§ Infrastructure

Ruff v0.16.0 enables 413 default linting rules (up from 59), breaking existing CI pipelines and catching syntax/runtime errors previously uncaught.

OpenForgeRL enables end-to-end training of harness-native agents with open infrastructure, addressing limitation of complex inference harnesses like Claude Code.

Windowed-MTP optimizes speculative decoding at million-token context by eliminating full-KV attention overhead in multi-token prediction draft heads.

Agentic context management frames token cost and memory bloat as lifecycle and architecture problems, not storage-retrieval, for production agent reliability.

Randomized KV-cache eviction with error certification via Hájek correction, proving deterministic eviction hides information loss.

TTEL: inference-time algorithm using token-level error localization and environment feedback for efficient test-time scaling.

KroQuant: Kronecker-structured block transforms for W4A4 post-training quantization of diffusion transformers with efficient inference.

Euclid-MCP: open-source MCP server coupling LLMs with SWI-Prolog for deterministic logical reasoning in safety-critical domains.

MemTools: interoperability framework decoupling memory system components for standardized agent architecture research.

Information-theoretic analysis measures LoRA adapters store ~2 bits per parameter, less than full fine-tuning, with capacity decoupled from parameter count.

Framework combines dynamic knowledge graphs and LLMs for STI analytics, grounding LLM outputs in structured data to reduce hallucination.

Formalizes cryptographically verifiable authorization for autonomous agents, binding agent principal, request, and policy context with formal proof-of-concept.

Laguna S 2.1, a 118B MoE model from Poolside AI, achieves Deepseek v4 Pro performance at lower cost than v4 Flash.

Poolside AI co-CEO Eiso Kant describes building a model factory enabling efficient training of 118B MoE models competitive with 1T open-weight alternatives.

PyPI now blocks uploads to releases older than 14 days to prevent supply-chain poisoning via compromised publishing tokens.

PyroDash enables cost-efficient SLM-LLM collaborative inference by training SLM to emit control tokens requesting frozen LLM handoff during generation.

ELSAA: efficient attention mechanism combining low-rank and sparse approximations without decomposing projection matrices, extends Transformer context length.

Statistical framework for LoRA rank allocation: formulates importance-score design with explicit statistical interpretation for parameter-efficient fine-tuning.

Neuromorphic Receptron classifier for edge microcontrollers enabling non-linear decision boundaries without multi-layer networks.

Google DeepMind commits $40M in AI compute credits to Genesis Mission, supporting AI-driven scientific discovery across multiple domains.

OpenAI announces Project Camellia infrastructure investment in Effingham County, Georgia with commitments to energy efficiency, job creation, and Codex access.

LangGraph practitioner guide with three executable recipes for stateful multi-step agent workflows in business processes.

AdaFlash improves speculative decoding via adaptive diffusion drafters with on-policy distillation for LLM inference acceleration.

Report on drone computing infrastructure vision addressing software-hardware capability gaps for large-scale logistics, disaster response, and infrastructure inspection.

Google DeepMind releases Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber models for inference efficiency and security tasks.

Nativ wraps MLX in a macOS app for local inference, offering chat UI and localhost API similar to LM Studio.

OpenAI and Hugging Face disclose security incident during model evaluation, sharing findings on advanced cyber capabilities and defense lessons.

Patch Policy: robot control policy leveraging dense ViT features for fine-grained spatial reasoning without billion-parameter VLM overhead.

SWE-Pruner Pro: prunes long context in coding LLMs by extracting internal relevance signals, improving efficiency over external classifiers.

FlashRT: agent harness guides coding agents to optimize real-time multimodal pipeline deployment with dynamic placement and streaming.

Differentiable Logic Gate Networks enable low-latency EEG classification on edge devices via Boolean circuits instead of floating-point ops.

ClouDens detects anomalies in high-dimensional cloud system telemetry using context-aware methods for large-scale distributed infrastructure.

Heterogeneous adaptation pipeline enables on-device model personalization by offloading INT8 backbone inference to Hailo-8L accelerator.

VDAR-Router selects LLMs via verbalized query difficulty analysis for cost-performance-aware routing without embedding-only heuristics.

SelectInfer: neuron-level optimization framework for efficient LLM inference on edge devices via selective neuron loading.

Sobek: streaming optimization for equivariant tensor product convolutions on graphs via memory-efficient execution scheduling.

Cohere releases W4A8 quantization with vLLM integration; compares W4A16 and W8A8 schemes on NVIDIA Hopper.

Claude Code v2.1.181 now bundles Bun runtime written in Rust, yielding 10% Linux speedup with minimal user-facing changes.

PagedWeight: dynamic weight quantization for MoE LLM serving balances model precision against KV-cache memory in inference.

Thermodynamic computing blueprint uses Langevin dynamics in analog hardware for energy-efficient ML workload execution.

Honest Quorum Problem introduces epistemic faults for Byzantine fault tolerance in agentic validators, extending BFT guarantees to reasoning errors.

JoyNexus: multi-tenant post-training service for Vision-Language-Action models with efficient GPU resource pooling.

Loopie: MoE Transformer models (20B and 6B parameters) that outperform vanilla scaling by efficiently looping under fixed compute budgets.

Tabular foundation model for pre-fault dynamic security assessment in power systems using in-context learning to reduce labeling and improve contingency generalization.

Satirical proposal: hyperscalers mitigate data center water consumption by converting golf courses to parks, citing Google's 10.9B gallon 2025 usage vs. Coachella Valley golf water footprint.

Puter compiled Firefox to WebAssembly, enabling browser-in-browser execution; project cost ~$25k in Claude Opus tokens.

In-place tokenizer expansion for LLMs: reallocate vocabulary post-training to reduce latency/energy for underrepresented languages.

NIFA extends FPGA-integrated ReRAM in-memory computing to support nonlinear operations for efficient ML inference.

Hierarchical Global Attention + tiered KV storage enables 16K-token fine-tuning on 16GB VRAM, 8× longer than dense attention baseline.

Simon Willison compiled Go's mermaid-ascii library to WebAssembly for ASCII diagram rendering with color support.