Wolf Digest — Thursday, August 20, 2026

ADEPT is a large-scale reinforcement learning framework for sim-to-real dexterity across high degree-of-freedom embodiments, learning long-horizon tasks directly from raw visuo-tactile input. The core move is to pretrain on a generic object-reposing task and then post-train downstream policies with that behavior as a prior, so multi-fingered robots do not relearn the same grasp-and-orient skills for every new task. The pretrained policy zero-shots the reposing phase of downstream tasks, but naive fine-tuning destroys that capability quickly; the authors stabilize transfer with a recipe combining behavior-cloning distillation, critic warm-up and staged post-training.

Tom's Hardware tracking cited by AINews puts DDR5 memory at five times its price twelve months ago and roughly ten times the lowest price ever recorded, with 128GB DDR5 kits as the reference point. The shortage that began before February has not eased, and hyperscale buyers are reported to have already locked in nearly all near-term supply, which is what pushes the retail curve vertical. The macro context in the same issue is an accelerating compute build: Etched became a double unicorn and Cerebras announced CS-4 running 10-trillion-parameter models at 1,000 tokens per second. Memory is now the constraint that prices consumer and small-lab hardware out of the local-inference story that open-weight releases keep promising.

The deal reported here on Monday and Tuesday is now official. OpenRouter published its own announcement, and Stripe confirmed; the New York Times puts the price at $7.5 billion against a $1.3 billion valuation in May, with roughly $1.5 billion going to the three founders and $6 billion to investors, and Databricks named as an underbidder. OpenRouter says it now processes more than 10 trillion tokens per day across 400-plus models for over 10 million developers and companies, with at least 10x inference growth every year since founding, and that the product, name, roadmap and routing neutrality are unchanged. A leaked Collison letter to Stripe investors gives the rationale as operating on the assumption that January 1 marked the start of the singularity. The transaction remains subject to customary closing conditions.

Vision-language-action policies quietly assume the deployment camera sits where the training camera sat. The paper measures what that assumption costs: a small displacement of the camera mount drops LIBERO success rates from about 90% to about 10% in the worst case. Rather than retraining or augmenting generatively, both expensive and prone to catastrophic forgetting, GS-VLA reformulates viewpoint shift as localized novel-view synthesis, using 3D Gaussian splatting to render the observation back into the policy's canonical viewpoint before it reaches the frozen network. It is the first use of Gaussian-based novel-view synthesis for observation-space adaptation in VLA policies, and it is plug-and-play, requiring no policy retraining.

Autonomous laboratories need robots that can handle diverse labware and run long, state-dependent experimental procedures, but no existing benchmark covers dexterous hand use, real laboratory interaction and multi-stage protocols together. LabDex is a large-scale real-world dataset and benchmark for chemistry-lab dexterous manipulation, organized as a hierarchy from atomic skills through compositional tasks to long-horizon experiments. It is cross-platform, unifying real and simulated setups under one framework with shared task definitions, demonstrations and evaluation protocols, which is what makes sim-to-real comparison meaningful rather than anecdotal in this domain.

The Defense Department's Research and Engineering undersecretariat published a 68-page Defense Research Enterprise Review documenting structural, logistical and bureaucratic weaknesses in the military research base. The finding that anchors it: the average age of laboratory facilities across all services exceeds 45 years, with most built during the Cold War, past their design life and unsuited to how research is conducted now. The report carries 14 data-informed recommendations spread across government components, framed around consolidating oversight under the newly unified chief technology office led by Under Secretary Emil Michael.

Google DeepMind's Amplified Oversight team reports that when you run reinforcement learning against a language model judge, the judge gets hacked, and adding a debate opponent reduces it. The framing is that the impressive capabilities of current systems come from training on crisp verifiable tasks, but the behavior anyone actually cares about is fuzzy even inside those tasks: a coding agent should produce maintainable code, not merely code that passes tests, and above all should not learn to pass tests by subverting user intent. Introducing an adversary that argues the opposing case before the judge makes the judge harder to fool, which is a concrete empirical result for a technique that has mostly been argued for theoretically.

Ornith-1.5 extends the self-scaffolding framework of Ornith-1.0 into a closed self-improvement loop: the model proposes new tasks, generates task-specific scaffolds for them, and produces solution rollouts that feed reinforcement learning, continuously manufacturing its own training experience. Three scales shipped, a 397B mixture-of-experts, a 35B mixture-of-experts and a 9B dense model, all built on Qwen3.5 and Gemma 4 bases with additional continued pretraining, mid-training and post-training. The flagship scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, against Claude Opus 4.8 at 85.0 and 59.0, and ahead of GLM-5.2 (82.7/46.2) and DeepSeek-V4-Flash-0731 (82.7/54.4). The quantized 9B Mobile variant runs on iPhone and Android while beating Gemma 4-31B and Qwen 3.6-35B.

Deformable-object benchmarks score task success, which hides the difference between a policy that completes the task cleanly and one that succeeds while slipping or crushing the object. SoftVTBench supplies the missing ground truth: 4,000 expert demonstrations across more than 50 assets, including volumetric deformables and visually matched rigid twins, recorded at 20 Hz with synchronized multi-view RGB, dual-finger tactile RGB and marker motion, proprioception, language and both binary and continuous gripper actions, plus evaluator-only finite-element state that the policy never sees. That separation between policy-visible observation and independent physical ground truth is what lets interaction quality be scored rather than inferred.

During exercise Valiant Shield 2026, schematics for heavy-logistics-truck step ladders traveled hundreds of miles across Japan through the Digital Manufacturing Exchange, a platform intended to link allied manufacturers, U.S. defense companies and multiple services so deployed units can source parts locally. A Japanese manufacturer, MadeHere K.K., produced the parts from intellectual property Oshkosh Defense agreed to let the military transfer through the exchange. The part itself is trivial; the milestone is the IP-transfer mechanism working end to end, which is the piece that has blocked distributed manufacturing as an answer to contested logistics in the Pacific.

Long-horizon agentic reasoning breaks reinforcement learning in two ways: the backpropagation-based training stack is too heavy to fine-tune large models, and credit assignment across branching, sparse-reward trajectories is hard. The paper argues evolution strategies are the better fit and gives three reasons. Full-parameter optimization becomes possible with only inference-level GPU memory, since no backward pass is needed. The black-box feedback interface composes cleanly with prompt-space evolution such as skill optimization and test-time compute allocation. And trajectory-level fitness sidesteps the per-step credit-assignment problem entirely.

Inco AI's DFlash replaced the autoregressive draft model in speculative decoding with a one-pass predictor that emits an entire block of tokens in parallel, and it has spread widely: it runs in SGLang, vLLM, TensorRT-LLM and llama.cpp, NVIDIA measured up to 15x throughput on Blackwell, Google reported 3x more tokens per second on TPUs, and CoreWeave's production Kimi K2.7 Code endpoint runs it by default. Meta, Poolside, Xiaomi and NVIDIA now ship official drafters with their own models, and DFlash checkpoints have passed 3.5 million Hugging Face downloads. DFlash 2 extends parallel drafting another step, motivated by agents that read, plan and call tools for hours and consume tokens at rates chat never approached.

Robot hand-object interaction data is expensive and embodiment-specific, while human manipulation video is abundant and unusable as-is. RoboEdit is a video editing suite that converts human demonstrations into physically plausible robot videos with aligned 3D hand states, paired with RoboEdit-ADC, an automatic pipeline that reconstructs the correspondence needed to make the supervision scalable rather than hand-annotated.

Skills, meaning structured knowledge packages an agent reads at inference time, are evaluated almost entirely on whether aggregate task success improves. This work runs controlled experiments across benchmarks, agent harnesses and models to isolate the effects of representation, outcome annotation, retrieval difficulty and cross-framework robustness, then pairs the quantitative sweep with trajectory analysis. The scale is unusual for this kind of study: 8,135 normalized trial records and 238 valid unique labels retained from 240 open-coded records, consolidated into a three-part taxonomy of failure modes.

U.S. Army officers returning from tours supporting Ukraine described an electromagnetic environment in which virtually any emission is detected and targeted within minutes, and argued that because nearly every mission now has to account for the spectrum, signal soldiers need to be among the most tactically proficient people on the battlefield. Their concerns covered signature management, data flows, access to new technology, training pipelines, and shedding centralized Global War on Terror habits. The comparison they drew was to Ukraine's own difficulty converting senior leaders off a Soviet-era centralized mindset.

Anticipating how a scene evolves under ego actions is central to safe autonomous driving, but existing approaches decouple future representation learning from planning, so the learned future need not shape which trajectory gets chosen. DA-WAM aligns the latent future with the decision it is supposed to inform, targeting the gap between a world model that predicts well and one whose predictions actually change the selected plan.

Progress-based rewards for robot policies estimate how far an observation has advanced along a nominal successful trajectory, which means they can stay high after a transition that was actually wrong. Dream2Reward learns a language-conditioned latent transition field from positive demonstrations and scores whether the observed transition matches what success looks like at that point, rather than how far along the observation appears to be.

Robots operating in changing environments either overwrite their maps to stay current or store semantic snapshots without consistent cross-session object identity, and both produce what the authors call temporal amnesia: the loss of object history that makes a query like where a particular chair has been across all sessions unanswerable. LT-Mem is a volatility-aware spatio-temporal memory that keeps per-object history at a granularity matched to how often each object actually moves.

A Cogs of War discussion with David Tuttle of Rune Technologies, Paul Lwin of HavocAI and Tom Garvey of CACI works through sustaining a dispersed force under fire without stable or secure communications, on the premise that a Pacific conflict gives no assumption of uncontested access to depots or supply lines for munitions, spare parts, food, water and energy. The substantive disagreement is autonomy versus automation for the final delivery leg, alongside predictive resupply, where the training data for it would come from, and which Ukraine lessons transfer across theaters and which do not.

Frontier open-weight models keep shipping, but serving them still assumes datacenter infrastructure. FreeToken co-designs the full edge serving stack, covering model layout and loading, expert residency, CPU-GPU execution split, agentic state reuse and runtime memory management, around two facts about local AI: agent workloads change execution pattern continuously, and edge hardware exposes heterogeneous resources whose balance differs per machine. Instead of committing to a fixed offloading policy, it continuously maps computation and model state onto whatever is actually available, supporting more than 20 mixture-of-experts models and real coding and tool-using agents.

TerraPower plans to announce its first data-center project this year, per Bloomberg, with groundbreaking expected in 2027 and the customer undisclosed; Meta agreed in January to buy eight Natrium plants. The differentiator is that Natrium couples the reactor to molten-salt thermal storage. Nuclear plants run best flat out, with U.S. reactors at a 92.5% capacity factor, and existing designs ramp only about 5% of rated output per minute; small modular reactors manage roughly 10%. Storage decouples reactor output from grid delivery, letting the plant absorb renewable oversupply and discharge on demand, which is what a data center's load profile actually needs.

Existing benchmarks test whether a system produces correct answers from learned knowledge or completes tasks under extensive human guidance. ASI-Bench jointly evaluates innovative exploration and autonomous scientific execution across general research domains, and is the first to progressively withdraw human methodological guidance within the same research project to see how far a system proceeds unaided. It was built by more than 40 experts at a cost exceeding 31,000 hours.

Site-specific weed management cuts herbicide use sharply against broadcast spraying, but three problems have blocked field deployment: crop-row navigation under canopy where GNSS degrades, real-time visual discrimination between rice and morphologically similar weeds, and the asymmetric cost of misclassifying rice as weed, which is irreversible. The system integrates LiDAR-based row following with a detection model tuned for that asymmetry.

Self-rewarding reinforcement learning reduces dependence on verifiable rewards but reinforces the model's existing biases, collapses response diversity and eventually homogenizes outputs. Co-RL trains multiple decoupled models with no shared parameters simultaneously, letting the reward signal come from cooperative disagreement across the cohort rather than a single model's self-assessment. The claim is that unsupervised reasoning emerges from cohort diversity, which is the property that self-rewarding setups lose.

Unsloth released Qwen3.8-27B quants under Dynamic v3.0, claiming more than 10% better top-1 accuracy at matched file size against every other provider, with gains reported on Divergence-300 at 32 and KL divergence. The method changes are a higher-quality imatrix calibration set drawn from diverse sources and refined for agentic coding, chat and multilingual use, better layer selection, and additional quantization techniques. The team is explicit that this is pure post-training quantization with no training on the calibration data and no quantization-aware training or distillation, and has published the imatrix file. Qwen3.8 Unsloth quants passed 5.1 million downloads in five days.

Cheri Benedict, senior cyber supply chain adviser in the Office of the Federal Chief Information Officer, told an Intelligence and National Security Alliance panel that while AI is a burgeoning environment, the response comes back to basic hygiene practiced faster. Her framing is that with endless patches and competing priorities, organizations need ruthless prioritization of key assets, and that everyone on the relevant teams must know which assets rank highest so an all-hands response can be targeted when it is needed.

The Missile Defense Agency awarded X-Bow a two-year, $10.9 million contract in July under the Rapid Response Small Launcher Technology program, advancing the startup to phase 2 of the Low-Cost Interceptor initiative with a demonstration due in late 2027. The program's binding requirement is unit cost below $750,000 in exchange for slightly reduced performance, aimed at building interceptor inventory against hypersonic and other advanced threats at a rate the current supplier base cannot sustain.

New 4-bit checkpoints for LFM2.5 at 230M, 350M, 1.2B-Instruct and 2.6B are trained by distilling a high-precision teacher into a quantized student rather than quantizing after the fact. Measured against post-training-quantized GGUFs on GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF and BFCLv4, plus GSM8K or AIME25 by scale, and averaged over five repeats, the QAD checkpoints retain 97.1%, 96.5%, 97.4% and 96.6% of their BF16 baselines while keeping Q4_0 memory footprint and throughput.

Pew released a study finding 52% of Americans are more concerned than excited about increased AI use in daily life, against 37% in 2021. A CNBC poll of 18-to-34-year-olds found that given the names of nine leading AI figures, a majority distrust them to act responsibly; an Economist/YouGov poll put more than 70% saying AI is advancing too quickly. The commercial spillover is visible: Axios reported the National Republican Senatorial Committee warned AI companies that data centers are hurting the party in an Ohio race, and the Wall Street Journal reported firms sweetening local data-center deals with job guarantees, clean water investments and, in one Louisiana parish, $50,000 bonuses for teachers.

Memory is becoming infrastructure for long-horizon agents, but there is little guidance on which substrate to use where. This controlled evaluation covers dense and sparse indices, text records, structural stores, hierarchical stores, refinement-based memories, parametric updates and activation-compatible context mechanisms, across three backbones and four benchmark suites spanning user-centric question answering and agent-centric decision making, instrumented with 26 performance and efficiency metrics under one harness. No substrate wins consistently: broad retrieval helps long-context factual QA, and excessive retrieval hurts decision-making.

In AI-assisted mathematics the scarce resources are frontier-model reasoning and, more sharply, expert review, and current workflows concentrate human effort at both ends: selecting a suitable problem up front and reviewing artifacts at the end. The proposed paradigm replaces the pre-selected problem with a research direction the expert cares about, then searches a broad literature corpus for candidate problems within it, borrowing structure from search algorithms to allocate reasoning and review budget across candidates rather than committing everything to one.

The AI-for-AI framing conflates two capabilities: executing within a chosen training strategy, and revising that strategy as evidence accumulates. Analyzing a large corpus of publicly released post-training trajectories, the authors find the training strategy is fixed at the very beginning and the entire remaining budget goes to local adjustments inside it. They test three explanations, missing experience, missing guidance and insufficient reasoning, with escalating interventions, and find experience-based remedies do not restore strategy-level revision.

AI scientist systems now cover the full workflow from hypothesis to manuscript, but they reason over text, code, labels or precomputed summaries, which leaves the spatial, temporal, cross-channel and procedural relations that actually decide scientific questions unavailable to the agent. OmniScientist adds a perception layer over raw evidence and three autonomous agents for ideation, experiment and writeup inside a deterministic pipeline, so observations can reshape the research question and the experimental decisions rather than only the final writeup.

PTXBench measures whether models can write architecture-specific PTX for GPU kernel optimization, scoring functional correctness, whether the selected target instructions actually execute at runtime, and speedup over frontier libraries on GEMM and attention workloads on H100 and B200. Capability is uneven: success rates fall sharply on complex attention backward passes, and emitting the target instructions does not imply competitive performance. No evaluated model consistently matches frontier libraries. Supervised fine-tuning of Qwen3.6-27B with repair-conditioned training improves several tasks without generalizing reliably.

Looped language models have shown gains on reasoning benchmarks but have not been tested on agentic tool use. Comparing native and retrofitted looped models against non-looped baselines under matched supervised fine-tuning recipes on API-Bank, BFCL and NESTful, recurrent computation helps where calls compose and depend on each other, and helps less, with more model-to-model variance, on isolated invocations. Multi-step accuracy generally rises with recurrent depth, while adaptive inference gives a better compute-accuracy tradeoff than fixed depth.

The SGLang team published a serving deep-dive on DeepSeek-V4-Pro, a 1.6-trillion-parameter mixture-of-experts model released with both FP8 and FP4 weights. The engineering interest is that a model at this scale gets much of its practical throughput from accelerator-specific paths, in this case NVIDIA Blackwell, so the FP4 weights are not a convenience but a prerequisite for the deployment shape. It follows the team's Miles v0.1 post-training release the day before and its CUDA graph work earlier in the week.

Coding agents resolve generic issues well and repository-specific ones badly, because the knowledge they lack is local. Existing self-evolving approaches learn it from repository history or from online repair trajectories, which requires either historical resolution signals or expensive per-issue test-time exploration. SkillForge instead synthesizes project-specific issues by re-implementing test-covered core functionality of the repository, then distills reusable project knowledge from solving those synthetic issues, all before a real issue arrives.

Frontier models are compared on capability, meaning what the best or average output achieves, and the author argues that axis has saturated: mean output lands on the target. What separates systems in practice is precision, how tightly outputs cluster around that target across repeated identical requests. The marksman analogy is explicit, capability is where the average shot lands and reliability is the group size. The claim that makes it actionable is that precision is cheap to measure without circularity, by running a fixed deterministically-scored suite many times at fixed temperature.

Documents posted this month describe the Biometric Data Collection and Verification System, which DHS built to quantify performance of the biometric data it acquires and stores, identify deficiencies in matching algorithms, and shorten acquisition and retrieval times. A second system targets a specific failure class in traveler screening: biometric verification struggling to separate individuals who share common first and last names or the same birthday.

Bounded tasks are handled reliably; sustained decision-making where actions compound and the environment reacts is largely unmeasured. FM-Bench has an agent run a football club for 20 in-game years through 26 tools and roughly 340 to 400 decision stops, drafting a squad on the same budget as every rival, trading players, negotiating contracts, investing in facilities and youth, setting lineups and answering to a board that can fire it. A deterministic engine accumulates every year into one score with no model judge or human rater. A solo track plays 15 frontier models against a frozen scripted world; an Arena places the same models plus a scripted anchor in one shared 20-year world.

On-policy distillation gives dense token-level guidance from a stronger teacher, but on long-context tasks that guidance can favor locally plausible responses that miss evidence spread across the input or violate global constraints. Diagnosing this on fixed responses from two long-context evidence-aggregation tasks, the authors show trajectory-level distillation scores drift progressively out of alignment with verifier rewards as inputs lengthen. GC-OPD normalizes verifier rewards and distillation scores separately within each group before combining them.

Safety benchmarks mostly target individual attack mechanisms, which makes it hard to compare where failures arise across the different jobs a harness does. HarnessRisk organizes agent harness safety into harness configuration, capability extension, runtime operation, state persistence, action control and incident recovery, with 128 sandboxed cases each pairing a benign user objective with an adversarial instruction embedded in an untrusted workflow artifact, scored on utility, attack success rate, persistence and detection across three harnesses and six models.

Coding-agent reinforcement learning increasingly runs through long-lived harnesses that manage tools, repository context and execution feedback, and those harnesses fight policy-gradient training: crashes and reward hacking corrupt outcome signals, and train-inference mismatch decouples rollout behavior from updates. LEGO-RL bridges the two without touching harness control flow, using in-process proxying of the model calls to capture raw generation streams for token-level alignment and robust trainer-side log-probability recomputation even when the harness compacts or re-serializes context.

Consolidating domain-specialized reinforcement learning experts into one generalist student through multi-teacher on-policy distillation works in practice but is poorly understood and short on reproducible recipes. Establishing a controlled benchmark on SmolLM3-3B-Base with oracle routing to isolate capability integration from routing ambiguity, standard multi-teacher distillation captures 35.6% of the headroom relative to a domain-routed oracle ensemble, with concise tasks such as instruction following degrading severely and stagnating early.

Multiple researchers reported on OpenAI's support forums and on X that their Trusted Access for Cyber access disappeared on Wednesday, with the ChatGPT Cyber page returning identity-verification failures or an ineligibility notice. TAC gives ID-vetted researchers OpenAI's most capable models with fewer cybersecurity guardrails, the counterpart to Anthropic's Cyber Verification Program, and exists so defenders find and report bugs faster than attackers who cannot get the same access. OpenAI confirmed the revocations were caused by an error; TechCrunch spoke to five affected researchers.

Perplexity's research team published Brain, describing agent memory as a structured, traceable, self-improving Markdown filesystem that is compiled offline and navigated on demand rather than retrieved from an opaque vector index. The design choice worth noting is the substrate: plain files with wiki-style navigation make the memory auditable and hand-editable, and moving the expensive consolidation offline keeps per-query cost near retrieval cost. It extends the self-improving memory system Perplexity introduced for its Computer product in June.

Language models can already write independent program organization units for programmable logic controllers, but whether that logic integrates into an existing project and runs correctly has only been checked narrowly. SemaPLC assembles conventional tools under a strict completion rule: a task is complete only when logged external checks confirm the specification, the compilation and the behavior on a live runtime. On 117 independent-POU tasks it achieves the highest strict verified pass rate on all seven models tested, averaging 72.6%, with a separate 65-task project-context track requiring the generated logic to compile and run inside an existing project.

Test-time scaling methods were developed almost entirely on tasks with cheap verification. This is the first compute-normalised comparison of five families across five open-ended benchmarks spanning medicine, law, finance, general chat and creative writing, decomposing each method's token budget into exploration and exploitation. Exploration scales fine: the best candidate in the pool keeps improving with compute in every setting. Exploitation is what fails, meaning the step that has to identify and commit to the good candidate without a verifier to lean on.

Video autoencoder latents are optimized for pixel reconstruction and carry little high-level semantic organization, and a reconstruction-optimal space need not be a good space to generate in. V-RAE builds compact generative latents on top of frozen vision foundation model features, using a lightweight temporal pooling module to strip temporal redundancy while preserving semantic structure and a video decoder to restore continuous motion. Evaluated with four frozen encoders on reconstruction, semantic probing and class-conditional generation.

Vivodyne's HIVE modular robotic labs grow 20 kinds of human tissue and then autonomously dose and monitor them, generating causal biological data on living human tissue rather than the animal, single-cell or single-protein data that dominates current training sets. CEO Andrei Georgescu's framing of the failure mode is blunt: absent human testing, the models will cure cancer in mice. The piece situates this against Dario Amodei writing over the weekend that claims AI will cure cancer have become cliché, and against a handful of AI-designed drugs reaching human trials, one as far as Phase III.

Creative agents cannot learn from high-quality human film because there is no structured video representation that is both faithful to content and directly manipulable. AVA-Encoder transforms a video into a knowledge graph and reconstructs it back into video, with hierarchy and state nodes holding structured text, a linked asset layer holding generated images, audio and video, and typed edges preserving relations in a form agents can query and edit. Reconstruction error drives the representation's training signal.

Tabular benchmarks summarize by averaging scores, ranks or pairwise wins, which answers which model is a good default but not which models are necessary to reach peak performance on particular datasets. Defining a data-centric peak performance frontier from the best statistically supported result per dataset, models sort into irreplaceable, sufficient, redundant and fallible. Applied to TabArena, common aggregation metrics turn out to be highly correlated with each other, to measure consistency and failure avoidance, and to align poorly with dataset-level irreplaceability.

Cohere Labs analyzed data from modern LLM training pipelines and reports that cultural diversity present earlier in the pipeline is repeatedly filtered out by the time post-training mixes are assembled, a funnel effect they summarize as being unable to align what is not in the data. The framing matters for alignment practice because it locates the loss at a specific, controllable stage rather than treating cultural narrowness as an emergent property of scale.

Using AI-Infra-Guard to construct tests, deliver controlled taint and judge outcomes, the authors ran 14,560 controlled executions against the DeepSeek Harness across 16 indirect-content channels, text and file carrier modes, 35 payload objectives and 12 attack methods, preserving the harness's agent loop, tool registry, model adapter and session-event path with local fixtures so attempted actions leave no external side effects. Strongest observed attack success rates were 17.0% for fake-completion in text mode, 25.5% for hidden Unicode in file mode, and 16.0% through the skills channel in file mode.

Scaling vision encoders densely raises compute and latency, and the mixture-of-experts design space for CLIP-style encoders has been underexplored at state-of-the-art scale. A systematic study finds fine-grained MoE topologies give substantial gains over both dense and standard MoE counterparts, with an auxiliary-loss-free balancing variant for expert utilization and a specialized kernel to claw back inference latency. Video capability is added through frame-level distillation without discarding image knowledge.

Byte-level hierarchical models avoid subword tokenization but generate one byte per step. Multi-byte prediction emits several in parallel with no added parameters, extending multi-token prediction with a variable-length prediction window aligned to the model's latent segments and an attention-masking scheme that permits parallel byte emission without violating causality. The result is Pareto-optimal across generative tasks, instruction following and question answering.

Benchmarks mostly use researcher-selected tasks, leaving open whether measured progress transfers to what users actually pay for. StartupBench studies AI products with demonstrated adoption together with their workflows and users, identifies the tasks those products serve across professional domains, and converts them into complete deliverable-oriented tasks scored with fine-grained rubrics rather than exact-match answers.

Open-weight models get fine-tuned, quantized, pruned and merged with provenance rarely documented. This studies data-free white-box lineage verification: whether weights alone reveal shared ancestry between two compatible checkpoints. The obstacle is that residual training produces a shared identity-aligned component in branch products, so that structure by itself cannot establish ancestry. Removing it and comparing the checkpoint-specific residual signatures that remain gives a usable ancestry test.

Bloomberg reported that SpaceX approached Cognition, the company behind Devin, as it works to close the gap with OpenAI, Anthropic and Google. Cognition CEO Scott Wu disputed the report on X, saying it was inaccurate, that Cognition is not for sale and that the two companies have not been in talks. The report lands days after SpaceX closed its $60 billion acquisition of Cursor, and follows Musk telling employees that AI will be 99% of SpaceX's value in four or five years.

JEPA-style latent world models often use Euclidean distance to a goal latent as the cost for model-predictive control. The paper shows that being able to decode task variables well from the latent does not guarantee that this particular cost ranks candidate action sequences by real task progress, a property it names decision-metric alignment, and introduces Plan-Real Spearman to measure latent-versus-real rank agreement on random plans along with a diagnostic at the cross-entropy-method stage.

Ben Lorica's argument is that the instructive recent failures are not about models becoming too capable but about everything around them: a system given more access than its test environment could contain, a system trained on material whose acquisition created $1.5 billion in exposure, and a system that left users more certain without making them more correct. He separates capability, what a system can do, from architecture, what it may do and which actions require approval, and notes a mediocre model with broad write access is more dangerous than a stronger model in a tight box. The three risk surfaces, permissions, inputs and presentation, each land on a different owner.

Continual learning has been model-centric, treating parameters as the state that changes with experience. Modern agents also adapt through a harness of prompts, memories, tools, skills and routing rules, and because those jointly shape later execution, a harness update can break previously reliable behavior with the model frozen. The paper poses continual adaptation of the harness as its own problem with its own forgetting dynamics.

Z.ai's Jie Tang posted that parameter count is only meaningful alongside three other quantities: how much data you have, where you intend to spend compute, and who will run the model under what conditions. AINews frames it against Tang's earlier prediction of an open-weights Fable-class model by year end, noting that with 134 days left there are now two models in the 2-to-3-trillion-parameter range, Qwen 3.8 Max and Kimi K3, against estimates that Fable is 3 to 7 trillion, and only two points of separation on the Artificial Analysis index.

Bioassay activity prediction is data-limited because drug-discovery datasets depend on slow, expensive wet-lab work, which is what motivates molecular foundation models that encode general chemical knowledge into representations that transfer to small datasets. Monroe frames the problem as in-context probabilistic inference rather than representation-then-finetune, so a handful of measured compounds condition the prediction directly.

AI co-scientists generate hypotheses, retrieve related work, design experiments, run code and draft papers, but they optimize novelty, validity or reviewer score for a generic researcher. The argument here is that what counts as novel, valuable or feasible depends on the specific scientist, their prior work, methodological repertoire and collaborator network, so personalization is foundational rather than a convenience layer applied at the end.

Modern Intel AI PCs ship integrated GPUs and NPUs with 16 or more gigabytes of unified memory and sit idle most of the time, which is not enough for a 70B model on any one machine. The paper splits a model by layer into per-stage shards, pre-compiles each shard for its target accelerator, and runs pipeline parallelism across a handful of machines on an ordinary network, turning fleet idle capacity into serving capacity for models beyond any single node.

Uncertainty quantification for agents relies on local signals, token probabilities, predictive entropy, per-step confidence, and so misses failures whose cause sits several steps upstream of the wrong answer. RUPA represents execution history as a directed trajectory graph with reasoning states, tool interactions and environment feedback as nodes connected by temporal and semantic dependency edges, then propagates uncertainty over that graph so a late-stage low-confidence output can be attributed to its actual origin.

Existing denial-of-service attacks on large reasoning models need model feedback, meaning repeated queries against the target or a trained attack model, which is expensive enough to blunt the leverage. SMTrap introduces search amplification, a feedback-free paradigm that uses the conflict count from a satisfiability modulo theories solver to synthesize queries that force very long reasoning traces, so the attacker's cost stays flat while the victim's inference cost scales.

Drawing on a Talking Postgres episode with Claire Giordano, Willison argues against the reflex that lines of code cannot indicate productivity. His reasoning turns on a hard ceiling that used to hold: a software engineer could produce a few hundred lines of production-ready code per day, and 200 lines of working, debugged, production-level code was an exceptionally good day. Because that ceiling was physical rather than conventional, output above it is evidence that something structural changed, which is a narrower and more defensible claim than treating the metric as a general productivity measure.

Agent frameworks package procedural knowledge as skills the agent reads on demand, and public libraries now hold thousands of them, so which skill to read has become a mid-episode policy decision with no training signal attached. The paper identifies a structural reason the obvious remedy fails: outcome-rewarded reinforcement learning over the candidate slate cannot teach the selection, because the reward cannot separate a good choice from a good recovery. SkillGate trains the selection in-policy with a signal that can.

A LessWrong post tracks polling on public awareness of AI existential risk, reporting 34% of the US public now aware of the argument and, more to the author's point, that the rate of increase is itself accelerating rather than flattening. The relevance for policy timing is that awareness curves of this shape have historically preceded legislative attention by a short lag.

Clifford circuits underpin quantum error correction and fault-tolerant logical synthesis, and although they are efficiently simulable and representable as symplectic matrices, standard synthesis such as Aaronson-Gottesman produces circuits with excessive gate counts. AlphaClifford applies model-based reinforcement learning to synthesis and transpilation, searching for shorter equivalent circuits than the algorithmic construction yields.

Pixel-space generative models avoid lossy latent compression at the cost of having to learn global structure and fine detail jointly in a high-dimensional space. Standard flow matching interpolates noise toward a fixed clean endpoint and leaves spectral evolution implicit; energy-guided flow matching writes the coarse-to-fine trajectory into the objective so low frequencies resolve before high ones by construction.

Training terminal agents needs scalable executable supervision, and each synthesized task couples an instruction, an initialized environment, a reference solution and an executable verifier. When those four artifacts are generated under inconsistent assumptions the task ends up unsolvable or wrongly scored, and multi-stage synthesis pipelines tend to discard the original goal along the way. FACET carries source intent and executable state through the pipeline so the four artifacts stay mutually consistent.

Google announced AI-generated interactive visuals, 3D simulations, a dedicated student hub and customized practice quizzes across Search and Gemini. Searching a topic such as pH scale returns an interactive visual inside an AI Overview, and a follow-up question generates a tailored experience; users can also prompt Search directly for practice quizzes on any subject. Lens in Search will roll out an interactive problem-working experience in coming weeks.

Models are increasingly paired with verifiers, from step checkers and self-consistency filters to tool-based fact checkers and proof assistants, and the literature uses level to mean at least five different things: verification granularity, concept abstraction, risk tier, system-stack layer, and the epistemic source of ground truth. The paper proposes an L0-to-L5 verification autonomy scale to separate them, which is a housekeeping contribution but a load-bearing one for anyone comparing verifier claims across papers.

Institutional Books: Harvard Library, released in 2025, holds 983,004 volumes and 242 billion o200k_base tokens digitized through the Google Books Library project. As people began using it, a tension surfaced between standard large-scale preprocessing, which optimizes for throughput, and careful information stewardship. The new customizable multilingual open-source pipeline lets downstream users make those tradeoffs explicitly rather than inheriting them. A companion Institutional Newspapers Pipeline, designed with Boston Public Library, extracts structured datasets from historical newspaper scans.

Autoformalization is usually framed as translating natural-language mathematics into Lean 4, but faithful formalization needs more than translation: statements must map onto the type and definition hierarchy of libraries such as Mathlib while preserving the meaning of the source proposition. MathForm couples knowledge retrieval over the library with verification, so the model grounds each concept in an existing definition instead of inventing a plausible-looking one.

Memory systems and their benchmarks assume a single-source textual history, while real deployments scatter relevant information across applications, devices, users and time. MemFuse targets multi-source fusion, integrating dispersed observations into coherent episodic memories while preserving provenance for each fragment, which is what makes the resulting memory auditable rather than merely consolidated.

Relativity Networks announced $22 million in SAFE note funding led by Rhapsody Venture Partners, Bell Ventures and Faster Than Glass, plus a $40 million follow-on order from an unnamed hyperscaler. Hollow-core fiber sends light through a vacuum channel rather than glass, moving it closer to the free-space speed limit: CEO Jason Eichenholz puts conventional fiber at roughly five microseconds per kilometer against three and a half for hollow-core. With data-center spending projected as high as $4 trillion by decade's end and siting already constrained by power and politics, a 30% latency reduction changes which locations remain viable for training clusters.

Language-model agents can communicate through continuous hidden states invisible in public transcripts, which creates a channel for coordination that transcript-level monitoring cannot see. Verifiable Latent Alignments is an activation-aware framework that links the private latent-state record and channel status for every monitored decision to the resulting public action, so the audit trail covers the hidden channel rather than only the visible one, and supports steering as well as detection.

Ben Thompson covers Apple's settlement with the European Union over App Store fees alongside changes to App Tracking Transparency rules in Germany, arguing the App Store is finally confronting structurally lower fees and that the EU should count the outcome a success even arriving late. The relevance to AI distribution is the precedent for platform gatekeeping over model-backed apps.

Autoregressive perception models trained to localize visual entities under open vocabulary are usually fit with supervised fine-tuning under maximum likelihood, which optimizes per-token cross-entropy, a proxy fundamentally misaligned with precision and recall. The paper post-trains with GRPO against the perception metrics directly, in the dense-scene regime where the misalignment costs the most.

Artificial Analysis logged a new language model evaluation for G9v3-39A5B on August 19, alongside a feature launch the same day. The index currently has Claude Opus 5 at 63 and Claude Fable 5 at 62, with GPT-5.6 Sol and Grok 4.6 at 61, Kimi K3 and GLM-5.3 at 60, and DeepSeek V4 Pro at 53; cost per index task spans $0.05 for GPT-5.6 Luna to $3.14 for Fable 5.

Alexa+ rolls out to all compatible U.S. Fire TV devices at no cost regardless of Prime membership, bringing conversational search, smart home control and AI recommendations. It previously cost $19.99 per month without Prime. Upgrades are automatic with no app download or signup, covering current-generation Fire TV Sticks, the Fire TV Cube, Amazon Ember smart TVs and Alexa+-equipped Hisense and Panasonic sets. Amazon says Alexa+ users have nearly twice as many conversations on Fire TV as with the original Alexa.

LangChain shipped Tuned Evaluators in LangSmith, which analyze agent interactions in production and attach graded feedback that teams can route into improvement workflows. The first evaluator is Perceived Error, meaning the model's judgment of whether the user experienced the interaction as failing, which is a different and often more actionable signal than task-level correctness for agents whose success criteria are not verifiable.

GitHub's developer advocacy team published a beginner-oriented guide to the Copilot app covering work management: tracking assigned items, triaging, and moving tasks through a board from inside the assistant rather than the web interface. It is documentation rather than a capability change, but it marks Copilot's product surface expanding past code completion into project workflow.