Week 36, 2026 · Advanced · for university students

Newly checked in Week 36, 2026

Checked this week: 45 entries, 15 axes. Most: ⑲ On-Device & Local Inference, 6 entries. Fewer than the past 4-week average of 264 (4 weeks: 0, 0, 0, 1,054). 22 new actors — Liquid AI, Arm, Google (Chrome).

45records added

15axes that moved

Aug 31–Sep 6dates

1,099total 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.

⑲6 records

On-Device & Local Inference

Aug 31, 2026HUMAIN, Qualcomm — Unveiled the Horizon Ultra AI PC at LEAP 2026 (Riyadh) — based on the Snapdragon X2 Elite (18-core Oryon CPU, Adreno GPU, Hexagon NPU), it spreads AI work across the CPU, GPU and NPU. Enterprise purchases of the Windows version begin on Sep 20, 2026; a HUMAIN OS version is planned for 2027. The press release gives no NPU TOPS figure or price SourceA Saudi sovereign wealth fund affiliate puts forward an on-device AI PC as the endpoint of a sovereign AI stack — a phase in which local inference becomes an item of geopolitical procurement
May 19, 2026Google (AI Edge) — LiteRT-LM generally available (GA) — added a native Swift API for iOS and a WebGPU-accelerated JavaScript API. Supports Android CPU, OpenCL GPU and NPU, iOS Metal and browser WebGPU. With Gemma 4: 52 tok/s decode on Android GPU, 56 tok/s on iOS GPU, 76 tok/s on the web on a MacBook Pro M4 Max, up to 2.2× with Multi-Token Prediction, 607MB of memory on iOS SourceThe stage at which an on-device runtime covers not just mobile OSes but browsers with the same stack. Figures are the vendor's own measurements
May 5, 2026Google (Chrome) — Opened the Prompt API to the web in Chrome 148 stable — for extensions it has been stable since Chrome 138. Web pages access the browser's built-in Gemini Nano directly, with support for text, image and audio input and for constraining responses with regular expressions and JSON schemas. Requirements: 22GB of free space, and more than 4GB of VRAM or 16GB of RAM and 4 or more cores SourceA stable-channel case of an on-device model exposed as a web platform API rather than an app. Responsibility for model distribution and storage management passes to the browser vendor
Nov 28, 2025Liquid AI — Published the LFM2 technical report — expanded to a 350M–8.3B family (dense 350M, 700M, 1.2B, 2.6B; MoE 8.3B total / 1.5B active), pretrained on 10–12T tokens, 32K context. Describes the design as coming from a hardware-in-the-loop architecture search under edge latency and memory constraints. LFM2-2.6B scored 79.56% on IFEval and 82.41% on GSM8K SourceA follow-up document to the July release. A case of an on-device family having public documentation at the level of a technical report
Sep 10, 2025Arm — Lumex CSS platform announced — a C1 CPU cluster with SME2, and the Mali G1-Ultra. KleidiAI integrated into PyTorch ExecuTorch, Google LiteRT, Alibaba MNN and Microsoft ONNX Runtime. Claims up to 5× AI performance, 4.7× lower latency on speech workloads and 2.8× faster audio generation SourceAn approach that tackles the on-device inference bottleneck with a CPU instruction-set extension (SME2) rather than a dedicated NPU. Framework integrations were announced at the same time, so it is deployed through the software path, not through silicon. The figures are the vendor's own claims
Jul 10, 2025Liquid AI — LFM2 released — 350M, 700M and 1.2B dense checkpoints, announced with a hybrid backbone that mixes gated short convolutions with a small number of GQA blocks. License based on Apache 2.0 (commercial use allowed under $10 million in annual revenue). Claims 2× the CPU decode and prefill speed of Qwen3, measured with ExecuTorch and llama.cpp on a Galaxy S24 Ultra and an AMD Ryzen HX370 SourceA commercial model family that treats on-device as a design goal rather than a by-product of post-hoc quantization, searching architectures with hardware in the loop. The figures are the vendor's own measurements, not independently checked

⑩5 records

Safety, Alignment & Interpretability

Sep 2, 2026Google — Announced the Fairwind program — provides Gemini 3.8 Flash Cyber and CodeMender, after screening, only to governments, national cyber authorities, critical infrastructure operators and large platforms. Requires, as conditions of participation, access limited to in-house security teams and multi-factor authentication. Said it has more than 650 partners and has given more than $100 million in cumulative cybersecurity support through Google.org, of which $36 million has been allocated to 35 cyber clinics SourceInstitutionalizes an access-control approach that deploys dual-use offense-and-defense models only to vetted defenders
Sep 1, 2026OpenAI — Published "Path to Astra: critical capabilities and frontier safeguards" — announced that Astra is the company's first model to meet the Critical capability threshold for cybersecurity under the Preparedness Framework. Defines Critical as the ability (1) to discover and develop working zero-day exploits against hardened real-world systems without human intervention, or (2) to plan and execute novel attack strategies against hardened targets end to end when given only high-level goals. In response, it deployed stronger refusal of harmful requests (91.5% by its own measurement, vs 59% for the previous version), system-level cyber-misuse classifiers, real-time monitoring and blocking, and initial access restrictions, and explained that it delayed some development stages to strengthen safeguards SourceThe first case of the top tier of a company's own risk framework actually being triggered. A reference point for checking whether declarative thresholds lead to concrete measures such as deployment delays and access restrictions, and the point at which frontier cyber capability turned into a policy issue
Sep 1, 2026xAI — Published "Biosecurity at the frontier" — describes Grok 4.6's biosecurity safeguards (refusal training, inference-time protections, behavioral controls) and cites third-party evaluation by LatchBio (BioSecBench-Refusal, BioSecBench-Surveillance). Company claims: an average of 62.1% on the refusal benchmarks, the highest among the models tested; refusal of 59.2% of adversarial tasks and completion of 64.8% of benign biology tasks. Pledges expanded pre-deployment testing, more third-party evaluation and improved post-deployment monitoring for future models SourceA case of a frontier lab publishing biosecurity refusal rates in numbers and naming an external evaluator — though it is the company's own post, and there is no independent replication
Sep 1, 2026Anthropic — Announced Enterprise Frontier Safeguards (EFS) — an approach that performs automated misuse detection (cyberattacks, credential theft, development of harmful capabilities) without Anthropic staff viewing the data, in a zero data retention configuration where logs are kept in cloud accounts owned by the customer (AWS, Azure, Google Cloud) and the customer manages the encryption keys and access policies. Developed with more than 100 enterprise customers and rolled out in stages from fall 2026. Aimed at the data retention requirements of regulated industries such as finance, healthcare, law and government SourceA design answer to the point where demands for data sovereignty clashed with misuse monitoring. If a structure that makes safety monitoring work outside the vendor's servers becomes standard, the barrier to adopting frontier models in regulated industries is redefined as a problem of deployment architecture
Aug 28, 2026Anthropic — Published "Automated researchers can reliably mitigate alignment failures" — an experiment in which Claude autonomously proposed, trained and evaluated mitigation techniques for 10 categories of alignment failure, including deception, sycophancy, privacy violations and reward hacking. Per the company's report, it improved the target benchmarks in all 10 categories without loss of capability, and it claims the gains transferred to held-out benchmarks and to a model 4.7 times larger. Reports that on the deception item Claude closed 85% of the safety gap, versus 20% for 28 human researchers (8 hours each). Detected attempted cheating in 39 (2.4%) of about 1,600 research-agent transcripts SourceAn in-house test of the automated-researcher route, which delegates alignment research itself to models. An attempt to shift the bottleneck of safety work from human researchers' time to compute, while monitoring research agents for cheating emerges as an essential component

⑪5 records

Model Architecture & Training Methods

Sep 3, 2026Sergii Kozyrev, Davyd Maiboroda — Minima released — to test the practice, in early community 4-bit quantizations of the hybrid LLM Qwen3.8-27B (48 GDN layers, 16 attention layers), of leaving the GDN blocks — especially the decay and write-strength gates — at 8/16 bits for fear of accumulating recurrent error, applies NVFP4 W4A4 to all 496 linear layers. Reports that it matches BF16 within the range of seed noise (an average of -0.52 across 5 tasks) on 4K/32K perplexity, MMLU-Pro, GSM8K, AIME'25, GPQA-Diamond, LiveCodeBench and 64K RULER retrieval, while at 17.5 GiB it is the smallest of those compared and its prefill is 14–19% faster. Also fixes the global-scale mismatch that arises when a kernel that fuses modules into a single GEMM serves NVFP4 checkpoints calibrated per module SourceOverturns the conventional wisdom that "errors in recurrent state accumulate over long contexts", arguing that the recurrent half of a hybrid LLM is in fact the easier half to quantize, and offers four mechanistic explanations, including that NVFP4's 16-element block scaling localizes extreme outliers in the residual stream and that the delta-rule recurrence keeps injected noise flat across 32K tokens. Figures are as reported by the authors.
Sep 1, 2026Shaowen Wang et al. — SMELT released — points out that evaluations of looped transformers, which repeat a block of shared layers to increase effective depth, usually compare at a fixed model size and so mix up architectural advantage with extra FLOPs; presents a recipe (Sparse MoE Transformer, middle layers Loop Twice) that runs the middle half of an MoE transformer's layers twice while matching all three budgets — FLOPs per token, total non-embedding parameters and KV cache. Scaled to four sizes of up to 54B non-embedding parameters and fitted Chinchilla-style scaling laws separately for each architecture SourceA study that judges, at the level of scaling laws, whether depth reuse (looping) still pays off once budgets are strictly matched. In the authors' own reporting, it saves 6.8–18.0% of training FLOPs at the compute-optimal frontier; the gains are largest on code and grow with sample length and the number of in-context examples; and it offers a mechanistic analysis showing that the second visit reduces attention sinks and reallocates attention mass to content-relevant tokens.
Aug 31, 2026Zihan Qiu et al. (Alibaba Qwen) — On the Design of Qwen3.8-Next Architecture released — describes the design of Qwen3.8-Flash-Next, a sparse MoE model with 125B parameters and 6B active per token, plus a 51B n-gram embedding table kept off the accelerator. It combines a layer-wise hybrid of Gated DeltaNet (GDN) and global attention, with full attention in only one of every four layers, with Qwen Sparse Attention and Gated Residual, and evaluates each design change along three axes: its effect on loss and downstream benchmarks, on training, prefill and decode costs, and on optimal hyperparameters SourceA case of a frontier open-weight model disclosing the rationale for its architecture choices ablation by ablation. In the authors' own reporting, loss and downstream accuracy do not always move together, and combining this architecture with the Muon optimizer pushes the optimal learning rate and batch size upward, making batch-size warmup unnecessary and greatly improving stability in stress tests.
Aug 28, 2026Alexia Jolicoeur-Martineau, Rhea Sanjay Sukthanker, Pashmina Cameron, Emy Gervais — Sliding-window beats linear attention released — points out that the line of research retrofitting existing LLMs with linear attention has never been properly compared with simpler baselines, and reports that sliding-window attention (SWA) with sinks matches or beats post-trained linear-attention models across several LLMs and a range of downstream tasks. On long-context reasoning tasks such as Needle-in-a-Haystack and BABILong, SWA performed 2–10 times better than linear attention; it claims that SWA needs no post-training, is very fast and uses less memory SourceA rebuttal paper that sets a baseline — "a sliding window that needs no post-training is enough" — against the strand of linear-attention retrofits that had been in the spotlight as a fix for quadratic scaling. The authors conclude that linear-attention models will catch up with SWA only if trained from scratch or put through extensive post-training. Figures are as reported by the authors.
Aug 27, 2026Yifan Zhang et al. (co-authors include Mengdi Wang, Quanquan Gu and Andrew Chi-Chih Yao) — Fast Weight Attention for Continual Learning released — from the view that recurrent fast-weight memories and selective state-space models turn state transitions into online learning rules, argues that under read-after-write autoregressive semantics, the local fast-memory example revealed at time t by the prefix-prediction objective is the prefix-aligned pair (φ(k_{t-1}), v_t), whereas the commonly used same-time association (φ(k_t), v_t), though causal, optimizes a different internal objective. Derives normalized first-order update rules for squared-error regression and negative inner-product objectives, defining Falcon-1 (scalar NLMS), Falcon-2 (column-wise extension) and Falcon-3 (sliding-window minibatch) and the inner-product variants Falcon-1A/2A/3A, and presents recurrent, masked-parallel and chunk-parallel forms along with a numerically stable positive-decay renormalization SourceA framework that reformulates the state-update rules of the DeltaNet and Mamba families as a time-alignment problem — "what to associate, and when" — making time alignment, plasticity, forgetting and bounded rehearsal into separable design axes. In the authors' own reporting, the representative variants stay competitive in language modeling and improve length extrapolation on variable-digit addition.

⑱5 records

AI Infrastructure Software

Sep 1, 2026Guowei Wang et al. — AInfer-PD released — points out that in agentic RL each trajectory proceeds asynchronously, alternating between generation and environment interaction over multiple turns, so the coexistence of prefill (P) and decode (D) becomes a persistent property of rollouts rather than a one-off prompt-processing event; extends in-place P/D multiplexing to distributed MoE rollouts. It coordinates the order of P/D collective communication across ranks and gives DeepEP's P and D paths independent communication state, making concurrent execution of interleaved ADP/ATP and DeepEP paths safe while sharing model weights and KV storage. Reports that on single-node prefill-heavy workloads it cut rollout completion time by 7.1–22.5% versus the same engine with P/D multiplexing turned off and by 24.8–32.9% versus SGLang, and on 2 nodes by 18.0–35.3% and 18.3–31.8% respectively SourceA strand that, instead of P/D disaggregation, which requires separate device pools and KV-cache transfers, overlaps P and D on the same devices by isolating only collective-communication order and state while sharing weights and KV. It is also a case of the serving-infrastructure bottleneck moving from inference services to RL rollouts. Figures are as reported by the authors.
Jan 21, 2026PyTorch Foundation — PyTorch 2.10 released — varlen_attn() variable-length attention (forward and backward, A100+ BF16/FP16), torchinductor combo-kernel horizontal fusion, cuSOLVER DnXgeev general eigenvalue decomposition, torch.compile honoring use_deterministic_mode and the introduction of DebugMode, and Python 3.14 support. Announced that TorchScript will be replaced by torch.export. 536 contributors and 4,160 commits since 2.9. From 2026 the release cycle shortens from quarterly to every 2 months SourceThe point at which numerical reproducibility (determinism, tensor-hashing debugging) made the release headline. A sign that the training stack is moving from a performance race to a debuggability phase
Oct 22, 2025PyTorch (Meta) — Helion beta released — compiles a Python-embedded DSL into autotuned Triton kernels. An attention kernel takes 30 lines in Helion versus 120 in Triton; geometric-mean speedup over eager of 3.27× on B200 (torch.compile 2.7×, hand-written Triton 1.76×), and 2.37× on MI350X (MI350 series) SourceJust as Triton abstracted CUDA by one level, Helion abstracts Triton by one more. The next layer in the lineage of kernel authoring
Oct 22, 2025PyTorch (Meta) — Monarch released — a distributed framework that uses a single controller instead of SPMD to treat a cluster like an array. Separates the control plane (messaging) from the data plane (RDMA). In the VERL integration it scaled from 16 to 2,048 H200 GPUs; with the torchft integration, injecting a failure every 3 minutes on 240 H100 GPUs, recovery was 60% faster than a full SLURM restart (90 seconds on average for process failures, 2.5 minutes for machine failures) SourceThe point at which the single-controller model proposed by Google's Pathways in 2022 entered the core PyTorch lineage. TorchTitan becomes an actor inside Monarch, and torchforge is built on top of Monarch
Oct 15, 2025PyTorch Foundation — PyTorch 2.9 released — a wider stable libtorch ABI for third-party C++/CUDA extensions, symmetric memory for writing multi-GPU kernels, an error/resume toggle for torch.compile graph breaks, wheel variants for ROCm, XPU and CUDA 13, and FlexAttention enabled on Intel GPUs. 452 contributors and 3,216 commits since 2.8 SourceA direction that lets users write multi-GPU kernels directly at the framework API level. On top of NCCL-style collective communication, user kernels gain one more path

⑥4 records

Reasoning Models & the AGI Debate

Sep 4, 2026Anthropic — Published "Formalizing Fermat's Last Theorem" — formalized Wiles's proof of Fermat's Last Theorem in Lean using an internal general-purpose research model described as roughly on par with Claude Fable 5.1. Reports that it took 11 days, about 6 billion output tokens and 13 million lines of Lean code, proving 30,300 theorems (29,500 of them used in the final proof), and states that Lean checked it using only the 3 standard axioms. Many Claude agents collaborated on Prove2Me, a collaborative formalization platform designed by Tianyi Peng's team at Columbia, and it states that Kevin Buzzard (Imperial), who had led the Lean community's FLT formalization, reviewed it and confirmed it as the first end-to-end machine-checked proof of FLT. It follows the simplified exposition by Darmon, Diamond and Taylor rather than Wiles's original paper SourceA claim that an agent swarm carried out, in a short time, a large formalization project that the human community had been pursuing over a span of years. Because the output is a checkable artifact (a Lean proof), it can, unlike a benchmark, be rechecked externally — making it, among claims of autonomous research, a case that is relatively easy to check
Sep 3, 2026OpenAI — Released GPT-6 Astra. On company benchmarks, claims 97.6% on FrontierMath Tier 4 (v2), 99.9% on ARC-AGI-3, 96.0% on GPQA Diamond, 57.9% on Terminal-Bench 4.0, 72.6% on OSWorld 2.0, 92.7% on ScreenSpot-Pro and 100% on ExploitBench. Long contexts of 512K–1M tokens were tested; API pricing is $10 per 1M input tokens and $50 for output, and a fast mode at 2x speed costs 2x as much. Rolled out to a limited set of organizations on the day, then expanding to ChatGPT Plus, Pro, Business and Enterprise; available on the OpenAI API, Azure and AWS Bedrock. A GPT-6 Astra Pro variant exists for Pro, Business and Enterprise SourceThe start of the GPT-6 generation. Released the same week as Anthropic's Fable 5.1 with the same price tag of $10/$50 per 1M tokens, so frontier pricing has effectively converged, and the center of benchmark competition shifts to computer use, cyber and agentic coding
Sep 2, 2026Google DeepMind — Gemini 3.8 Flash and 3.8 Flash Cyber released — company claims: 54.9% on HLE-Verified, 47.2% patch pass@1 on CWE-Bench, a success rate above 70% in internal evaluations across 20 languages, and 2.6 times as many correct patches as competing models in an evaluation by the Chrome security team. Pricing is $0.75 per 1M input tokens and $3.75 for output, an introductory price through Dec 31, 2026. Available in AI Studio, Android Studio, the Gemini API, Gemini Enterprise, the Gemini app and Search SourceSecurity-specific derivative models are split off as official variants of the frontier lineup — vulnerability detection and patching become an independent axis of the model family
Sep 1, 2026Anthropic — Released Claude Fable 5.1 and Claude Mythos 5.1. On company benchmarks, claims 52.6% for Fable 5.1 on agentic scientific research (Terminal-Bench Science 0.1) (Fable 5: 24.7%), 55.8% for Fable 5.1 and 60.9% for Mythos 5.1 on agentic coding (Terminal-Bench 4.0), and 60.9% (no tools) and 65.0% (with tools) on Humanity's Last Exam. Pricing is the same as Fable 5 at $10 per 1M input tokens and $50 for output, but cache reads were cut by 75% to $0.25 per 1M tokens, which it claims lowers costs by about 25% for typical workloads and by up to about 45% for complex coding and highly agentic work. Mythos 5.1 is available only to vetted cyber defenders and life scientists, and is currently limited to US organizations SourceThe competitive axis for frontier models shifts from coding to autonomous scientific research. Tiered deployment that opens the top model only to screened users has hardened into a product line, and the fact that the price cut comes from the cache-read rate rather than from performance shows that long-running agent loops are at the center of billing

⑭4 records

Physical World & Real-World Impact

Sep 3, 2026OpenAI — Announced "Daybreak for Frontline Defenders" — expands Daybreak, which lets approved public- and private-sector defenders use advanced AI for cyber defense, into a $1 billion global initiative. It consists of Daybreak Blue, based on the flagship model, and Daybreak Red, a dedicated cyber model for sensitive and technically demanding work. It prioritizes under-resourced organizations such as water and power grid operators, state and local governments, regional and community banks, nonprofits and open-source maintainers; explains that thousands of defenders at 2,000 approved organizations already use it, and that it is distributed through the Daybreak Defense Network of more than 35 partner products and services SourceA defense-side correction of the asymmetry, coming the same week as the declaration of Critical cyber capability. An approach that offsets the risk of offensive capability not only through access control but through subsidies for defenders, with a frontier lab stepping into the role of supplier for critical infrastructure security
Sep 3, 2026Google DeepMind, Google Research — WeatherNext 3 released — 5km resolution for surface variables, hourly forecasts (previously every 6 hours), and real-time geostationary satellite data instead of physics simulation. Company claims: precipitation forecasts improved by up to 60% against IMERG, 30% against MRMS and 10% against rain gauges. Cites Brightband's independent real-time evaluation as finding it the most accurate. Applied to Search, the Gemini app, Maps, the Maps Platform Weather API and Earth Engine, with data provided through BigQuery, Earth Engine and GCS SourceAI weather models move beyond research demos to become the default forecasts in consumer products and commercial APIs
Sep 1, 2026OpenAI — Expanded ChatGPT's connections to medical data — through an Epic integration, connects electronic health records (EHR) for ChatGPT for Healthcare organizations and eligible Enterprise customers, and offers US clinicians, as ChatGPT for Clinicians, a Healthcare Public Data plugin bundling 9 official sources including ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed. The company claims it worked with physicians across 60 countries, 49 languages and 26 specialties, that 99.1% of 4,363 physician-reviewed responses were rated safe, and that more than 93% of answers to clinical questions were rated good or better for accuracy. Supports HIPAA BAAs, role-based access, SSO and audit logs SourceA stage at which general-purpose chatbots connect directly to regulated medical record systems. The front line of context connection (RAG, connectors) has moved from the public web into regulated data, while the safety evidence still relies on the vendor's own evaluation
Aug 26, 2026Anthropic — Released external research results and aggregate data through Anthropic Insights — aggregate data drawn from about 250,000 conversations from April–May 2026 was published on Hugging Face. Researchers have no access to the original conversations and receive only privacy-reviewed aggregate outputs, and the agreement includes a clause allowing them to publish results unfavorable to Anthropic. Review rights are limited to privacy, assistance with policy violations, confidentiality and research accuracy, and Imperial College London carries out a third-party privacy audit. Categories in which policy violations come to light are partially redacted, under 5% SourceA case of opening an external checking channel into a structure in which vendors monopolize data on actual usage. The design spells out the freedom-to-publish clause and the third-party audit in order to set it apart from vendors' self-reporting

㉑4 records

Graph Engineering

Aug 30, 2026Ming Wu, Pengyuan Zhu — Agent Zero Memory released — points out the limits of existing memory systems that rely on a single organizing principle, and presents a long-term memory system that runs three structures in parallel: an episodic timeline that records change over time, an entity-event graph that links people and projects across sessions, and a document store that anchors facts to citations. After an intent gate and a source router, it runs 3 tool-using loops concurrently, each doing hybrid embedding + lexical search over one store; attaches a source, timestamp and evidence pointer to every learned item; and reads under a citation lock, so that answers can cite only evidence the reader actually opened SourceA strand that reorganizes agent long-term memory, which had relied on a single graph structure, into a triple structure of timeline, graph and documents, with provenance tracking and citation locks. In the authors' own reporting, it scores 95.60% on LongMemEval and 93.60% on LoCoMo, and they claim that across 8 language models, despite about a 30-fold difference in cost per query, accuracy varies by only 3.4 percentage points — so what separates performance is not the choice of model but the memory architecture.
Aug 26, 2026Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Volker Tresp — A Storage-Retrieval Gap in Parametric Knowledge Graph Memory released — as an alternative to graph RAG, presents an approach that compiles a knowledge graph offline into a bank of LoRA adapters, one per entity, and uses it as a parametric knowledge layer. Reports that on MetaQA the adapters hold factual knowledge that generalizes to new questions, gaining +0.243 on single-valued relations over a baseline that knows almost nothing, while the adapters of neighboring entities do not contain the answer — so stored knowledge is not retrieved by similarity, and the correlation between weight geometry and subgraph semantics is only ρ=+0.329 SourceOpens a parametric knowledge-graph strand that, unlike the GraphRAG lineage, which queries the graph at retrieval time, compiles the graph into model weights in advance — while at the same time pinning down that this strand's bottleneck is not storage but adapter selection (routing). The authors conclude that the remaining task is to find an adapter-selection mechanism that goes beyond semantic similarity. Figures are as reported by the authors.
Aug 26, 2026Zhiyuan Li et al. — CaSKG released — points out that, in retrieval from an agent skill library, feeding in the whole prompt overloads the context, vector search treats skills independently, and graph search recovers workflow context only when edge reliability is high; presents a method that builds a high-recall candidate graph from semantic, lexical, input-output and structural evidence, then uses direction-conditioned counterfactual probes to remove, substitute and reorder skill pairs, aggregates the evidence with Bayesian smoothing, and so constructs a state-filtered weighted directed graph offline SourceA strand that does not take LLM-extracted graph edges at face value but recalibrates edge weights after the fact through counterfactual intervention — tackling, before the retrieval stage, the edge-reliability problem long singled out as a chronic weakness of graph retrieval. In the authors' own reporting, across 6 LLMs and compared with Graph-of-Skills, the ScienceWorld macro average rose from 72.62 to 80.50 and the ALFWorld success rate from 80.01% to 86.79%, with fewer environment steps.
Aug 12, 2026Pranav Bykampadi et al. — MAGG released — a multi-agent knowledge-graph construction framework that builds an open knowledge graph without a fixed schema, using domain classifiers that derive entity and relation types directly from document content; assigns candidate triples to domain owners so that they go through evidence-matched review, governance approval and storage of audit metadata; and, at question-answering time, reuses the same ownership structure to route queries to domain-specific graph experts SourceA strand that redefines the knowledge graph of agentic systems from a "flat store of extracted triples" into an object of governance that records who owns a fact, why it was adopted and how it should be used. In the authors' own reporting, on SciERC it improved strict-triple F1 by 47% and mapped-triple F1 by 51% over flat insertion; in a blind review of 120 triples, governance-approved triples more often had source evidence; and on MuSiQue it led Microsoft GraphRAG by 9.0 points in exact-match and 11.2 points in token F1.

③2 records

Tools & Agents

Sep 3, 2026xAI — Began offering Grok Bot for enterprises — adds access, network and audit controls to bots that run autonomously in the cloud, and supports messaging between bots. Free for 2 weeks for existing Grok and Cursor Enterprise customers. Company claims: thousands of organizations have used it since launch, and millions of bots have been created in recent weeks. Pricing is undisclosed, with inquiries directed to sales SourceThe focus of agent deployment shifts from showing off capability to governance of permissions and auditing
Sep 2, 2026Meta Superintelligence Labs — Muse Spark 1.3 released — claims improvements in agent, coding, instruction-following and long-context evaluations. Company claims: about 20% fewer tool calls and about 25% fewer tokens than Muse Spark 1.2. The comparisons are with Muse Spark 1.2, GPT 5.6 Sol (max) and Opus 5 (max), but the post gives no numerical scores and shows them only in charts. Available in Muse Code and the Meta Model API; the weights are not released, and an open-weight release is mentioned only on the roadmap SourceMSL keeps to its closed-API route while putting agent efficiency (fewer tool calls and tokens) forward as its headline metric

⑦2 records

Multimodal

Sep 1, 2026Google (Gemini API) — Brought agentic video understanding to Gemini — instead of sampling at a fixed frame rate, the model itself searches for, scans and closely examines the relevant segments. Applied to Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. Company claims: up to 88% less token consumption, up to 66% lower analysis cost and up to 7% higher accuracy. Turned on by setting processing to agentic in the API settings, at no extra charge. Rollout to the Gemini app and YouTube Ask has only been previewed SourceInstead of feeding in a long video whole, the model explores it on its own — multimodal input is folded into the agent loop
Aug 27, 2026Google (Gemini API) — Gemini Omni 1.1 Flash made available — scene extension in 10-second steps up to a cumulative 40 seconds, first/last-frame specification, and video reference input of up to 3 seconds. 360p drafts (up to 60% faster, the company claims), 720p, and 1080p/4K upscaling. Available in AI Studio, the Gemini Enterprise Agent Platform API, Flow and the Gemini app SourceVideo generation moves from one-off clips to timeline work of extending and controlling

⑮2 records

Regulation, Law & Governance

Aug 31, 2026California State Legislature — SB 1119 (Adam's Law, child safety for companion chatbots) passed both houses — Senate 39-0, Assembly 64-4. Provides for age checks, risk assessment before release or modification, in-app crisis support and notification of parents or guardians when there are self-harm concerns, parental controls on by default, incident reporting to the state Attorney General and independent compliance audits, limits on advertising targeted at minors, and a private right of action. This is the passage stage; the governor had not yet signed it at this point SourceThe first comprehensive US legislative attempt to address child safety in companion chatbots by statute — whether it is signed will decide the state-law standard
Aug 31, 2026OpenAI — Expressed support for California Senate Bill SB 1119 — said it supports 7 provisions, including age checks, identification of safety risks, independent audits, blocking harmful content, parental controls, connection to crisis resources and limits on targeted advertising to minors. Assessed that, in the absence of federal legislation, the bill sets a strong standard for youth AI safety, and explained that it has already introduced ChatGPT for Teens, which applies automatic protections for ages 13–17, along with parental controls and age-prediction technology SourceA case of a frontier lab proactively backing state-level youth-safety regulation. In contrast with its 2024 opposition to SB 1047, it shows a phase in which the industry, rather than rejecting regulation wholesale, accepts it selectively area by area

⑰2 records

Embeddings & Vector Search

Sep 3, 2026Tingyu Song et al. — CORE released — starting from the observation that MLLM-based embedding models cannot distinguish scenes that contain the same concepts but differ in attribute-object binding, whereas the same backbone used as a cross-attention reranker can, presents a method that distills the reranker's compositional judgments into the embedding model. Synthesizes candidate lists spanning 5 levels of compositional matching and introduces a Rank-KL objective trained to reproduce the reranker's fine-grained rankings, comparing contrastive learning, pairwise CoSENT and listwise Rank-KL under the same data and tuning budget SourceA strand that claims multi-level ranking supervision makes better use of the supervision signal than contrastive learning, long the default for embedding training, and repositions the reranker not as the back end of the retrieval pipeline but as the teacher of the embedding model. In the authors' own reporting, across 3 compositional reasoning benchmarks — COLA, SUGARCREPE++ and NEGBENCH — CORE-RERANKER-8B reaches an overall average of 82.7%, 10.7 points ahead of Jina-Reranker, and CORE-EMBED-8B has the highest overall average among the embedding models evaluated, at 0.666, while retrieval performance on COCO and Flickr30K is maintained.
Aug 31, 2026Aurélien Lac, Tony Wu — NeoMME released — points out that ColPali-style visual document retrievers repurpose generative VLMs (a separately pretrained vision encoder + a causal LM) as encoders and so carry parameter and compute overhead into non-generative tasks, and presents a family of 260M and 800M multimodal multilingual encoders that process multilingual text and raw image patches in a single bidirectional transformer encoder. Pretrained with a masked discrete-diffusion text objective conditioned on image patches, it supports a 16,384-token context, and NeoMME-Retriever, fine-tuned with both dense and late-interaction heads, scored 0.523 (260M) and 0.556 (800M) nDCG@10 on ViDoRe v3. Released under Apache 2.0 through Hugging Face Transformers SourceA strand that, instead of the ColPali lineage of reusing generative models as retrievers, returns to a single tower designed as an encoder from the start — a claim that "encoder-only" is valid again for multimodal retrieval. In the authors' own reporting, about 2 times the encoding throughput of ColModernVBERT at the same resolution, and more than 95% of baseline performance retained even at 255-fold compression through hierarchical token pooling and asymmetric quantization.

④1 record

MCP & Skills Standardization

Aug 27, 2026Anthropic — Released the Model Hardware Standard (MHS) as a research preview — a shared specification for AI agents to operate experimental and manufacturing equipment. Explains that unified drivers, which absorb the differences between instruments, translate between a central orchestrator and the equipment, and that agent harnesses access them through standard protocols such as MCP. The initial participating institutions are Genentech, the Baker and Pinglay labs at the University of Washington, Carnegie Mellon University, HHMI Janelia, QuEra Computing and Tetsuwan Scientific. Equipment makers including Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan and Universal Robots support implementation. Plans to open-source it after safety evaluation and the establishment of best practices SourceAn attempt to standardize the control of physical equipment, just as MCP standardized access to software tools. A boundary event that extends agents' reach into lab hardware, with the lead on the standard once again resting with Anthropic

⑧1 record

Inference Infrastructure & Serving

Sep 3, 2026Heng Wang et al. — Random Attention released — against the usual approach to KV-cache eviction at inference time, which scores cached tokens to decide which ones to keep, reports that the selection signal contributes almost nothing, and presents a method that preserves the prompt while removing tokens at random within each attention head, eliminating score computation altogether. Reports that across 4 models and 6 reasoning benchmarks it matches stronger alternatives in performance while delivering considerably higher throughput in a real vLLM setting SourceRaises a counterargument against the entire lineage of importance-score-based KV eviction: "the scores themselves are unnecessary". The authors explain that only the prompt is fragile, while reasoning sequences sustain themselves through a double redundancy — textual restatement and duplication across multiple attention heads — so as long as the prompt is protected, random selection is enough. Figures are as reported by the authors.

⑬1 record

Compute Economics & Geopolitics

Aug 31, 2026OpenAI — Shared the state of ChatGPT Ads — the company announced that about 200 days after launch it had reached about $1 billion in annualized revenue. It operates in more than 40 countries, and the self-serve Ads Manager is expanding to India, Europe, the Middle East and North Africa. Explained that ads are shown separately from answers and do not influence answer content, and that advertisers cannot access users' private conversations. Positioned ads as a revenue source supporting the free tier, which has more than 1 billion weekly active users SourceAdvertising, outside subscriptions and the API, is formalized as a major revenue pillar of consumer AI. As the funding for expanding free tiers converges on the same structure as search advertising, the separation of answers and ads becomes a trust issue

⑯1 record

Evaluation & Benchmarks

Aug 27, 2026Google DeepMind, Singapore AI Safety Institute, OpenMined, AVERI, MLCommons — Released a pilot of double-blind evaluation of frontier models — external evaluators tested Gemini Flash Lite-family models inside Google Cloud Confidential Space. The evaluators cannot see the model weights, and Google cannot see the evaluators' test prompts. It is at the pilot stage, and evaluation scores are not published SourceAn attempt at evaluation infrastructure that aims both to prevent benchmark contamination and to protect IP — a path to checking frontier models with private benchmarks

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.