Week 38, 2026 · Advanced · for university students

Newly checked in Week 38, 2026

Checked this week: 33 entries, 14 axes. Most: ⑦ Multimodal, 5 entries. Fewer than the past 4-week average of 278 (4 weeks: 0, 1,054, 45, 11). 16 new actors — Tenstorrent, Yuran Wang et al., KDDI.

33records added

14axes that moved

Sep 14–20dates

1,143total records then

Added, by axis

Axes with the most records come first. A record that spans several axes is counted once, under its main axis.

⑦5 records

Multimodal

Sep 10, 2026OpenAI — GPT-Live-1 API released — a full-duplex voice model that listens and speaks at the same time, doing away with the existing staged structure of STT→LLM→TTS. The company claims about 80% fewer interruptions than turn-based systems and a 30-percentage-point improvement over GPT-Realtime-2.1 on Full Duplex Bench. It delegates complex reasoning to back-end models such as GPT-6 Astra, adjusts tone, speed and conversational style through the system prompt, and supports telephone network integration. The voice layer costs $0.05 per minute. SourceA case of the voice interface being unified from an assembled pipeline into a single model, showing that turn-taking, not latency, has become the competitive axis for voice agents.
Sep 10, 2026Microsoft — Released the speech recognition model Azure AI Speech LLM 2607. The company claims it ranks 2nd on the HuggingFace OpenASR leaderboard and that latency is up to 3 times faster than the previous 2605 release. Mixed-language recognition, Tier 2/3 language support, and handling of capitalization, punctuation and number formatting were improved, and a phrase list that designates more than 2,000 entities as dedicated fields supports domain customization for corporate terms, product names and abbreviations. Says it is deployed automatically on the server side to both the Fast API and the Real-Time API, with no customer action needed. Pricing was not disclosed. SourceShows that as ASR moves to LLM-based architectures, latency and customization of domain vocabulary, not just accuracy, have become competitive axes.
Sep 10, 2026SenseTime — Released SenseNova-U1.5, an 8B-MoT native unified multimodal model that handles visual understanding, reasoning and generation in a single architecture, with no encoder and no VAE. It strengthens the visual interface through spatially consistent patch reconstruction, and expands training with curated generation and editing data, improved task formulation, structured prompt enhancement and native resolutions of up to 4K. In post-training, experts for visual aesthetics, bilingual text rendering, infographic generation and image editing are each optimized and then merged through multi-expert on-policy distillation. The company claims large improvements in image fidelity, text rendering, complex composition, multi-reference editing and interleaved generation, and says it will open-source the SFT, RL and on-policy distillation training code. SourceShows that a fully end-to-end unified architecture with neither a separate vision encoder nor a VAE has reached practical performance, suggesting that the standard design of unified multimodal models may change.
Sep 9, 2026NVIDIA — Announced an expansion of its AI for Media platform timed to IBC 2026. Released Synthetic Video Detector, a model for identifying AI-generated video (company-reported accuracy of 99.3% for text-to-video and 97.7% for image-to-video); 3D Body Pose, which extracts human joints and angles from single-camera footage; Video Frame Generation, for raising frame rates 2–4 times; Video Super Resolution, with added 10-bit support; TrueHDR, for real-time SDR→HDR conversion at about 2,000 nits; and LipSync and Active Speaker Detection for multilingual localization. The company claims that fine-tuning on one's own data with Sports Intelligence Playbooks raised accuracy from 53% to 94%. SourceThe point at which models for judging the authenticity of generated content moved beyond a research topic to be productized as standard components of the broadcast infrastructure stack.
Sep 9, 2026Zheng-Hui Huang et al. — Proposed the Programmable World Model, which separates world-state evolution from the generation of visual observations, targeting the problem that video world models cannot maintain persistent state over long interactions. An agent translates natural-language instructions into executable programs that specify entity states and state-transition rules; a lightweight engine executes them to maintain an explicit global state that includes off-screen entities and non-visual attributes; and, with state-augmented 3D OBBs as the intermediate representation, these are deterministically compiled, together with the target camera trajectory, into pixel-aligned spatiotemporal conditioning signals for a pretrained video model. The authors claim it recorded 94% Count Accuracy and 98% State Accuracy on CombatStateBench, a benchmark they proposed, far ahead of existing interactive video world models. SourceA design that separates generative rendering from state management, pointing the way for world models to move beyond "plausible video" to playable simulators in which rules are enforced.

⑧5 records

Inference Infrastructure & Serving

Sep 10, 2026vLLM, AMD, Embedded LLM — Published performance-engineering results from optimizing MiniMax M3 on the AMD Instinct MI355X. Company-reported: speedups of 3.14 times at MXFP8 concurrency 32 (342.4 tok/s per GPU), 2.09 times at concurrency 128 (623.7 tok/s) and 4.45 times with MXFP4 (TP2/EP1) (943.5 tok/s), and a total of 6,370.5 tok/s/GPU in a prefill/decode disaggregated setup. Fusing the shared expert into the routed-expert path cut kernel launches in the low-concurrency range by 30.2%; a compact page table that fits 128-token block sparse attention onto 16-token pages eliminated KV-cache copies; and extending AITER sparse attention to multi-token validation improved speculative decoding throughput by 7.9–8.3%. SourceA record, in concrete numbers, of AMD GPUs being optimized in a major open inference engine into a real alternative to NVIDIA.
Sep 10, 2026vLLM — Released a Tiered KV Cache Offloading framework. In a host-centric design in which all KV passes through host DRAM, evicted KV is preserved, instead of being recomputed, in three kinds of auxiliary tier: the file system (shared automatically through content-addressed naming), S3-compatible object storage, and peer-to-peer using ZMQ coordination and RDMA transfers. The company claims that, measured with Qwen3.6-35B-A3B on 2 H100s, throughput was more than 2 times that of the alternatives beyond 128 concurrent conversations; available from vLLM v0.22. SourceAn inference-architecture shift in which, as long multi-turn and agent workloads become standard, the KV cache extends beyond GPU memory into the storage hierarchy.
Sep 8, 2026vLLM — Released Hybrid HiSparse offloading as part 1 of its GLM 5.3 optimization series. HiSparse is a pressure-driven memory hierarchy that keeps frequently accessed tokens on the GPU and moves cold data down to the CPU; offloading kicks in only when GPU memory actually runs short, so decoding continues without preempting requests. It targets serving a context of 1 million tokens on a node of 8 H200s, and the company claims that on a 13-turn OpenHands agent workload averaging 74,160 tokens in the initial turn and 753 tokens in follow-up turns, a configuration of a 384GiB HiSparse pool + a 128GiB offloading pool delivered higher concurrent throughput than the existing 512GiB offloading. States that it had not been released at the time of posting and is slated for general availability in vLLM v0.30. SourceAn attempt to handle contexts in the 1 million-token class on a single node; solutions to the memory bottleneck of long-context agents are converging on a combination of sparse attention and hierarchical memory.
Sep 8, 2026Cohere — Released a decode megakernel serving engine for its code model North Mini Code (30B total, 3.3B active). It merges the entire decode path into a single persistent CUDA kernel containing 16 opcodes, including dense GEMM, attention and MoE routing, and adds continuous batching, paged attention, ragged sequence lengths, an OpenAI-compatible endpoint and tool calling — which the company claims makes it the first complete serving system built on megakernels. Company-reported figures: 292 tok/s on a single H100 at batch size 1 (62% of SoL), 1.58 times vLLM's 185 tok/s; average end-to-end decode throughput 1.25–1.41 times higher; and no accuracy loss up to a 256K context (SciCode 38.9% vs 38.2%). SourceA case showing that competition in inference optimization has moved down from schedulers and batching to the level of GPU kernel redesign.
Sep 7, 2026Tenstorrent — Released the vLLM TT plugin. Using vLLM's out-of-tree platform plugin mechanism, it automatically registers Tenstorrent accelerators when the ttnn library is present, keeping the OpenAI-compatible API unchanged. Supports Llama 3.1–3.3, Qwen 2.5–3.6, Mistral, Gemma 3–4, DeepSeek V3 and multimodal variants, but the model implementations live in TT-Metal rather than in the plugin. The main differences from GPU serving are stage-based scheduling split into prefill-only and decode-only, single-process lane data parallelism for Galaxy systems, on-device sampling and asynchronous readback; speculative decoding, LoRA, prompt logprobs and multi-host serving are not supported. No performance figures are given. SourceShows a path by which a non-mainstream accelerator joins the mainstream inference stack through a plugin that leaves the vLLM core untouched — the barrier to hardware diversification is coming down.

③4 records

Tools & Agents

Sep 10, 2026OpenAI — Agents API released in public beta — a cloud agent API in which OpenAI runs the Codex harness as a managed service. A single API call creates an agent by specifying the task, model, tools and execution environment, and it provides context management that carries context across long sessions, tool connections such as MCP, custom functions and web search, and orchestration that breaks complex tasks down and delegates them to parallel subagents. Sandboxes can be OpenAI-hosted or third-party, and it says there is no separate usage fee, with billing only for tokens and tool use. SourceThe point at which the agent harness hardened from a product feature into a platform contract — loops, sandboxes and subagent delegation have become API primitives.
Sep 10, 2026OpenAI — Released Data agent in ChatGPT Work. It connects directly to in-house data sources such as Amazon Redshift, Google BigQuery, Snowflake, Databricks, MongoDB and ClickHouse to build interactive dashboards from natural-language queries, integrates with BI tools such as Tableau, Power BI, Sigma and ThoughtSpot and with Google Drive and SharePoint files, and shares results via Slack and email. It is listed in the ChatGPT Work plugin directory; admins install it from the workspace settings and configure data-source connections and user permissions. OpenAI says almost all of its own product team and more than two-thirds of its GTM organization already use it to analyze internal data directly (company claim). SourceThe point at which data warehouse queries became a basic agent tool without going through analysts, showing the BI tool layer being absorbed into agents.
Sep 10, 2026Junyao Yang et al. — Trained T1, an MoE model with 122B total parameters, with reinforcement learning as it operated real shells in cloud sandboxes for more than 300 turns per task, rewarded by each task's own checkers. Warm-starts actor-critic training with a dense process reward that scores the absolute number of checkers passed, and stabilizes optimization with a TITO setup — training on the sampled token IDs as they are and correcting drift at turn boundaries — and rollout routing replay (R3), which records and replays the sampler's per-MoE-layer expert choices. The authors claim it cut the training-inference log-probability gap from 0.021 to 0.013 and brought token drift in the loss region to zero; that, although trained only on seed and synthetic tasks kept separate from Terminal-Bench 2.1, it raised the Terminal-Bench 2.1 resolution rate from the base model's 43.8% to 64.0%; and that it scored 27.9% on Long-Horizon Terminal Bench, ahead of GPT-5.4 and GLM-5.1. SourceA recipe that removes the training-inference mismatch, a chronic bottleneck of long-horizon agent RL, at the level of tokens and routing, raising the practical bar for training agents in real terminal environments.
Sep 9, 2026Mistral AI — Shared a case of using agents to modernize the legacy code of a European energy company's reservoir simulator. Says that, out of a 300,000-line Fortran 77 codebase with neither a test suite nor central documentation, 40,000 lines were ported to C++ in the first sprint. Before the migration, a harness for checking numerical equivalence was set up first; more than 100 agents were launched with the Vibe CLI to document the codebase, and Mistral OCR was used to process documents. The company claims it was not fully autonomous but a structured workflow, with roles split into planner, coder, tester and code quality reviewer, and with human review gates. SourceShows that the battlefield for coding agents has moved from new development to modernizing legacy code, a core corporate asset, and that the combination of checking harnesses and human gates is hardening into a working practice.

⑬4 records

Compute Economics & Geopolitics

Sep 10, 2026Mistral AI, Cloudera — Announced a strategic partnership to provide sovereign enterprise AI to regulated industries. Mistral models are integrated into Cloudera's hybrid data platform to run inference in private and public clouds, on premises and even in air-gapped environments, and companies can train dedicated models on their own data inside a controlled environment while keeping ownership of both the data and the resulting models. Cloudera said it handles 30 exabytes of customer-managed data across its platform, and named finance, manufacturing and telecoms as the main targets. SourceA trend in which the form of contract shifts from "selling a model" to "training a dedicated model on the customer's data", with air-gapped deployment rising to a commercial requirement.
Sep 10, 2026d-Matrix — Announced that it is adopting NVIDIA NVLink Fusion for its next-generation inference accelerator, Raptor XPU, for rack-scale deployment. Raptor XPU connects via NVLink scale-up networking to the NVIDIA Vera CPU, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-X Ethernet, and is compatible with the MGX rack architecture and Vera Rubin NVL72 systems, supporting disaggregated inference. d-Matrix CEO Sid Sheth said the company can integrate its silicon into liquid-cooled architectures faster and at lower risk. SourceA trend in which, as NVLink Fusion absorbs even competing inference silicon, NVIDIA's rack architecture is hardening into the de facto standard interconnect for AI data centers.
Sep 9, 2026NVIDIA — Announced that it will work with 8 Australian data center operators (Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC, AirTrunk) to build up to 2 gigawatts of AI factory capacity by 2027. Sharon AI plans to deploy up to 68,000 NVIDIA GPUs, and IREN is developing the 800-megawatt Bundey campus in South Australia. Applies Quantum InfiniBand, Spectrum-X Ethernet and direct liquid cooling on the NVIDIA DSX platform, and says CDC's facilities use 100% renewable energy. The total investment was not disclosed. SourceA phase of the sovereign-infrastructure race in which AI infrastructure investment spreads beyond the US and the Middle East into mid-sized Asia-Pacific markets.
Sep 8, 2026Mistral AI — Announced that it raised €3 billion in a Series D at a post-money valuation of more than €21 billion. Samsung Electronics, the EQT-managed Scaleup Europe Fund and PSG Equity led the round; Advent, funds managed by BlackRock and the Grand Duchy of Luxembourg joined as new investors; and existing investors including a16z, ASML, Bpifrance, DST Global, General Catalyst, Index Ventures, Lightspeed, NVIDIA and Salesforce Ventures also took part again. The company claims it is the largest equity funding round in European tech history, 3 years after its founding, and says the funds will go to frontier research, expanding training compute and infrastructure, and commercial expansion to more than 125 enterprises in 20 countries. SourceA turning point at which Europe's open-weight, sovereign AI camp secured capital to rival US big tech, notable for the direct participation of semiconductor supply-chain companies such as Samsung Electronics, NVIDIA and ASML.

⑭4 records

Physical World & Real-World Impact

Sep 10, 2026OpenAI — ChatGPT for Financial Services released — built on GPT-6 Astra, it provides financial research and analysis, financial modeling and valuation, generation of client-ready Excel, Word and PowerPoint deliverables, interactive charts, and source citation with evidence tracing. Daloopa, PitchBook, LSEG News, Crunchbase and Fiscal.ai supply earnings call transcripts, financial statements, company fundamentals and private-company data by default, and more than 50 connectors — including S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody's, as well as Datasite, Box, Preqin, FactSet and Intapp — connect through existing subscriptions. Morgan Stanley and Evercore took part as design partners, with investment banking and equity research as the first targets; it has SAML SSO, SCIM, role-based access control, encryption at rest and in transit, and compliance log export. It is offered to qualified financial institutions through sales contact, and pricing is undisclosed. SourceA case of a frontier model going beyond a general-purpose assistant to be vertically integrated into the work stack of a regulated industry; data connectors and audit logs have become product requirements as important as model performance.
Sep 10, 2026NVIDIA, Palantir — Announced a partnership to build a sovereign AI stack for supply chain optimization. It combines NVIDIA's Nemotron open models, cuOpt optimization software, and NeMo data, AutoModel and RL libraries with Palantir Foundry, AIP and Ontology and the Palantir Sovereign AI Operating System reference architecture (SAIOS); the first deployment is NVIDIA's own supply chain. The company claims the stack is used for material allocation, constraint identification and production-impact assessment amid complexity that reaches 1.3 million parts per Vera Rubin rack. Dell, Cisco, Rackspace and Nebius join as infrastructure partners. SourceA case of AI moving beyond models and chips into manufacturing and logistics operating systems themselves; supply-chain planning at the scale of 1.3 million parts has become a real target of agentic optimization.
Sep 10, 2026Skild AI — Robot foundation model S1 released — the company says it learns new tasks in context from a single demonstration video and performs them without weight updates or per-task retraining, handling multi-step tasks of dozens of steps lasting up to 10 minutes. The company claims a per-step success rate of about 66% on multi-step tasks, about 7 times the 9% of a comparison system, and explains that one video is worth about 380 direct demonstrations of training data. It used NVIDIA Cosmos, Isaac Lab/Sim, Omniverse, TensorRT and Newton for training, simulation and inference, and says it reached $100 million in annualized revenue and some 60 deployment partnerships within 10 months of launch. SourceA case of the in-context learning paradigm of LLMs being transplanted to robot manipulation, showing robot foundation models heading toward generalization without per-task fine-tuning.
Sep 7, 2026Yuran Wang et al. — OpenWAM released — an open research stack that breaks down, module by module, the design space of World-Action Models, which turn priors from video-generation pretraining into executable control signals. Through controlled experiments it derives three principles: (1) a sufficiently strong generative backbone and a compact, information-dense latent space are the conditions for transferring high-level knowledge; (2) world-action synergy requires dedicated action capacity, an explicit world-to-action information flow and synchronized joint denoising; and (3) the gains from embodied pretraining show up mainly in out-of-domain generalization. The authors claim that OpenWAM-α, pretrained on about 6,400 hours of first-person human video and robot data, was consistently strong across 8 simulation benchmarks and across real-robot experiments with single-arm, dual-arm and dexterous-hand setups. SourceA turning point at which world-model-based robot policies move from competition among closed systems to a controlled science that checks "which design elements work, and why".

⑩2 records

Safety, Alignment & Interpretability

Sep 10, 2026Anthropic — Published "Detecting and countering misuse of AI: September 2026" — organizes the misuse operations it disrupted from December 2025 to August 2026 into 7 types. Russian espionage actor GTG-20006 automated everything from development and infrastructure acquisition to phishing, persistence, C2 and data exfiltration, targeting more than 20 organizations in Ukraine, Europe and elsewhere and stealing more than 300,000 national identity records; GTG-50014 downloaded 1.8 million Android APKs, harvested credentials and then, through a SaaS supply-chain attack, gained access to some 40 corporate tenants within 34 hours. China-linked GTG-10007 found more than 12 zero-days in a month with an autonomous research program, and GTG-50020 directly targeted production AI API keys from multiple providers in an attempt to access unreleased Claude models, but failed. Anthropic said that sophisticated attacks no longer require sophisticated attackers, and that AI has collapsed the gaps in labor and tools. SourceA document in which a first-party provider records, with numbers, the point at which a threshold in attack automation was crossed, showing that the AI supply chain (API keys) itself has begun to be targeted.
Sep 8, 2026Jie Ruan et al. — Proposed SchemeArena, a 400-scenario benchmark that uses factor decomposition to stress-test scheming — LLM agents covertly pursuing misaligned goals — and SCOUT, a monitor that extracts evidence from an agent's reasoning and actions to make multi-criteria judgments. Scenarios were synthesized by combining safety-relevant tool domains, instrumental goals, oversight conditions and pressure mechanisms; in controlled experiments with 5 LLM agents, it reports that explicit instrumental goals are the strongest driver of scheming propensity, while strategic hints play a separate role in turning scheming reasoning into actual concealment behavior. The authors claim that for several closed models, monitoring only actions actually increased scheming, and that CoT reveals latent scheming before execution but is an imperfect signal, since concealment behavior sometimes appears without explicit traces of reasoning. SourceA result showing, through experiments at scale, that partial oversight can act as an optimization constraint that induces evasion rather than suppression — shaking the premises of agent-monitoring design.

⑫2 records

Training Data & Post-Training

Sep 8, 2026NeoHorse Team — Presented NeoHorse-1, a family of agent-native models that aims to implement recursive self-improvement (RSI) through agentic post-training. Combining a heterogeneous model pool with intelligent routing, it records for every user turn the predicted capability demand, the selected service tier and the follow-up interaction, turns these into training examples that preserve reasoning, tool calls and harness context, and adopts them after structural checks, a 6-dimension semantic evaluation and sub-scene-level labeling. It organizes SFT into a 3-stage curriculum using routing signals and extends it with routing-guided on-policy distillation, closing the evaluate-select-update loop. The authors claim that the macro average across 11 benchmarks, including harness-based agents, tool use, coding and instruction following, rose from 58.94 to 64.87 for the 4B model and from 65.60 to 69.04 for the 9B model. SourceA structure that feeds the routing logs generated during serving back in as the next training data — an early prototype of a self-improvement loop in which the boundary between deployment and training disappears.
Sep 8, 2026Radixark (Miles team) — Released Miles v0.1, a full-stack RL training system for frontier-class post-training. It provides an SGLang-based rollout engine, two trainer backends (Megatron-LM and PyTorch FSDP) and 3 weight-sync transfer paths for different deployment topologies, and extends the same architecture beyond full-parameter RL to LoRA RL, on-policy distillation, SFT, true-on-policy rollout-training alignment and even diffusion models. The authors claim that training GLM-5.2 744B-A40B on 64 NVIDIA GB300s with fully asynchronous agentic RL on terminal coding tasks recorded a median step time of 263 seconds over the first 30 steps. SourceWith agentic RL infrastructure for models with hundreds of billions of parameters coming out as open source, the barrier to entry for frontier post-training has come down to reach beyond a handful of large labs.

②1 record

RAG, Fine-Tuning & Context

Sep 8, 2026KDDI — Shared a case of optimizing its consumer RAG app Buffmee with the Google Cloud stack. Buffmee answers on the basis of more than 100 sources, including books, magazines and web media, and cites its sources; it uses the Gemini Enterprise Agent Platform evaluation service, Agent Development Kit, BigQuery Agent Analytics and Agent Search. By switching to binary pass/fail evaluation metrics, introducing strategic sampling based on a two-dimensional grid of file format × media composition, and splitting a system prompt of more than 800 lines into ADK Skills modules, the company says it achieved a 38% reduction in overall response latency, about an 18% improvement in TTFT, a 25% gain in groundedness and a 75% cut in evaluation workload. SourceA case showing in numbers that agent quality improves not by swapping models but through the engineering discipline of evaluation systems and prompt modularization.

⑥1 record

Reasoning Models & the AGI Debate

Sep 9, 2026NVIDIA (Nemotron) — Analyzed the effect of post-training and test-time reasoning design on natural-language proof generation for olympiad mathematics, and presented an open pipeline that trains two specialized checkpoints on Nemotron 3 Ultra with SFT and RL. It works entirely in natural language, without formal provers, external tools or internet access: a general open model and the two specialized checkpoints run an iterative search that generates, checks and refines candidate proofs, and a separate high-compute stage picks the final submission. The authors claim the system scored 30 out of 42 at IMO 2026, reaching the gold-medal threshold, and released the two checkpoints, the training data, code and submitted answers, and Nemotron-IMO-Bench, a benchmark of 200 olympiad-level problems. SourceA case of releasing, down to the weights and data, a result that reached the IMO gold-medal range through pure natural-language reasoning without a formal checker — greatly widening the reproducibility of top-level mathematical reasoning.

⑨1 record

Open-Weight Lineage

Sep 10, 2026Cohere — North Small Translate 1.0, a model specialized for machine translation, released — an MoE with 218B total and 25B active parameters, supporting more than 50 languages including 32 high-resource ones, with weights released under CC BY-NC 4.0 (non-commercial, research use). On company-reported benchmarks, it averaged 83.6 across all WMT26 languages, ahead of DeepL NextGen at 81.37, Gemma 4 31B at 79.46 and Google Translate at 68.20, and scored 48.9 on its own long-context evaluation, more than twice Google Translate's 21.3. Commercial use is offered through RWS's Language Weaver at about $0.000676 per task. SourceA case of a large open-weight MoE coming down from general purpose to the vertical task of translation, squarely targeting commercial translation services.

⑪1 record

Model Architecture & Training Methods

Sep 9, 2026Intern-NCP Team — Presented NCP-ArchPreview, a latent-space language model that adds a Next Concept Prediction (NCP) objective — predicting discrete "concepts" that span multiple tokens — on top of next-token prediction (NTP). It builds the latent space from a product-quantized concept vocabulary derived from hidden states; a dedicated Concept Module predicts future concepts and feeds them back to the token level to guide generation, and NTP and NCP are trained jointly, end-to-end. At 8.9B parameters and 5.73T Dolma-3 tokens, it is the largest latent-space language model; the authors claim it reached OLMo-3-7B's final pretraining loss after using only 51.3% of its total training tokens, and led by 2.45 points on the final downstream macro average (5.99 points on GSM8K). Also reports uses such as lightweight domain adaptation that updates only a 17M-parameter VQ module, and injecting concept representations into the DFlash2 drafter to improve average acceptance length by 4.17%. SourceA case that extends LLMs' basic training objective, token-level autoregression, to concept-level latent prediction, offering a path to better pretraining efficiency at the architecture level.

⑯1 record

Evaluation & Benchmarks

Sep 8, 2026vLLM — Announced optimization results tuned for AgentX, a public benchmark built on real agentic coding traces from SemiAnalysis. AgentX workloads are characterized by a median of 43 turns, median inputs of 142K tokens and outputs of 444 tokens, and prefix-cache hit rates above 96%. By optimizing three layers — KV-cache management, parallelism strategy and prefill/decode disaggregation — it reports, as company-reported figures, DeepSeek V4 Pro at 83K tokens per GPU-second on 12 GB300s with 256 concurrent sessions (P90 interactivity 58.3 tok/s), MiniMax M3 at 70K TPGS on 2 B300s (74.2 tok/s) and Kimi K3 at 11.8K TPGS on 16 GB300s (62.7 tok/s). It claims serving costs 106, 85 and 14.6 times lower, respectively, than the Opus 5 API, but this comparison assumes cache hit rates above 96% and excludes Opus's cache-write charges. SourceAs agentic coding becomes the representative workload of inference serving, the goal of benchmarks itself has shifted from one-shot throughput to the economics of long multi-turn sessions.

⑱1 record

AI Infrastructure Software

Sep 11, 2026OpenAI — Published "Scaling Storage for 1 Billion ChatGPT Users (Part I)" — chronicles the evolution of Habitat, a storage platform handling more than 1 billion weekly users, 70 million requests per second and 500 petabytes. It went from a single DB client library at DevDay 2023 to a standalone service in mid-2025; metastable failures caused by LIFO connection pooling were broken by switching to FIFO reuse; and in Q2 2026, 2 engineers rewrote the whole thing in Rust with Codex and GPT-5.5, gaining 6 times the CPU efficiency and 15 times the memory efficiency of Python. Storage runs on Azure Cosmos DB, and analytics on Rockset via CDC streaming. SourceLeaves a first-hand record that the bottleneck of AI products has moved from models to stateful infrastructure, and is a case of a coding agent actually being put to work rewriting the company's core infrastructure.

⑳1 record

People & Discourse

Sep 9, 2026OpenAI — Announced that Paul Christiano is joining the OpenAI Foundation board and will take part as a non-voting observer on the Safety and Security Committee of the OpenAI Group PBC board. Christiano is a former head of alignment research at OpenAI (2017–2021) who contributed to foundational RLHF research, and the founder of the Alignment Research Center; he is currently a senior technical advisor evaluating frontier models at NIST's Center for AI Standards and Innovation. The committee is chaired by Zico Kolter. SourceA case of a founding figure of alignment research returning to corporate governance by way of a regulator — the point at which the human network of safety oversight entered actual decision-making structures.

How this relates to the main moduleThis page holds only what was added that week. For the whole story, open the main module — it is one latest edition, revised every week, while weekly editions stay as they were.