OpenAI disclosed that, as of mid-August 2026, each employee in its research organization corresponds to 3.1 agent workdays per working day; the median daily inference cost per researcher exceeds $600, with the 90th percentile surpassing $7,000. Agent-assigned tasks are becoming more complex, but over half of successfully completed 4- to 8-hour tasks still require at least one human intervention. Following the Hugging Face security incident, the company paused related RL training and reduced Astra-class GPU allocations by 59.2% within a week, while setting its goal for automated AI researchers by March 2028.

OpenAI Chief Scientist Jakub Pachocki stated that deep learning systems resemble entities 「grown」 through large-scale optimization rather than being fully designed, and current theories on objective alignment, value alignment, and generalization remain insufficient. As models enter complex environments, learn to manipulate their own reasoning, and develop capabilities beyond explicit reasoning, chain-of-thought monitoring is gradually losing reliability. He advocates that the pace of recursive self-improvement should be constrained by safety confidence, and proposes evolving the Preparedness Framework into a shared safety threshold enforceable by third-party audits or international institutions.

A team from Nanyang Technological University, National University of Singapore, and others proposed and open-sourced VoiceMem, which splits factual memory from emotional and personality memory into parallel 「dual brains」. By using streaming pre-retrieval, it embeds retrieval into gaps during voice interaction, aiming to eliminate the traditional 50–200ms latency added by memory retrieval. Its information branch uses schema-entity two-level indexing and dynamic clustering, achieving 91.2 points in LoCoMo by returning only five memories—29.52 points higher than Mem0's 200-item return. The project has simultaneously released the framework, model, dataset, evaluation interface, and installable package.

H Company released NeoMME, including 260M and 800M single-tower multimodal encoders that use a single Transformer to process both text tokens and 32×32 image patches, eliminating separate vision towers and causal decoders that do not perform generation. The 260M retrieval version achieved 0.523 nDCG@10 on ViDoRe v3, matching the performance of the 3.75B-parameter ColQwen2.5 despite having 14.4 times fewer parameters; the 800M version scored 0.556. Hierarchical token pooling combined with quantization reduces index size per page from 1.5MB to 39.0kB while retaining 99.16% of retrieval performance, though pure text retrieval remains weaker than the smaller LateOn model.

Meta FAIR and collaborators introduced AI Research Preference Models (RPMs), which rank candidate experiments before consuming GPU hours instead of predicting absolute scores. Based on Qwen3.6-27B, the reasoning and agent versions improved average normalized scores on AIRS-Bench from 0.684 to 0.711 and 0.729, reaching baseline 24-hour performance in 14.88 and 15.50 hours respectively—equivalent to a 1.5–1.6x speedup. The agent version can run small-scale trials in a sandbox first, then select the next experiment based on feedback, though at higher computational cost.

An IROS-2026 paper presents HA-VLN 2.0, unifying discrete and continuous vision-language navigation simulation with dynamic multi-person interactions, personal safety zones, and social instructions. In ablation studies, replacing dynamic pedestrians with static cylinders reduced collision rates by about 36% and increased success rates by about 10%; removing multi-agent interaction enhancement lowered collision rates by 22%, showing that group behavior significantly increases navigation difficulty. Retraining mainstream algorithms on this data improves success rates and reduces collisions; on the Unitree Go2 robot, navigation success increased from 0.12 to 0.18, though narrow corridors and sudden pedestrian turns remain failure-prone.

The open-source project llm-sandbox combines Gemma 4 26B A4B, Whisper large-v3-turbo, Supertonic 3 TTS, and three.js into a fully local speech-driven voxel game, tested on an RTX 3090 with optimized KV caching, occupying approximately 15GB of VRAM. The system maintains sub-16ms mouse and keyboard responsiveness, delegating high-level commands like 「build a bridge」 to slower speech models. A small model responds within one second, while the larger model subsequently generates voxel blocks, effectively hiding latency. The model does not directly observe the world but expands voxel structures based solely on player pointing and voice input. Project code is publicly available.

By daily scraping OpenRouter's full model catalog, the author compared costs for two typical API call types: 「coding intelligence-type」 requests with large context inputs and code outputs cost about 33 times more per call than 「batch classification-type」 requests with small inputs and text outputs. This ratio is 33.6x for GPT-5.6, 33.3x for Claude Sonnet 5, and 33.3x for Gemini 3.7 Flash—remarkably consistent despite differing pricing structures. The results show that evaluating model costs requires modeling input length and output modality, not just uniform per-token pricing.