isolation to Titus, its Kubernetes-based container platform, for workloads that cannot share hosts.

Cerebras explains how it deploys trillion-parameter models across wafer-scale chips that each hold only 44 GB of on-chip SRAM.

Quant-Arena benchmarks LLM agents on quantitative trading research.