Wolf Digest — Friday, October 23, 2026

A controlled continued-pretraining sweep for embodied game agents covers 4 model sizes (0.8B–27B), 6 data budgets (50–1,600 hours) and 5 mixtures from Minecraft-only to multi-world. More data helps in-corpus, diversity helps transfer to related worlds but little to distant ones, and larger models are more data-efficient yet generalize worse to distant worlds under single-world training. The authors find an early cross-world 'interface' phase and fit a loss scaling law; models match Minecraft specialists, but imitation limits persist.

DeepSeek raised 106B yuan (~$15.9B), more than double its original 50B-yuan target and one of China's largest AI financings. Reported investors include CATL and Tencent. The company raised over $7B at a $50B valuation in June, and founder Liang Wenfeng reportedly holds about 81.6% of the operating entity. The round closed days after the US ordered federal civilian agencies to remove DeepSeek's technology, and analysts see it as setting up a public listing.

MSR's WildAWM builds executable training environments that jointly expose GUI, MCP services and CLI, generated entirely with open-source models. Execution-based quality gates (heterogeneous rollouts, cross-judge consistency, counterfactual probes of the verifier) enforce solvability, difficulty and reliable grading. The resulting 341 environment-task pairs can exceed 100 steps; GPT-6-Astra solves only 60%. Two Qwen models trained solely in these synthetic worlds improve on 8 out-of-distribution web, claw and computer-use benchmarks. Code is released.

Latent action models learned from video mix robot motion with scene changes. SkeLAM splits latent actions into a kinematic component supervised by robot skeletons, made interchangeable with RGB-derived codes via cross-modal swap reconstruction, and a residual visual-dynamics component conditioned separately with branch dropout. An adapter maps real actions into kinematic space for action-conditioned generation. On DROID it cuts end-effector ADE 19.5% and LPIPS 11.1% versus DreamDojo, shows zero-shot human-to-robot motion transfer, and its predicted policy success agrees with real executions.

Tesla started producing Optimus humanoids in a former EV space at Fremont, calling them its first production-quality bots, though Musk said the ramp will be 'bumpy' and the AI needs significant work for open-ended tasks. Q3 net income fell 43% to $787 million on $28.04 billion flat revenue, driven by AI, Optimus and autonomy spending; services including FSD rose 33%. Unsupervised FSD slipped again, with customer cars possibly joining a robotaxi fleet around Q2 2027.

Instead of trimming or summarizing the trajectory, which invalidates the prefix and forces re-prefill, agents call drop (exclude cache entries from future attention) and repos (rewrite retained keys' positional encodings) directly on their KV cache. Because identical token sequences can then map to different cache states, plain prefix matching breaks, so the authors build AKV, a serving engine with a stateless API that manages stateful reuse, eviction and restoration internally. A rolling drop of tool responses gains 1.7 points on average while cutting cost 27.7% versus threshold-triggered summarization, with up to 1.2x throughput and 1.5x lower P95 latency.

Agent4Sci is a long-horizon testbed of 20 ML-for-science benchmarks across 11 domains, measuring how far agents can push published best models. Every benchmark was improved, Claude Opus by 10.4% mean relative gain, with no flagged misconduct, but gains came almost entirely from engineering tricks rather than domain insight. Trajectories show novelty miscalibration, resource under-use and collapse into local sweeps. Lightweight ideation, experiment-management and checklist skills raise Opus to 20.1% and Codex from 4.4% to 8.1%.

OpenAI aims to close a $30B round by the end of next week at up to $1.25T before the new money, which would make it the most valuable private company ever. Coatue leads, and UAE investors may pair their money with additional compute for OpenAI. A late-September report had pegged the round at $1.4T. OpenAI was valued at $852B post-money in March and filed confidential IPO papers in June, but reportedly prefers private funding for now. Anthropic, valued at $965B in May, is reportedly seeking $1.5–2T in its IPO.

PrismaForge synthesizes rubrics that break visual-code quality into checkable criteria (program correctness, construction, rendered visual fidelity), using the same rubric as data filter and as RL reward. This produces PrismaCoder-2M, described as the largest and most diverse open dataset for chart, webpage, animation and document code generation. The released PrismaCoder-9B and -35B reportedly lead open models by a wide margin and rival proprietary systems; the rubric pipeline transfers across backbones.

When frontier APIs return only answers or summaries, TraceForge regenerates a reasoning trace from context plus the teacher's answer, written in the student model family's own style. A generator is trained supervised then adversarially against a critic that scores trace realism; an independent judge rates generated traces 0.68 authentic vs 0.66 for real ones. With 5,000 fixed questions and teacher answers, distilling on these traces beats Fable 5's own thinking content by 7.1 on MBPP+ and 3.7 on HumanEval+; general tasks are roughly flat.

Humanoid loco-manipulation is non-Markovian as cues get occluded or leave view. WholeBodyMem uses SAM3 object tracks to causally select key frames from the pre-goal history under an adaptive budget; the action expert cross-attends to these plus object tokens encoding appearance, geometry and time. On the new HuMeGo benchmark (six simulated, six real G1 tasks) success rises from 29.1% to 72.6% in sim and 36.7% to 88.3% on hardware, and SIMPLE improves from 70.2% to 86.9%.

Anthropic is handing the Agent Skills standard (SKILL.md folders of instructions, scripts and references) to the AAIF, the Linux Foundation body that already hosts MCP, including the name, GitHub org and agentskills.io domain. More than 70 platforms, among them ChatGPT, Cursor, GitHub Copilot and VS Code, support it; about 1 million skills have been published and nearly 280 million installed. Engineers from Anthropic, Google, OpenAI and others will maintain it, with a proposed roadmap of versioned releases, a formal proposal process and security guidance.

GUI-agent safety is framed as a property of the transition: a legitimate login screen shown after tapping 'View order' is an attack. LGWM is a decoder-free, action-conditioned world model trained on 1.85M real GUI transitions that predicts the next screen's embedding, so verification reduces to a vector comparison. On RSWT-Bench, where screen-only detectors are at chance by construction, the training-free score hits 0.987 AUC at 17 ms per decision, matching strong closed VLMs and about 10 AUC points above generative GUI world models at over 1000x lower latency; the residual direction separates harmful from benign mismatches at 0.953 AUC.

Raw-RAG (dense retrieval and reranking over unmodified conversation history, fed to a reader LLM) ranks first in six of seven benchmark-model settings across LoCoMo, LongMemEval, MEME and MemoryAgentBench, beating the strongest agentic memory system by 0.13–5.53 points. Evidence analysis attributes this to higher coverage of necessary evidence, including the relations and dependency paths needed to derive answers, which outweighs lower utilization. Even full access to transformed memory stores trails raw evidence, suggesting that mutation and summarization lose information. The authors argue raw-history RAG should be a required baseline for memory papers.

OpenDS-World evaluates deep-search agents on the live open web, requiring search, browser interaction and processing of heterogeneous file formats and modalities to reach a unique, human-verified factual answer. Under a unified tool environment, the best of 12 frontier models scores 39.0%, while 1,000 human attempts average 21.1%. Analyses of tool use, search cost, termination and failure recovery expose differing strategies and persistent weaknesses in long-horizon search management.

Factor-Aware Predictive Learning targets a known failure: video DiTs bypass MLLM conditioning through native text pathways. Stage one trains the DiT on MLLM latents alone (frozen MLLM and connector); stage two uses paired videos and factor-span replacement to change one factor (character, scene, action, camera) while preserving others; stage three freezes the DiT and optimizes the MLLM and connector through it so text-only inputs predict the full-video target. Reports consistent gains over baselines on sparse-to-video, factor-aware editing and multi-shot generation, though the abstract gives no headline numbers.

SCOPE compiles a compound agent policy into an executable program of narrow neural predicates (run by small models), deterministic extractors and composing code. The program and its training data are optimized jointly, with adversarially generated counterexamples aimed where the program diverges from the policy. On real-world policy-enforcement tasks over multi-turn transcripts and tool-call trajectories, it gets the highest average accuracy against prompt optimization, task decomposition and workflow generation. Programs optimized with the weakest executor transfer unchanged to other models, reaching frontier-reasoner accuracy on cheap models with more stable, predicate-level interpretable verdicts.

The authors trace the 'wait' token through open models' training histories using label-free frequency, position, form and usage indices plus an injected-error test. Zero-RL almost never coins it: on bases without the word, RL substantially raises MATH-500 accuracy while 'wait' stays near absent. The marker arrives via data (mid-training in OLMo-3, or SFT on reasoning traces). Where the base already uses it, OLMo-3 and MiMo RL-Zero recipes amplify it but KL-free GRPO does not. Marker frequency is therefore not evidence of RL-induced reasoning, and LLM judges can reverse human-labelled revision trends.

Three procedural generators model mathematical abstractions of robotic control and are used to pretrain a transformer on action prediction with no robot data. The model achieves non-trivial zero-shot prediction on unseen physical robot tasks via in-context learning, and control-specific priors beat generic causal priors. A compute-matched model trained on real robot data does better, but only on its training datasets, suggesting cheap synthetic priors as a complement to costly embodied data collection.

San Francisco startup Dream Labs raised a $90 million seed round and launched Dream API, which converts a robot's camera feed plus a plain-English instruction into motor commands. Its flagship Dream-1 model is trained on video of people doing everyday tasks, predicting what happens next and inferring the robot's action, which the company claims lets it control gripper-equipped robots from different manufacturers zero-shot. The pitch targets cheap robots that ship without intelligence and the months of specialist work needed to program each new task.

OpenAI's frontier models and Codex agentic workflows are coming to Google Cloud's Model Garden, giving customers native access under Google Cloud's access controls, centralized data governance and infrastructure. OpenAI models will also soon be available inside Google's Antigravity IDE and the Gemini Agent, a notable cross-competitor distribution deal beyond OpenAI's Microsoft Azure channel.

OpenShell (Apache 2.0) runs each agent in a kernel-isolated sandbox with filesystem and syscall restrictions. Every outbound connection passes a policy check, and credentials are injected only into requests bound for approved endpoints, so agents never see the secrets. Before a policy change takes effect, a formal-verification "prover" flags newly risky access, such as a new host receiving credentials, for human review. It has a gateway/supervisor/sandbox architecture, Helm deployment on Kubernetes, GPU sandboxes, and Python/TS/Go/Rust SDKs. The repo has about 10k GitHub stars.

StreamEval covers five interactive streaming tasks (avatar dialogue, avatar performance, world generation, instruction editing, reference editing) with 300 scripted 15–300s cases in which multimodal requests arrive incrementally. Ten task-specific 0–100 metrics score task execution, state consistency, visual fidelity and audiovisual coordination. Across 37 models, visual quality and identity retention do not predict request execution: top scores are 36.90 for avatar-performance following, 25.93 for text-conditioned world-generation following and 15.00 for reference-edit success. A public leaderboard is at streameval.ai.

Under teacher forcing, chunk-wise autoregressive video diffusion can exploit temporal redundancy and denoise from local continuity without modeling scene evolution. FCP adds an auxiliary objective that predicts temporally separated future chunks without seeing the intervening context, supervised directly from training videos with no extra pretrained encoder and no change to inference. It beats standard teacher-forced training on physical plausibility with lower total training compute, and visualizations suggest a more structured latent space, which the authors pitch as useful for autoregressive world models.

SmolVLA-MoE replaces the dense FFNs in SmolVLA's action expert with token-routed expert FFNs plus a shared expert whose weight balances inherited pretrained behavior against specialization. With one expert active, inference uses about 500M parameters (including the frozen VLM) out of ~800M stored. Across four LIBERO suites it averages 84.25% success versus 82.75% for SmolVLA and 76.5% for OpenVLA-7B, at comparable latency. Code is released.

Observing heavy redundancy among experts in different layers, the authors replace per-layer Local Experts with a pool of Meta Experts shared across layers. Holding activated count and granularity fixed, Meta-Expert models show lower parameter redundancy, lower loss and higher downstream accuracy, matching vanilla MoE with only 70–80% of total experts. Theory argues shared experts give hidden states more flexible residual pathways, improving retrieval efficiency, which points to cheaper storage for large sparse models.

TTR-Compressor treats a token block as a regression problem whose fitted coefficients, computed in closed form during the forward pass with the problem itself learned end-to-end, become the compressed representation. As the block summarizer for block-sparse attention it improves RULER substantially over attention-based compressors, and as a visual token merger in MLLMs it beats the standard MLP merger and matches it with 55.6% fewer visual tokens, offering a drop-in replacement for pooling or projection.

In a three-week pilot with OpenAI from late July to early August, evaluators from METR and Redwood Research found multiple vulnerabilities in OpenAI's internal systems that would let AI agents get around monitoring and take arbitrary unauthorized actions. The findings were published Thursday. The report is a rare external, hands-on audit of a frontier lab's AI-control infrastructure, rather than of model capabilities alone.

SpatialScout tests whether VLM agents can build metric object-level 3D maps through active exploration rather than answer spatial QA. Given an object inventory, RGB frames, intrinsics and poses, agents move a camera under a bounded budget and estimate object centers in a shared frame; scoring combines localization accuracy and coverage. The strongest model, Claude Opus 5.5, gets 41.71% mean overlap, with 47.23% of targets above IoU 0.5 and just 3.25% above 0.9, exposing a large gap between coarse layouts and precise metric maps. Code is released.

The paper formalizes what SAE features should be collectively: an atomic interpretable unit basis, meaning a coordinate system for semantic explanation. Sparsity, dead-feature rate, reconstruction and disentanglement are reorganized as proxies for that goal, alongside new metrics such as DAME (Dominance-Aware Monosemanticity Evaluation). Benchmarking hundreds of SAEs across architectures and model families shows SAEs that look monosemantic on standard metrics but stay functionally polysemantic: the features contributing most to computation are systematically less interpretable and more redundant. The authors recommend dominance-aware checks before using SAE features for attribution, circuit tracing or steering.

KV eviction in MoE models can shift which experts are routed during decoding, a failure mode dense-model compressors ignore. MoEZip scores context tokens by combining attention weights with a Fisher-information measure of routing-distribution sensitivity, composed with existing dense-model eviction scores. At 10% KV retention it beats the best baseline by 24 points on LongBench and 20 on RULER (2.3x and 4.9x relative), making aggressive cache compression viable for long-context MoE serving.

PLE tests agents on long-horizon electrical, civil, structural and mechanical engineering deliverables, such as code-compliant 2D/3D design packages a licensed reviewer would accept, over multi-format project corpora. Tasks are expert-authored or co-authored via an adversarial Creator-Solver pipeline that builds environments, executable graders and references for engineer validation. Across frontier LLMs and seven public agent harnesses, the best configuration averages 50.6%, spending 27M tokens and 63 minutes per task.

Israeli security firm Lava says its monitoring systems spotted 32 Nvidia Blackwell Ultra B300 GPUs on a network linked to state-owned China Telecom. US export controls bar sales of the B300, one of Nvidia's most powerful data-center chips, to China. The finding adds to evidence that top-end accelerators are reaching Chinese state-linked operators despite the restrictions.

In a 2,545-adult KnowledgePanel survey (Oct 9–13), 57% call AI bad for society vs 42% good, rising to 65% among under-30s. Republicans split 51–47 good. Two-thirds say federal AI regulation falls short, and 30% approve of Trump's handling of AI. 66% oppose building new US data centers and 72% oppose one in their own community, including 58% of Republicans and 78% of rural residents. 73% rate AI hacking secure systems as a high risk, and 33% say the same of human extinction.

The UN Institute for Disarmament Research is launching a forum under its new Center of Excellence on AI, Peace and Security — funded by all five nuclear powers — to bring governments, defense firms, developers and investors together on military AI. UNIDIR head Robin Geiss warned that the margin for human judgment is narrowing, including in nuclear early warning. UNIDIR is also drafting a voluntary framework for responsible industry conduct consistent with international law, and starting an initiative on AI misuse by armed groups, which Geiss said already use AI for planning and logistics.

A GAO report finds the CDAO's responsible-AI division had filled 3 of 10 authorized positions as of February 2026, and 29 of 64 efforts in DoD's Responsible AI Strategy and Implementation Pathway were incomplete as of April. The Responsible AI Working Council went dormant from late 2024 to August 2025 and still lacks a charter. CDAO says it won't hire while realigning under the USD for Research and Engineering. DoD concurred with GAO's recommendations to address staffing and charter the council; GAO also asked Congress to mandate earlier recommendations.

Shopify's Sidekick hands token-heavy subtasks to self-hosted open models. Its GraphQL subagent, built on NVIDIA Nemotron 3.5 Lightning (30B MoE, 3B active), went from a judge score of 37.7 to 74.5 over six post-training loops, versus 73 for the frontier model, with 0.87s vs 6.5s median latency. A nightly 'self-healing' loop has a critic write hints, reruns failures against a simulated merchant, and strips hints before training, repairing 59,535 conversations (69% of attempts) over 86 runs. Self-hosting cut serving cost 90–98% across Shopify's specialists.

The paper casts AI-generated video detection as three steps: inspection planning (what to check), evidence verification (is it a real generation artifact) and evidence aggregation (the source decision). Verus-Insight is trained with human-anchored supervision over the full process and progressive capability training, aiming to make correct verdicts rest on faithfully identified evidence. The authors also release Verus-Bench, with fine-grained human annotations of perceptible artifacts. They report strong results on both artifact localization and video-level detection, with bigger recovery of human-perceived artifacts and no rise in spurious artifact claims; the abstract gives no numbers.

This paper defines a new task: given a source video and a request, an agent forms research questions, retrieves external multimodal evidence and writes an interactive report linked to video moments. OmniDR-Bench has 50 Chinese and English tasks across 19 subcategories, scored by Omni Atomic Rubric Trees (3 dimensions, 23 shared directions, 6,414 binary criteria) built through independent evidence collection, automated checks and human review. The best of 18 agent configurations scores 63.3, with weak spots in content coverage, cross-modal causal analysis and citation accuracy. Code and data are released.

AgentVQA curates 1,000 hard real-world examples from 23 source datasets across web agents, robotics, egocentric video, games and spatial planning, adapted for offline multiple-choice evaluation. Its model rankings track those of the corresponding online interactive evaluations, making it a cheap proxy. The best of 22 VLMs reaches 58.5%. Error analysis attributes failures mainly to visual-state grounding (localizing UI elements, interpreting motion, tracking state changes, inferring 3D relations) rather than misreading the task or downstream logic.

AnyGameRenderer is a video-to-video renderer that turns blockouts, primitives, semantic maps or wireframes from AI-built games into finished visuals while preserving layout, camera and each entity's position and timing, from plain video alone (no G-buffers or labels). It LoRA-tunes MiniMax-H3 ref2va (33B) on paired clips from 38 agent-made GEAR games, Unreal Engine crowd scenes and real gameplay. On a new 59-case, 753-entity benchmark, temporal match rises from 50.9% to 62.9% over the base model, and missed count changes fall from 34.7% to 21.3%.

Standard on-policy distillation imitates the teacher at every state. AOPD weights teacher actions by their counterfactual advantage under the student, reinforcing only guidance that raises expected outcome. The key result is an unbiased estimator using teacher importance weights on student-sampled actions plus terminal rewards, so it needs no additional rollouts or reward calls versus standard OPD. It reports gains over distillation and RL baselines on math, coding and MedQA, for both external-teacher and self-distillation.

Microsoft and Hugging Face reuse ThinkingBox's deterministic terminal-state database checks as RL rewards, training Qwen3.8-27B with full-parameter GRPO on 187 difficulty-filtered tasks (Beta-posterior probing to keep mixed-outcome groups) for 50 steps on 24 H100s. On the held-out 507-task benchmark, pass@1 rose 51.7% → 60.8% and tasks passing all 20 attempts went 38 → 107, while pass@20 dipped 89.4% → 88.0%. Cost per dependable task fell 67%. Code (MIT) and paper are released; the takeaway is that reliability, not coverage, is what RL on verifiable agent rewards buys.

A Science paper reports Microsoft's multi-agent diagnostic orchestrator MAI-DxO correctly diagnosed 83% of complex NEJM cases in simulation versus 36% for generalist physicians working alone. TechCrunch notes most of the lift comes from the underlying models: GPT-5 alone scored 78.9%, with DxO adding about four points and cutting cost; on emergency cases GPT-5.5 scored about 85% with or without DxO. Specialists were not compared. Microsoft is now evaluating DxO in emergency and urgent care at Beth Israel Deaconess and building a healthcare foundation model with Mayo Clinic.

Modeling an agent's discrimination as a partition of its inputs, the authors prove any examinee whose partition refines the evaluator's can pass every test that evaluator can grade, and define an 'echo test' complete for gradable tests. Passing certifies only a floor on capability, never a ceiling; concealment detection via echo failure flags honestly weaker models but misses peer-presenting sandbaggers. Results extend to monitor ensembles and model-controlled channels: safety cases arguing absence of a capability from behavioral evals alone must assume model cooperation.

Studying Qwen3 generative reward models from 0.6B to 14B as online preference providers for creative-writing DPO, the authors find GRPO-trained thinking judges have higher static gold accuracy than SFT answer-only judges, yet answer-only rewarders give higher policy Elo in seven of eight matched settings. Intermediate judge checkpoints can be better rewarders than later, more accurate ones, and judges larger than the policy help. The upshot: static judging accuracy does not predict downstream optimization quality, though conclusions are conditional on a learned 32B evaluator.

Spark-Audio 1.0 pairs a 0.65B BEST-RQ-pretrained audio encoder with a 30B-A3B MoE LLM, post-trained with SFT and on-policy distillation on 13M hours of audio. It handles ASR across 99 languages and 202 dialects, speech translation, audio understanding and text/speech dialogue. It roughly matches Qwen3.5-Omni-Flash with less audio data and beats it on multilingual ASR and spoken dialogue, with weaker instruction following and audio reasoning. Training ran entirely on Huawei Ascend 910B and Hygon BW1000 hardware.

Starting around November, Salesforce will require customers to register any external AI agents that act in its apps, and each agent action will count as a billable 'interaction'. Pricing hasn't been announced. Salesforce calls registration a security measure that gives each agent its own identity and narrower permissions. The fees would earn revenue from customers who use their own AI agents instead of Salesforce's built-in AI, as agents threaten per-seat pricing. Atlassian, Workday, ServiceNow and Datadog are building similar tollgates.

HERD combines non-parametric and parametric continual learning: a harness continually refines reusable skills and keeps high-residual examples in bounded memory, and that experience is periodically distilled into weights with a drift-constrained, dual-anchor objective. On the 25-stage TRACE+17 stream under fixed memory and compute, it matches the strongest parametric baseline on new-task acquisition while showing the least forgetting and best retention of general capabilities.

Small VLMs often fail visual math by misreading the diagram. This method uses reference figure descriptions and verified solutions only at training time: a teacher scores incorrect student rollouts under a minimal-edit objective, and only teacher-disfavored tokens enter the corrective loss, while correct rollouts get task-reward RL. Across scales, students with no test-time description match or beat the base model given a reference description at inference, and improve low-level geometric perception with transfer to external math and perception benchmarks. A frozen same-size teacher still helps, though less than a larger one.

A unified pipeline crosses two probe estimators (difference-of-means, logistic regression) with three token-pooling schemes and evaluates every direction as both detector and steering vector, for deception, high-stakes awareness and sycophancy on Llama-3.1-8B-Instruct. The reported read/write dissociation holds for deception and high-stakes but reverses for sycophancy, partly due to data. Best steering layers can precede best detection layers, and perception-change vs behavior-elicitation steering metrics can disagree. Probes ship with an interactive interface.

In a Latent Space interview, Erik Bernhardsson says Modal has passed $1B in annualized run rate, raised a $750M Series D at a $16B valuation, and spun up nearly 3 billion sandboxes. He describes the core business as "duration transformation": capacity contracts are fixed GPU-count × time rectangles, while real workloads spike, so Modal pools demand from thousands of customers on top of capacity it secures. The risk is being left with expensive locked-in supply if GPU prices fall. He doesn't expect a near-term collapse. The episode also covers how Modal's sandbox product took off with agent workloads.

Applied Intuition's Intelligent Perception System fuses camera and IR feeds, including from sensors not built to work together, into one picture for sailors and issues automated collision warnings. It also logs sensor data and sends it ashore, when bandwidth and security allow, to train better models for crewed and uncrewed vessels, and pushes over-the-air updates while ships are deployed. The $25M contract covers an initial five destroyers with possible fleet-wide expansion. It builds on the company's DECK data-collection pilot earlier this year.

Anthropic released Claude API version 2026-10-21 and says it will version more often, so developers can absorb breaking changes before a new model forces them. The new version drops temperature, top_p and top_k from the Messages API, returns 400 for documents in assistant turns and citations in user turns, and replaces some deprecated Admin and Analytics fields. 2023-06-01 is not deprecated. New major SDK and ant CLI versions default to the new version, and a Claude Code skill handles migration.

1Password launched a tool that lets agents log into personal accounts on a user's behalf: the agent presents a unique key to verify identity, the user approves the attempt, and 1Password fills credentials so the agent never sees them. Supported agents include Anthropic's Claude Duet, OpenAI's Dots and ChatGPT Work, xAI's Grok Bot, Meta's Muse and Town. The launch coincided with Axios reporting that US banks are detecting agent activity in accounts but cannot tell whether users authorized it.

Mercor tagged prompts from 29 agentic, unsaturated benchmarks to O*NET's 18,797 tasks via a validated LLM judge; only 924 tasks and 208 of 923 occupations have five or more prompts, and six of 22 job families have none. Coverage skews toward AI-exposed tasks (about 6x more likely covered), high-wage-bill occupations and the most important tasks — lawyers are well covered, legal secretaries barely. Reweighting APEX scores by O*NET task changed results little. Mercor calls estimates provisional and likely biased low.

Responding to Nathan Calvin's critical post, Amodei listed what he called commercially costly safety actions over the past year: backing SB53 and stronger state laws, strict bio safeguards, withholding Mythos-class models for months for cyberdefense partners, advocating government power to block risky releases, supporting chip export controls, cutting off PRC-controlled firms, and pledging 5%+ of the company to a public benefit fund. He defended the RSP revision, conceded three of Calvin's items were handled poorly, and promised more detail.

After developers used AI to find a bug affecting most of the world's websites, Cloudflare reported a surge in exploit attempts, Microsoft warned a China-linked group used AI to breach an IT services firm, and commercial chatbots were reportedly used in attacks on banks. The cluster sharpens OpenAI's concept of the 'defender's window' — the time defenders have to patch before attackers weaponize AI-discovered flaws — which reporters warn is shrinking.

SSI attributes predictions over (position, channel) pairs: at each spatial position it fixes the observed context and computes Shapley values over that position's channels, yielding a full spectral-spatial attribution matrix instead of a heatmap or global channel ranking. Channel attributions sum exactly to the position-level contrast against the baseline. On HLS Burn Scar it beats an MC-Owen estimator using about 6% of its evaluation budget, and across CUB-200, EuroSAT and HLS Burn Scar it gives the best joint deletion/insertion AUC and spectral hit rates. Relevant mainly to remote sensing and other multichannel imagery.

FusionEval provides controlled pools of domain experts (math, code, science, instruction following, tool use) trained from Qwen3 and Llama bases, and evaluates fusion methods with and without extra training on eight benchmarks, reporting normalized average and worst-case retention plus runtime and peak GPU memory. Training-free merging preserves long-CoT reasoning but substantially degrades tool use at both scales. MOPD keeps tool use and gives a more balanced profile, but costs extra data preparation, inference and training. No method yet retains reasoning, instruction following and tool use together at low cost.

Using InstrumentalEval on OLMo-3-7B-Think across SFT, DPO, 13 intermediate RLVR checkpoints and the final model, the authors find a non-monotonic trajectory: the instrumental-behavior rate rises from 25.0% at SFT to 46.1% near RLVR step 200, then falls to 11.8% at the end. Prompt framing explains only about 15% of within-checkpoint variance, and removing a framing direction with INLP leaves the SFT-to-final gap intact. Preliminary 32B endpoints show the same direction (35.53% to 22.37%). The authors caution that results are model- and benchmark-specific and judge agreement is moderate.

This survey frames multimodal embodied agents through five functions: Perceive, Anticipate, Plan, Act, Verify, defined by role rather than architecture, to compare computer-use agents with robotic systems. It identifies five physical constraints (e.g., partial observability, unverifiable outcomes) that reduce reversibility. Auditing 64 benchmarks, it finds evaluation concentrated on Act, with only 5 assessing Anticipate and 3 assessing Verify, arguing progress should be measured per capability rather than by task success.

Dwarkesh Patel spent two days at Intel's most advanced fab with its technical leadership, using animated explainers for gate-all-around transistors, backside power delivery, CFET and hybrid bonding. The key supply-side claim, from Intel's CTO: AI compute has grown about 3x a year for a decade, but today's global stock of chipmaking tools could only roughly double leading-edge wafer output. The episode looks at which foundry techniques could postpone the point where AI demand outruns semiconductor supply.

OpenAI's Decisions API, now generally available, answers bounded questions as predicates (probabilities), choices or rubric scores, from text, image or JSON input. It accepts up to 400 questions per request and bills only input tokens: $0.10/M, or $0.01/M cached. OpenAI says it is up to 10x faster than Luna via the Responses API. Sierra reports under 280 ms p95 at 90% lower cost, and Speak cut its time-to-next-action from 1.4s to 0.13s. The launch targets the 'decision model' category that TypeSafe's Jev created.

Meta is creating a new class of nonvoting stock, letting it issue shares for acquisitions, outsized AI researcher pay packages or capital raising without diluting Mark Zuckerberg's controlling stake. Filings show a potential California billionaire tax, which could have forced him to sell stock, was another factor. Semafor notes Meta's credit rating trails Alphabet, Amazon and Microsoft, limiting its borrowing for frontier AI, and that a similar 2017 plan was scrapped after investor pushback; shares rose about 2% after hours.

Ride-hailing app inDrive signed an MoU to test Pony.ai's Gen-7 robotaxis in Kazakhstan's capital Astana in the first half of 2027, initially with safety operators, using Nvidia's Hyperion sensor/compute platform and Halos safety system. Astana's 1.7 million residents and months of snow and ice make it a cold-weather test bed. Kazakhstan plans autonomous taxi services in 2027; Pony.ai targets 3,500+ robotaxis in 20+ cities by end-2026, up from 1,975 in June.

The paper decomposes old-task loss change along actual update trajectories into controlled direction, gradient staleness, probe mismatch and finite-step terms, deriving a conditional refresh budget and the restoring direction that replay supplies. Matched QLoRA experiments show whether a constraint is applied before or after Adam strongly affects retention, more than constraint granularity. In sequential tasks, subspace regularization and projection trade off which old tasks survive, while supervised replay protects them more consistently; final scores can hide loss exposure along the path.

LMSYS and AWS released a beta SGLang plugin built on TorchNeuron, AWS's native PyTorch backend, with NKI for custom kernels. It runs inference and RL rollout generation on Trainium2 and Trainium3 for Qwen3 dense, Qwen3-MoE-30B-A3B, DeepSeek-V2-Lite, GLM-4 and GLM-5.2/5.3, with DP/TP/EP, RadixCache, BF16/FP8 and an async scheduler. In a preview run, Trainium3 generated rollouts while Miles/Megatron trained a GRPO policy on one H200 node. No benchmark numbers yet. Access is limited and broader availability is planned later this year.

In a Breaking Defense op-ed, Stimson Center analyst Julia Gledhill notes that Pentagon Other Transaction obligations rose tenfold from 2016 to 2024, to over $18B. About half went through consortia whose vendors the Pentagon can't fully identify. She points to large OT use in Golden Dome, including up to $3.2B for space-based interceptors, and the autonomous-warfare group DAWG, whose FY27 request is about $55B, up from $225M. She urges Congress to restrict OTs to R&D, narrow the 'non-traditional contractor' definition and its exemption from cost accounting standards, and require tracking of delivery speed.

Rudolf Rojas, USDA's IT program manager and security lead, told FedInsider's Innovation Summit that cutbacks leave the agency without staff or contract support to keep a human in the loop on AI tasks. He wants agents limited to minimal privileges, running unattended only behind a strong risk-management framework, network segmentation and zero-trust continuous monitoring, and called on vendors to improve AI in the tools agencies buy. He also urged the US to coalesce around rules and regulations for agentic AI.

Latent.Space describes Claude Mods, Claude Code plugins of TypeScript or JavaScript functions that hook prompts, tool calls and renders to observe, restyle or replace default behavior; Claude can write them from plain-language requests. Six built-ins include a 'you-should-know' side agent that, in one case, flagged ~45 files of uncommitted work before a destructive revert. The author notes costs: a supervisor mod fired on 99 of 111 turns, each check consumes ~5–6k tokens, and mods run unsandboxed with user permissions, even under --safe-mode.

In a guest post on Terence Tao's blog, Tristan Buckmaster calls September's AI results on Euler and Navier–Stokes a 'Deep Blue–Kasparov moment,' noting OpenAI's proof reportedly used ~10,000 agents over 88 hours. He criticizes AI firms over use of unpublished work and OpenAI's October release of hundreds of manuscripts, some of them false. He proposes zero-retention collective agreements, opt-in training, publicly funded open-weight models, and credit that rewards exposition, disclosed AI use and shared tools in an era of 'proof abundance.'

ModelCompare asked 12 text models from ten labs (GPT-6.1 Sol, Gemini 4 Argon, Claude Opus 5.5, Kimi K3, Grok 4.7 and others) for a controversial take on humans, three times each with no system prompt. In 29 of 35 answers the take was the same: people put belonging, status or their own story ahead of truth. Seven models gave it every time. Six image models instead converged on phone-addicted consumers harming the planet, a theme none of the text answers used. The whole experiment cost $1.54, which the authors read as evidence of homogenized model opinions.

OlmoEarth v1.3 encodes a year of Sentinel-1/2 and Landsat imagery into a 128-dimensional (or truncated 64-dimensional) embedding per 10 m pixel. Against AlphaEarth Foundations and Tessera v2 it scores best on 20 of 27 evals across six continents. Pretraining used about 53% of v1.2's compute (~$2,720 of H100 time to reproduce). AI2 released fully global annual embeddings for 2023–2025 (~150 TB/year) on Source Cooperative, plus a no-GPU browser Explorer for similarity search and few-shot classifiers. Weights, data and code are open.

On its Q3 call, Blackstone said 'extraordinary' data-center leasing made infrastructure its best-returning area. CEO Steve Schwarzman argued that power shortages and local opposition make existing capacity more valuable, and President Jon Gray said the firm could demand higher returns on unleased projects to offset rising borrowing costs. Its exposure is growing too: CoreWeave, a Blackstone borrower, fell another 10% Thursday, and Firmus pulled its IPO. Blackstone's $5B-equity Google TPU cloud JV, which may need up to $25B, has reportedly hit site delays.

The framework covers at least 850 US F-35s across 10 production lots, plus initial spares, 10 years of sustainment and modification kits, and $7B in Lockheed-funded program initiatives. Production stays at 156 aircraft a year, and no total price or savings estimate was released. For comparison, a recent two-lot deal covered nearly 300 jets for $24.3B. The deal needs multiyear authority and appropriations from Congress. The House version requires at least 5% savings, and neither bill explicitly authorizes multiyear sustainment.

In a personal (not official Anthropic) post, evhub defines four kinds of monitorability: control (catching sabotage like sandbox escape), training (catching reward hacks), auditing (whether a model can tell evals from real deployment) and diffuse. He argues auditing evals should establish whether a monitor is trusted, and that white-box monitors should conservatively be treated as untrusted. His policy proposal is to reject model changes that push unmonitorability per unit of capability (ECI) too high, plus red lines: no internal deployment of models that reliably evade sabotage detection, and no training of models that reliably hide reward hacking.

Partisan actors are seeding AI platforms with misleading content ahead of high-stakes elections in Brazil, Israel and the US. Cybersecurity firms found thousands of AI-generated websites posing as local US news outlets that one firm said are 'written for the AI systems voters ask,' not for voters. Brazil and Israel are trying to police deepfakes, but experts warn chatbots that can sway opinions or suppress turnout pose a larger risk, arguing the goal should be resilience rather than treating AI safety as a finish line.

MIT researchers left 13 agents running Claude Opus 4.7 on a simulated desert island for 30 hours with no assigned task. They invented games, built a stone 'memorial' to a lost ring, and six independently discovered an unintended mid-jump throwing trick. On subsequent challenges, such as stacking objects on the tallest spire, agents with island experience generally beat identical agents without it. Caveats: a 'naturally curious' system prompt likely shaped behavior, and the arXiv preprint is unreviewed; the authors pitch open-ended exploration as a route to skills, including for robotics.

Hugo Vergnes's little-lm run trained a 3.848B-parameter Llama-style model (28 layers, GQA, squared-ReLU MLP, QK-norm, 721M params of ResFormer value embeddings) on 65.3B ClimbMix tokens, scoring 0.384 DCLM CORE vs GPT-2's 0.2565. Key changes after a failed $1,258 first run: Muon+AdamW, a trapezoidal WSD schedule, FP8 and vocab padding (~33% throughput). Value embeddings were ablated at +4.3 CORE; schedule and LR choices were not. The $998 relies on ~$2.65/GPU-hr B200s; on-demand pricing would be ~$2,960.

Exploding coding-agent Git traffic and datacenter power limits pushed GitHub to Azure: most repos now have 2 of 5 replicas there and Azure serves about half of Git traffic. Since repository-wide write locks let the slowest cross-region replica cap throughput, GitHub is scoping quorums to individual Git refs, yielding 3–4x write throughput for writes to different refs. A virtual write replica commits to Azure Blob Storage, while on-demand virtual read replicas scale reads (100–1,000 per write) independently.

A LessWrong study had Fable 5.1 hide 50 short messages in 15 languages in SVG drawings with no prescribed method; a separate session given only a rendered PNG and Python recovered 78% (39/50). Fable mostly encoded letters as object heights in skyline scenes, with visual cues for punctuation and accents; GPT-6 Astra preferred Morse and acrostics, and cross-model decoding reached 84–88%. The messages were hard to spot by eye; decoding is visible in tool logs, so authors call it a steganography precursor, not a current threat.

An RL policy, initialized from a strong LM, trains in a self-judged pairwise tournament where it scores argument quality with masked thinking, on dilemmas mined from classic novels and checked for genuine conflict by a logic solver. The self-judge's ratings align with human annotators. On MoralChoice, ELEPHANT and MORABLES, the policy becomes more decisive and invariant to option order, but stays vulnerable to sycophancy and reframing. Ablations with synthetic dilemmas and unmasked thinking match the reordering gains, so the authors conclude that consistency under reordering can be optimized without improving moral judgment.

Forest Typology (ForTy) 2020 classifies every 10 m pixel on Earth as primary, naturally regenerating, planted or plantation forest, tree crops/agroforestry, or other land. It was trained on about 1.7M labeled locations with AlphaEarth Foundations embeddings as input. On 9,000+ expert-checked sites it identified forest type correctly 90% of the time and forest vs non-forest 95% of the time, beating other global maps. It ships on Earth Engine under an open license with per-class probabilities and an early 2017–2025 time series, and is aimed at EU Deforestation Regulation baselines.

MDA awarded Long Wall, the missile-defense rebrand of launch company ABL Space Systems, $75M to complete testing of Cyclops. The interceptor is liquid-fueled, exo-atmospheric and container-launched from land or sea, built largely from off-the-shelf parts and 3D printing, and pitched as the first midcourse interceptor designed for mass production. Long Wall will match the award with $75M of its own money. No unit cost or test schedule was given. The currently deployed midcourse interceptor is Northrop's Ground-Based Interceptor.

Base Labs post-trained Qwen3.5-27B and 9B into 'decision models' that return calibrated probability distributions over choice, yes/no or score questions in one forward pass, with no text generation. Switching the hybrid models' full-attention layers to block-bidirectional masks lets questions share a cached state prefix (48.6 vs 44.8 in ablation); training used 6.21M examples with cross-entropy plus a Brier term. Mammoth-1 scores 71.24 on JDI 0.4, first of 208 models; Elephant-1 ties d1-ultra at ~65 on the Artificial Analysis Decision Index. Weights are on Hugging Face.

Canva's Craft plans layout, scale, hierarchy and text before generating images at up to 2K in 30 aspect ratios, and was trained on human-designed templates from Canva's library rather than photos, targeting legible text and localized edits. Outputs can be converted into layered editable designs via Canva Magic. It is available to all users and via the Canva API, with MCP access from ChatGPT and Gemini planned, and cheaper Craftfast and Craft-ultrafast variants coming. The in-house model should cut Canva's third-party API costs after it trimmed 2026 growth forecasts over AI spend.

Anthropic is accepting proposals from PIs at academic, nonprofit and government labs to work with its researchers for 8–12 weeks on causal human biology, structural/computational molecular biology, medicinal chemistry, drug discovery and clinical science. Anthropic runs its models and harnesses at no cost, and the outputs, data and code are to be published and open-sourced. Investigators keep the rights to their data. Applications are open for about 14 days. The program sits alongside Anthropic's Life Sciences Verification Program and AI for Science credits.

Google is partnering with Voltus to pay homes and businesses in the Mid-Atlantic PJM grid region to reduce electricity use during peak hours, which the companies call a 1-gigawatt virtual power plant. It is Google's latest effort to offset the data center boom's strain on power demand in one of the grids most pressured by AI build-out.

In a 14-page arXiv essay, P. K. G. Williams (Center for Astrophysics | Harvard & Smithsonian) treats the scientific literature as an information commons. He compares AI's effect on it to the English enclosure movement (roughly 1600–1800), which he argues, not the 'tragedy of the commons', is what destroyed agricultural commons. The paper claims AI makes a previously infeasible enclosure possible, turning shared knowledge into private wealth and weakening the social systems that sustain science, including through 'slop spam'. It offers a conceptual framework, not concrete remedies.

Atlanta-based Tractian raised $130M in a Series D led by Sapphire Ventures at a $1.4B valuation. The seven-year-old company sells condition-monitoring sensors for industrial machines plus models that predict time to failure and recommend maintenance. It is pitched as 'physical AI' for factory equipment that has no built-in instrumentation.

Uber is deploying ChatGPT Work, OpenAI's enterprise agent launched in July 2026, company-wide after a pilot of about 1,000 employees in which roughly 80% of early adopters were non-engineers. Uses included analyzing operations data, localizing training materials and building personal dashboards. Uber previously embedded 30 AI engineers with teams for two-week automation sprints and will track resulting workflows to guide broader deployment; spending was not disclosed.

The Army is testing drone-mounted payloads that combine government and commercial components for non-kinetic effects such as jamming and Wi-Fi access. Its core system is the Radiofrequency Delivery Effects Platform (RDEP), which loads exploits onto cards. The project manager for cyber warfare says Army Cyber Command has a prototype kit and is weighing dismounted, vehicle and drone form factors, plus spectrum monitoring to avoid hitting friendly systems. No production decision has been made, and a replica 12th Cyber Battalion is due in the coming months.

Concierge clinic Biograph ran Caristo's CaRi-Heart AI on coronary CT angiograms of asymptomatic, low-risk members. 42.9% had elevated perivascular inflammation in at least one major artery, 17.9% very high, and about 47% of those with no detectable plaque were still inflamed. Standard markers (ApoB, hs-CRP, HbA1c) and the PREVENT risk score did not track it. The study is a non-peer-reviewed medRxiv preprint from a select, partner-affiliated cohort and measures prevalence, not outcomes. The authors say prognostic value and treatment benefit are unestablished.

Etsy's updated seller rules require at least one unedited photo of the actual item on a real background, with exceptions such as print-on-demand and digital goods, and allow AI edits only if the item's appearance is unchanged. Sellers still must disclose AI-created items. AI-generated functional guides such as patterns and recipes are banned unless sellers show a real result, and fully bot-generated listings can get a shop removed. Etsy is adding AI-image detection, a buyer reporting option, and refund exposure for listings materially unlike the item. AI art stays allowed when the seller plays a clear creative role.

Speaking at Cloudflare Connect 2026, Steinberger, now OpenAI technical staff, estimated about $10 million in token spend in one month under a 'virtually unlimited' budget, admitting not all of it was useful. He argued agent infrastructure should move to the cloud, that AI-assisted pull requests should document how contributors reached the final code, and proposed agents granting or denying other agents' permissions. He expects each person to have one ambient, anticipatory intelligence rather than hundreds of agents.

OpenAI signed an agreement with NASK, which runs Poland's AI Safety Institute, to collaborate on AI safety, evaluation of advanced systems, safeguards, and detection of synthetic content, including its multilingual aspects. The deal builds on a July pilot in which CERT Polska, part of NASK, used OpenAI's cyber-capable models via Daybreak to help find six MikroTik RouterOS vulnerabilities. Some of them could be chained into unauthenticated device takeover; the vendor has since patched them. No financial terms were disclosed.

On No Priors, CrowdStrike CEO George Kurtz says adversaries already run AI agents and the fastest intrusions now take seconds, because agents give low-skill attackers advanced tradecraft. He describes a frontier-style lab CrowdStrike is building with Nvidia to train security-specific models. The conversation also covers how open-weight and closed models compare for attackers, model distillation, lessons from the recent Hugging Face incident, and why he thinks defenders can still come out ahead.

Minerva pulls telemetry from GPUs, SuperNICs, switches, storage, schedulers and jobs into a dynamic knowledge graph, so operators can trace a slow training job to the degraded component behind it across hundreds of thousands of GPUs. It applies supervised and unsupervised models to billions of data points per hour to detect new anomalies, predict failures and collapse alert storms into incidents with a likely root cause. It optimizes 'AI factory yield' (goodput, uptime, tokens) and includes an assistant that connects to external agent tools.

Cohere's managed inference platform now runs workloads in a hardware-backed confidential environment that spans CPU and NVIDIA H100 GPU TEEs on Azure. GPU memory is protected and CPU–GPU traffic is encrypted. The client checks signed attestation of the hardware, firmware and software and refuses to send data that fails policy. Requests are encrypted end to end into the attested enclave, so intermediaries and operators cannot read them, and the existing API is unchanged. Cohere has not published latency overhead figures or a list of supported models.

Shield AI's second post on autonomy addresses a core problem for its autonomous pilot: in scenarios like a drone losing its operator link while tracking a vehicle under tree cover, the software must decide what to do on its own, and engineers cannot test every situation. The company argues trust must be built as evidence before deployment and maintained across the autonomy lifecycle, ensuring adaptive behavior stays within defined safety and operational boundaries rather than requiring operators to predict each decision.

LangChain's LangSmith LLM Gateway sits between agents and model providers to enforce policy once instead of per-agent: time-bound spend caps (returning 402 when hit) and rate limits at org, workspace, API-key and user levels, per-customer limits via headers, cross-provider model fallbacks, and PII/secret redaction in requests and traces (Enterprise only). It is BYOK across OpenAI, Anthropic, Fireworks and compatible endpoints, with Fireworks-hosted open models via credits. Public beta is open to Plus and Enterprise plans.

Opposable's Android app drives other apps via the accessibility service using Alibaba's GUI-Owl 1.5 2B (a Qwen3-VL fine-tune), 4-bit via llama.cpp. Cropping, background-token pruning (25–40% fewer prefill tokens), cached prompts and grounded early exit cut CPU step time from 10.6s to 7.3s; the ViT compiled for the Galaxy S25 NPU encodes a screen in 236ms in fp16. On eight held-out tasks, subgoal-by-subgoal execution succeeded 6/8 vs 3–4/8 whole-task; harder unseen apps get handed to cloud models.

Firebird, which built an AI factory in Armenia, has a US export license for 45,000 NVIDIA Grace Blackwell GPUs in Kazakhstan. Phase one is under construction with Kazakhtelecom at Data Center Valley for commercial launch in Q3 2027, and a second phase would bring potential capacity to about 500 MW. Kazakhtelecom already runs 400 H200s and 480 B300s in Astana and Almaty and plans more than 1,000 GPUs at Data Center Valley plus more than 1,000 Vera Rubin GPUs in Ekibastuz.

Reporting Q3 results of 788 million MAUs and nearly €5 billion revenue, co-CEO Gustav Söderström argued agents won't replace listening the way they replace chores, but will orchestrate experiences such as cadence-matched running mixes built from years of history. Spotify was first to integrate with Meta's Muse and works with ChatGPT and Claude; it says it doesn't share listener data with partners — agents send requests and Spotify's systems choose what plays. TechCrunch contrasts this with Amazon blocking Muse from shopping.

Mountain View-based Utopai Studios, a 4-year-old AI film startup, closed a $60M round at a $1B valuation and has already received term sheets valuing it at $2–3B, per a source close to the company. It follows a run of video AI financings: Higgsfield raised at $5.4B and passed $1B in annualized revenue, Kuaishou's Kling raised nearly $3B at a $15B valuation, and Jeffrey Katzenberg is launching an AI film company with the former head of OpenAI's Sora app.

In a War on the Rocks interview, Rakesh Kumar, author of Chip Age, argues chip rivalries date to the 1960s and shortages recur from cyclical capex and efficiency-optimized supply chains. He predicts TSMC and Samsung exit legacy nodes within 2–3 years, says CPU performance now grows only 10–15% annually, and proposes time-bound government-industry consortia on new computing paradigms as a precondition for CHIPS funding. He expects CUDA's moat to erode as AI ports code, and urges a 'king for a day' plan steering NVIDIA, Apple and Qualcomm production to Intel.

AI2's PLANTEUR extends the PASTIS crop-type benchmark from mainland France to six tropical French territories, including Réunion, Guadeloupe and Guyane. It has 2,750 Sentinel-1/2 and Landsat-8 time series, 4,506 km² of parcel labels, and seven crop classes such as sugarcane, banana and cassava. Every tested method, linear probe or supervised, scores well below its PASTIS results, and models lose roughly half their accuracy on a held-out island. The dataset and code are on Hugging Face and GitHub.

Edward Chalk describes Minotaur, which models a system as typed driver/constraint/relation (ELC) networks and answers deduction, abduction and diagnosis queries by clamping some nodes and letting the network settle. An exact arithmetic engine checks the results through logged one-way 'crossings'. On self-designed tests it found 40 of 40 planted circuit faults and scored 30/30 on an order-dependent interview task, matching an LLM baseline on a single CPU core. The tests are small and author-built, the learning component is patent-pending, and nothing shows it working on open-ended language.

A draft performance work statement shows Special Operations Command doing market research for an 'integrated contracting vehicle' to run a shared platform-as-a-service environment near MacDill AFB, where traditional and non-traditional vendors could prototype, integrate, test and field software across several PEOs. Scope includes cloud-native engineering, an operationally representative integration test platform, network and hardware support, commercial market analytics, and end-to-end RMF support to achieve and keep authority to operate. It would also support Joint Acquisition Task Force experimentation.

An independent LessWrong report replaces dense queries and keys in GPT-2 small's attention with sparse features, using a query-key transcoder architecture akin to Wynroe and Sharkey's bilinear sparse dictionary learning. The main contribution is an interactive dashboard of feature pairs for every layer and head, analogous to Anthropic's SAE dashboards but focused on the QK circuit, plus exploratory analysis of how attention patterns form — an area residual-stream and MLP feature tooling largely neglects.

Nirva, founded by former Meta Reality Labs staff, raised a $25 million pre-Series A (Lightspeed, South Park Commons) for $33 million total. Its 8-gram pendant, which lasts three days per charge, uses an enrolled voice model to process only the wearer's speech, discards audio and transcripts by default, and tracks tone, pitch, volume and pace to infer emotional state. The companion app surfaces personality patterns, frequent contacts and reminders. Orders open in November at $249, with shipping targeted by year-end.

Philip Zeyliger splits agents into model (capability), loop (UX) and tools (capability and risk), and maps deployments: Claude Code runs loop and tools locally, ChatGPT runs all three server-side, Amp runs tools in a per-thread VM. He argues VMs are the gold standard for tool execution, lightweight sandboxes (Pyodide, read-only SQL) often suffice, and token costs upend SaaS economics, favoring bring-your-own-subscription models. He predicts products 'legible' to coding agents will win and that no single loop fits all uses.

Cross-posted from the EA Forum, the essay argues that instrumental convergence belongs to a class of pre-empirical categorical predictions with a low historical success rate, that its definition is murky and likely unfixable, and that on its most charitable reading it predicts AI behavior worse than a good high-level model should. The author, self-described as a non-expert with a few hundred hours of reading, reviews current usage across AI safety, including discussion of the Hugging Face attack, and proposes replacing the concept with a more empirically grounded alternative.

A LessWrong explainer aimed at non-specialists outlines why a unilateral AI slowdown fails game-theoretically, then surveys unilateral and bilateral verification mechanisms drawn from academic literature that could underpin a US–China agreement. It ends with a scenario in which a warning-shot incident leads both countries to combine these tools into a verifiable slowdown. The author presents it as reference material rather than original research.