Wolf Digest — Tuesday, July 14, 2026

The Marine Corps' Project Dynamis, its contribution to joint all-domain command and control, will evaluate Ditto's technology for turning radios, phones and drones into a peer-to-peer data mesh that keeps AI tools running locally when cloud access is denied. Ditto's claim is architectural: no server-client model at all, running over the Bluetooth and Wi-Fi radios already in consumer phones with no new hardware. The problem is real — transformers hallucinate the missing pieces when their data goes away, which is the basis of Anthropic's own objection to battlefield deployment and OpenAI's March statement that its military contracts exclude fully autonomous weapons because those require edge inference. What the public Dynamis documents still do not say is whether cloud connectivity itself has ever actually been cut during testing.

Murphy-Hill, Butler and Savelieva studied tens of thousands of engineers across Microsoft's early-2026 rollout of Claude Code and GitHub Copilot CLI. First use spread primarily through social networks rather than mandate or demographic segment; retention tracked baseline coding activity, not demographics; and adopters merged roughly 24 percent more pull requests than they otherwise would have, with the lift persisting across the full four-month window rather than decaying as novelty. The authors flag the proxy limitation themselves — a merged PR is not the value it delivers — but the persistence is what resists the easy dismissal. At Microsoft's scale, token spend runs into millions annually, so misreading adoption or impact makes a rollout expensive without moving velocity.

Arvind Narayanan released the annotated slides and transcript of his ICML keynote in Seoul. Three arguments: the AI-as-Normal-Technology framing holds unless and until a discontinuity such as recursive self-improvement arrives; even taking recursive self-improvement seriously, there is no single lab milestone whose achievement suddenly ends work; and future jobs will be radically different, requiring substantial adaptation. He grounds it in his Princeton group's work on agent evaluation, whose whole point is that benchmark capability is a poor predictor of real-world deployment, and closes on a picture of human-AI "co-superintelligence."

Playcode measured 16 fixtures through Anthropic's count_tokens endpoint and found the new tokenizer used by Sonnet 5, Opus 4.8 and Fable 5 emits about 32 percent more tokens than the prior one on the same content at unchanged list prices — TypeScript +31 percent, Rust +29, Python +23, English prose +34, Chinese roughly flat. Billing confirms it: identical content cost 2,541 input tokens on Opus 4.6 versus 3,191 on Opus 4.8. Against GPT's o200k as 1.00x, Claude now runs 1.36x to 1.73x, so Opus 4.8's $5/$25 list bills like $7.50/$37.50 effective, while GPT-5.6 Sol, Gemini 3 Flash and Grok 4.5 stay near list. Input tokenization only; cache pricing scales identically.

After Anthropic and Bun published a retrospective on Bun's agentic rewrite from Zig to unsafe Rust, Zig creator Andrew Kelley posted a blunt rebuttal attributing Bun's memory bugs to engineering culture — heavy AI-agent authorship and review, no meetings, no style guide — rather than to Zig. Ray Myers, writing at 1,454 points on Hacker News, sides with Kelley and contrasts Bun with TigerBeetle, a Zig financial-transaction database whose TigerStyle discipline (static allocation at startup, no dynamic alloc or free afterward) avoids the same class of bug without changing languages. He notes Bun's founder says the team hasn't typed code themselves in months, counts roughly four memory-bug fixes per week pre-rewrite, and points at DARPA's TRACTOR program as the serious prior art on C-to-Rust translation. Myers discloses he was chief architect at a coding-agent startup.

Existing terminal benchmarks grade only the final outcome on tasks that finish in minutes, which produces sparse reward and hides partial progress. Long-Horizon-Terminal-Bench introduces 46 tasks across nine categories — experiment reproduction, software engineering, multimodal analysis, interactive games, scientific computing — each decomposed into finely graded subtasks so agents earn dense intermediate credit. Tasks typically require hundreds of episodes and minutes to hours of wall-clock execution, stressing long-context management and iterative debugging rather than one-shot solving. Fifteen frontier models were evaluated.

The claim is that large-scale text-to-video generation is the vision analogue of next-token prediction: it supplies spatiotemporal priors, vision-language alignment, and scalability in one pretraining objective. GenCeption turns a pretrained video diffusion backbone into a feed-forward perception model steered by text instructions, and reports state-of-the-art results across depth, surface-normal and camera-pose estimation, expression-referring segmentation, and 3D keypoint prediction — matching or beating specialists including DepthAnything3, SAM3, VGGT-Omega and Sapiens. Under matched settings the video-generative backbone also outperforms V-JEPA and VideoMAE pretraining.

Distributional RL agents learn full return distributions that are increasingly read at face value — for interpretability, risk-sensitive control, and safety monitoring. This paper asks whether those risk claims are actually true, and audits them properly: a decision-relevant screening metric (the excess Wasserstein gap between the top two actions, which equals the mass by which first-order stochastic dominance is violated), ground truth from snapshot-restart Monte Carlo, and a statistical harness with permutation nulls, bootstrap refutation and FDR control — without which, they note, the audit itself manufactures false conclusions. Across QR-DQN, C51 and IQN on MinAtar over 33 runs, 40 to 95 percent of the strongest claimed risk trade-offs are refuted at 95 percent confidence, the placement of the strongest claims is statistically indistinguishable from truth-blind, and essentially no claim is confirmable. The learned "risk" reflects a training artifact rather than environment stochasticity. This is the most quietly damaging result of the day for anyone using distributional value heads as a safety signal.

Ottawa signed four agreements in June for the Arctic Over-The-Horizon Radar, sourced from Australia and worth about 1.75 billion US dollars — an export of the JORN lineage. The gap it fills is stark: Gordon Frazer, a former JORN engineer, says the current state is that a jumbo jet could fly from Beijing to Canberra with transponders off and go unnoticed, and that the same is likely true across Canada's north. HF over-the-horizon radar is a sensing problem where modern signal processing and learned clutter rejection have a large and under-covered role.

Artificial Analysis posted a new article comparing GPT-5.6 Sol, Terra and Luna on intelligence versus cost, and shipped Intelligence Index v4.1 (updating GDPval-AA v2, tau-cubed-Banking and Terminal-Bench v2.1) plus a new long-horizon knowledge-work benchmark, AA-Briefcase. Headline standings: Claude Fable 5 leads the Intelligence Index at 60 with GPT-5.6 Sol (max) at 59, but Fable's cost per task is $2.75 against Sol's $1.04. On the Coding Agent Index, Codex with GPT-5.6 Sol (max) tops at 80, with Claude Code on Fable 5 and Codex on Terra tied at 77 and Claude Code on Opus 4.8 at 73 — the first independent leaderboard where the Codex harness edges Claude Code at the top.

Nous Research is finalizing at least $75 million at a $1.5 billion valuation, led by Robot Ventures with USV participating, against $70 million raised previously. Its open-source Hermes agent — launched weeks after Openclaw went viral — differentiates on built-in skills (web search, coding, image understanding) and auto-learning new skills from usage without manual authoring. It runs locally or on a VPS, is driven via Telegram and Discord, and has roughly 214,000 GitHub stars and nearly 40,000 forks, with a hosted tier at $20 to $200 per month.

Standard pretraining converts visually rich sources — figures, typeset equations, page layouts — into plain text, discarding the layout signal. This paper trains directly on the visual documents instead and reports that, across multiple backbones and benchmarks, visual pretraining on the same underlying corpus consistently outperforms text-only pretraining. If it replicates, the implication is uncomfortable for every pipeline that treats PDF-to-text extraction as a lossless preprocessing step.

Singapore-based PixVerse closed a $439 million Series C including an extension backed by Alibaba, Mirae Asset and others, past a $2 billion valuation. Its line spans consumer video generation, a professional film tier, and R-Series world models for game development, generating up to 4K with baked-in audio at $4.80 per minute image-to-video, on claimed 150 million registered and 15 million monthly active users with 150 staff. Co-founder Jaden Xie attributes the edge to data-labeling methodology rather than data volume, drawing on co-founder Wang Changhu's ByteDance recommendation work, and claims OpenAI exited video generation when Sora 2 shut down.

Nearly all LLM-as-judge bias work is input-output: perturb the input, measure the score delta, propose a prompt fix. This paper gives the representation-level account instead, across seven judges, seven bias types and nine benchmarks. Geometrically, baseline judging inputs occupy a tight activation manifold while biased inputs are displaced along a low-dimensional, bias-type-specific subspace that sharpens with depth and is recovered consistently by three independent families of estimators. Causally, steering hidden states along that subspace drives scoring in both directions — forward shifts reproduce biased scoring on clean inputs, reverse shifts restore baseline scoring on biased ones — while matched-norm random directions move scores an order of magnitude less. Operationally, a simple linear projection onto the same bias-direction features anticipates judge failures. That last part is the payoff: it turns judge bias from a prompt-hygiene problem into something you can detect and cancel at inference.

After Musk called him a scammer, Altman replied that Musk is the one selling public-market investors on short-term space data centers. TechCrunch's point is that Altman's jab tracks the expert consensus: rival space-compute founders, Google's orbital-compute team, and independent engineers all conclude the economics don't close until launch gets much cheaper and high-power satellites can be mass-produced. SpaceX's orbital-inference plan is described as the main driver of its two-trillion-dollar valuation, yet SpaceX conceded on its IPO roadshow that Starship may need to expend second stages near-term. Starship's thirteenth test flight is expected as soon as July 16; at-scale orbital compute is characterized as a 2030s question.

GPTQ and its descendants build the quantization objective purely from input-activation statistics, implicitly assuming every output channel contributes equally to layerwise reconstruction. KronQ drops that assumption by folding gradient covariance in under a Kronecker-factored Hessian approximation, which yields two things: bidirectional incoherence processing (extending the input-side random rotation to the output dimension) and a mixed-precision sensitivity metric derived from gradient and activation Hessian traces. The headline result is 2-bit weight-only quantization of LLaMA-3-70B, where GPTQ and GPTAQ diverge outright — over 2,000 perplexity on WikiText-2 — and KronQ does not.

RLVR is expensive to repeat on every new strong model because the target has to generate the rollouts. Direct-OPD runs RL on a small model where rollouts are cheap, then transfers not the post-RL teacher's policy — which mixes the RL gains with the small model's limitations — but the log-ratio between the post-RL teacher and its own pre-RL reference, treating that as a dense implicit reward applied on the strong student's own on-policy states. In effect the checkpoint pair encodes which actions RL made the weak model more or less likely to take, and only that delta is transferred. This reuses weak-model RL supervision without running sparse-reward RL on the target at all.

A survey attempting to fix the fact that "metacognition" is used loosely across a dozen unconnected LLM literatures. It taxonomizes the field into methods and benchmarks for measuring metacognitive ability, techniques for eliciting and improving it, and applications, then lays out open problems. Its practical value is as a map — the confidence-calibration, self-verification, self-correction and introspection threads have been developing without a shared vocabulary, and this is the first attempt to give them one.

Olympiad math is saturating, and existing advanced-math benchmarks grade on final-answer correctness, which says nothing about whether the reasoning was valid. ProverBench contributes 296 undergraduate-to-doctoral-qualifying-exam proof problems, paired with an automatic verification pipeline trained on large-scale expert annotations that returns both a correctness verdict and fine-grained error localization, with strong agreement against human experts on held-out trajectories. VerifierBench adds 888 model-generated proof trajectories with expert ground truth to test whether models can judge proof validity at all — which is the harder and more diagnostic of the two tasks.

On-policy distillation is notoriously high-variance and unstable. TOP-D dynamically constructs a proximal teacher, and the authors give a formal argument that this inherently controls gradient variance, with a global convergence analysis and a monotonic improvement bound. Empirically it improves stability, sample efficiency and final performance on mathematical reasoning, and it adds zero computational overhead — which is the property that makes it a plausible drop-in replacement rather than an alternative worth arguing about.

A hybrid Mamba-Transformer MoE activating 3B of 30B parameters per token, keeping the inference cache near-constant as context grows — the throughput argument for long-context, high-concurrency serving. Pretrained on roughly 27 trillion tokens with German deliberately up-weighted, it matches dense 14-to-27B models on aggregate English and German benchmarks, posts the best code aggregates in both languages among 17 open base models, and beats every European sovereign baseline including ones with far more active parameters. Among fully open models it edges Olmo 3 32B and Apertus 70B. Built end-to-end on Deutsche Telekom's German Industrial AI Cloud, with weights, intermediate checkpoints, per-source data accounting, hyperparameters and training code all to be released.

Models memorize injected facts quickly during finetuning and then fail to use them for downstream reasoning — the authors formalize this as the Knowing-Using Gap, with both an accuracy gap and a temporal lag between memorization and generalization. Using a self-patching intervention that identifies activation locations where relocating representations rescues failed generalization, they argue for a knowledge-circuit misalignment hypothesis: the memorized representation exists internally but is not routed to computation-effective layers. A simple heuristic derived from the diagnosis recovers 58 to 75 percent of the oracle headroom on generalization failures, cross-domain.

AI2's maritime-domain-awareness agent is structured as soul, skills and config: a system prompt with hard boundaries, markdown skills following the same agent-skills spec as Claude Code and Codex, and a swappable harness (OpenClaw) and model (currently Claude Opus 4.6). The design lesson is that raw API calls produced malformed pagination and geometry errors in early prototypes, so Shippy drives a purpose-built typed CLI writing to local JSON files instead. Isolation is per-user Kubernetes deployments with the user's own JWT injected at provision time. Evaluation uses expert-written weighted rubrics with an LLM judge scored against live data on every versioned build; the latest run showed Shippy holding its guardrails but overstepping into tactical recommendations on patrol planning, and once inventing a nonexistent CLI command.

VLMs and VLAs improved perception and action prediction, but long-horizon embodied agents still lack a runtime layer for reasoning, memory, verification and cross-embodiment execution. ABot-AgentOS sits above low-level controllers and supplies scene-conditioned planning, context-isolated skill execution, multi-stage verification and edge-cloud collaboration, backed by a Universal Multi-modal Graph Memory that turns dialogue, observations, spatial context and task traces into typed nodes and edges. A failure-driven self-evolution loop promotes diagnosed memory failures into gated runtime assets only on later evaluation splits, avoiding ground-truth leakage. EmbodiedWorldBench accompanies it: 16 scenes, four difficulty levels, 200-plus executable tasks with trace-grounded scoring.

Monolithic observation-to-action navigation policies suffer coordinate drift, handle long-tail semantics poorly, and are opaque. ABot-N1 decouples cognition from control: a slow vision-language reasoner runs explicit chain-of-thought and emits a pixel goal — a compact set of image-space anchor points that serve as a single interface across point-goal, object-goal, POI-goal, instruction-following and person-following tasks — and a fast action expert converts those anchors plus text cues into continuous waypoints at native control frequency. The pixel-goal interface is the contribution worth watching; it sidesteps the coordinate-frame problem that has dogged VLN.

A dispatch from ICML in Seoul, where the work on show centered on running and training models more efficiently — and where, per the author, researchers also spent the week discussing the possibility of AI automating their own jobs. Five themes dominated. Two incidental observations worth logging: OpenAI's booth drew crowds while Anthropic barely had a presence, and the conference closed just before Apple filed its suit against OpenAI's hardware team.

Test-time training on a long context is prohibitively expensive over the full input and actively harmful over randomly sampled spans, because most spans are irrelevant to the question — on LongBench-v2, random-span TTT degrades the base model while oracle-span TTT substantially improves it. S-TTT closes the gap by having the model first identify the evidence spans it should learn from, then applying the standard language-modeling objective only to those. The result is that TTT's usefulness turns out to hinge almost entirely on span selection, not on the adaptation procedure.

Chinese humanoid maker LimX Dynamics said Tuesday it closed a pre-IPO round at a 15 billion yuan (2.2 billion dollar) valuation, raising 200 million dollars from Chinese and international investors, led by IDG Capital. The pre-IPO framing is the signal: Chinese humanoid firms are moving toward public listings while their Western counterparts are still raising private rounds at comparable or higher marks.

SymCrypt is building new verified cryptography in Rust, proving that the code safely and correctly implements the standard algorithms — notably the post-quantum ones — using Aeneas to translate Rust into a functional model and Lean to discharge the proofs, with AI agents used to scale the proof effort. They are releasing the verified code, specifications, and properties. The AI-agent angle is the part worth tracking: formal verification of production crypto has always been gated on proof-engineering labor, and agents attacking that bottleneck is a more concrete AI-for-software result than most.

Pullen argues an inference-first company doesn't need billions to compete — millions, a national compute allocation (Cosine trains on the Isambard supercomputer in Bristol), and a consortium feedback loop suffice. The technical core: why open-weight models still trail on size, active parameters and data; why active parameters dominate how a model actually feels, more than the MoE-versus-dense framing suggests; and why real coding trajectories are the scarce asset. On trustworthiness he argues for rewarding the process rather than the final answer to beat slop, reframes code review as runtime proof (spin the bug up in a VM and make the agent exploit it), and describes Swarm, Cosine's system running hundreds of sub-agents in one shot. He reads US export controls as an accidental gift.

Reward optimization and distribution matching both couple policy exploration to distribution alignment, forcing expensive exploration onto the primary model and blocking asynchronous generation, reuse, or cross-model transfer of the resulting signal. PUST uses a lightweight proxy as the exploration testbed, extracts the relative improvement between the proxy's initial and optimized states, and transfers that directional update to the primary model. Because it transfers relative improvements rather than absolute policies, the signals can be generated asynchronously, cached, and reused across models — which is the practical payoff.

Continuously captures egocentric visual and audio streams, aligns them on a shared timeline, and organizes them into current, short-term and long-term tiers, routing retrieval to the appropriate level per query and grounding answers in multimodal evidence. It runs on smartphones and AI glasses, covering object finding, conversation recall, life summarization and routine discovery. The hierarchical tiering is the practical concession — a flat memory over an always-on sensor stream is not retrievable at wearable compute budgets. Code released.

Video motion transfer has been stuck on predefined human skeleton structures and skeleton-conditional training, which generalizes badly to animals of different species and is bottlenecked by scarce labeled multi-skeleton data. Motion4Motion abandons skeletons entirely and models the character's motion flow, which makes cross-species transfer tractable and requires no training at all. Reported to outperform baselines by a wide margin.

An open-sourced suite across K80, M10, M40, M60, P40, P100, V100 and T40 covering ResNet50 train and inference, Blender, ViT throughput, llama.cpp on Qwen2.5-1.5B and Llama 3 8B, Whisper Medium FP16, Folding@Home and gdsio. Findings worth knowing: a 16GB V100 (under $200) performs comparably to the newer, rarer T40; the P40 beats the P100 for LLM inference, confirming community consensus; the ~$50 M60 is a standout for Whisper throughput; three P100s hit 566.58 img/s on ResNet50 training against a 624.6 theoretical sum; and CPU single-core clock barely matters except for Whisper and ViT attention. llama.cpp notably failed to scale with added GPUs — an unresolved config issue.

TechCrunch takes the maximalist "total user alignment" position — that a model should serve its user's stated goals without imposing its own refusals — and pushes it to the case where the user's goal is homicide. The value of the piece is as a reductio: it forces the total-alignment camp to specify where the line actually sits, which is precisely the specification that camp has avoided writing down.

The Defense Department has paid out millions to personnel affected by anomalous health incidents — the first HAVANA Act payments under any administration — and renamed its cross-functional team the Directed Energy Bio-Effects CFT, now under Research and Engineering rather than Policy, with a mandate covering treatments and countermeasures. The department did not say how many individuals have been compensated. The reorganization signals a shift from attribution to defensive research.

Claude users in India are beginning to see rupee-denominated subscription plans. India being Anthropic's largest market after the United States is the fact worth carrying — purchasing-power-adjusted pricing in the largest developer markets is where the consumer-tier volume war actually gets fought, and Anthropic moving here suggests the elasticity is real.

Chief Product Officer Sachin Kansal walks through Uber's financial-services ambitions, its increasingly complicated relationship with Waymo, its new AV Labs data operation, and where AI surfaces for riders and drivers. AV Labs is the part to watch: an aggregator building its own autonomy data operation is positioning to stop being a pure demand channel for someone else's driver.

Waze ships new AI features and customization updates, several powered by Gemini — part of Google's broader push to fold the assistant into every surface it owns, and a competitive move against Apple Maps. Unremarkable on its own; useful as a marker of how quickly assistant integration has become table stakes in consumer navigation.