§ Infrastructure

OpenAI scaled Habitat storage system with multi-tenancy and Azure Cosmos DB to achieve 12× read speedup for 1B+ ChatGPT users.

OpenAI details containment measures for frontier model training: offline environments, compute isolation, and red-teaming infrastructure.

OpenAI launches Decisions API for low-latency, cost-efficient inference in applications requiring fast structured outputs.

Proposes low-precision storage for n-gram embedding tables in language models, targeting 196B-parameter tables in DeepSeek-V4.1-Flash without performance loss.

Identifies and fixes two numerical stability bugs in softmax/cross-entropy implementations: Softmax Collapse and Softmax Cancellation.

BeNice addresses looped architecture inference memory via KV-cache sharing across recurrences, investigating shareability mechanisms.

Modal CEO Bernhardsson analyzes why AI compute infrastructure defaults to rectangular resource allocation, examining elastic GPUs and sandbox economics for agent deployment.

TTR-Compressor uses test-time regression for sequence compression in LLMs; learns end-to-end via backprop, computes coefficients in closed form at inference.

Empirical evaluation of LLM inference optimizations (KV-cache pruning, speculative decoding, quantization) on Apple Silicon vs NVIDIA GPUs.

T2Quant: task and time-aware mixed-precision quantization for diffusion language models using label-free sensitivity signals.

RaSFNet: ray-aligned fusion network for LiDAR semantic segmentation in autonomous driving and robotics perception.

RIFT addresses decode-time KV cache staleness in speculative decoding via complementary compression training, improving long-context draft model efficiency.

CAKE system automates AscendC kernel generation for Ascend NPUs via adaptive multi-DSL routing and agent-guided evolution, addressing data scarcity gap.

Neural network transistor models for 3nm nanosheet accelerate SPICE simulation but require validation in full EDA tool flows for standard cell characterization.

Study measures RAG chunk update costs under real document revisions; content-defined chunking outperforms fixed/recursive strategies on Chinese and English Wikipedia edits.

OpenAI explains forward deployed engineering (FDE) role: embedding engineers with customers to drive AI product adoption and integration.

Cohere Model Vault enables verifiable confidential inference using CPU/GPU computing with end-to-end encryption and attestation.

HDC-FL clusters federated learning clients via classifier heads to personalize shared encoder training without explicit communication overhead.

Open-source telecom alarm root-cause generator using real Internet topologies and RFC 8632 standards; dataset benchmarking contribution for infrastructure diagnostics.

Energy-driven pathfinding for RAG with knowledge graphs addresses semantic loss and modality mismatch in multi-hop QA.

OpenAI models now available via Google Cloud Marketplace through Gemini's Model Garden, expanding enterprise access.

OpenAI details LLM-powered ad stack architecture for ChatGPT, separating ads from answers via conversational intent understanding.

DeepSeek releases DeepInfer, open-source inference stack processing 155T tokens/day at ~80% margin, enabling $22M/day revenue with 81-day payback period.

SF Compute's Evan Conrad outlines programmable data centers with GPU markets, physical settlement, and decentralized compute trading infrastructure.

IndustrialRemembr extends NVIDIA ReMEmbR spatiotemporal memory system to robot fleets with on-demand scene encoding to reduce compute costs.

Edge-aware attention layer for graph generation incorporates structural information while leveraging hardware-optimized Transformer implementations.

Multi-agent framework using property graphs for autonomous root-cause analysis in distributed microservice logs.

Cache engineering for LLM agents optimizes KV storage and prefilling costs across long trajectories without invalidating cached states.

Token-Energy Law models GPU energy per decoded token as function of active bytes and realized time, enabling efficient LLM deployment.

Analysis of cluster collapse in federated learning and control-theoretic interventions for stable client routing.

Freshness-weighted gossip aggregation protocol for decentralized federated learning handling client outages and stale local models.

Vector quantization and calibration data selection enable 2-bit post-training quantization of large models while maintaining usability.

BitMTP combines ultra-low-bit quantization with multi-token prediction for efficient LLM decoding without quality loss.

AsymLoop separates context encoding from iterative refinement in looped Transformers to improve effective depth without scaling parameters.

Systematic study of DPO hyperparameter landscape across model scales reveals optimal learning rate and β-regularization shift with size; direct transfer between scales unreliable.

PhyNet framework for industrial fault diagnosis treats physical variables as global constraints and response signals as diagnostic evidence in heterogeneous sensor streams.

VISH framework selects compact hotword subsets for speech-LLM contextual ASR via frame-level CTC posterior verification; training-free and scalable via shared-prefix Viterbi search.

COMPASS agentic router selects optimal LLMs by profiling capabilities via tools and pairwise selection, achieving 69.85% accuracy at lower cost than trained baselines.

Privacy-preserving inference embeds user prompts locally via 55.6M-param transform before passing to frozen Qwen 27B, preventing host access to plaintext.

RAG-NIMBF benchmarks RAG pipeline configurations via non-invasive ablation to guide chunking, embedding, retrieval, and generation model choices.

ProVe reverses speculative decoding: large model proposes token blocks, smaller model verifies in parallel, improving LLM inference speed.

LatBridge enables safe model replacement in multi-agent systems by bridging latent-communication interfaces across heterogeneous agent architectures.

SliMOO applies multi-objective evolutionary search to LLM depth pruning, preserving task-diverse capabilities while reducing memory and latency.

AC/DC on a Budget reduces dense training phase cost in sparsification by zero-initializing masked parameters to optimize FLOPs-performance tradeoff.

OpenAI acquires Orchestra Bio to apply AI to life sciences R&D, targeting drug discovery and development acceleration.

LatComm introduces reusable latent context compression between LLMs, enabling compact encoding that is readable across multiple models and questions.

WildAWM synthesizes executable GUI/web/code environments for training computer-use agents without real-world security risks.

EdgeDRM enables decentralized federated learning inference-time collaboration via learned routing potentials allowing queries to reference neighboring clients.

TokenGate enables learnable token pruning in LLMs for faster, memory-efficient long-context prefill using a differentiable top-K approximation.

Polar Switch achieves faster polynomial orthogonalization for gradient descent in language model optimization with lower FLOP cost than quintic methods.