Wolf Digest — Tuesday, June 30, 2026

Jack Clark's newsletter leads with ENPIRE, an NVIDIA harness that puts physical robots through the same autonomous experiment-execute loop coding agents use, via four modules for environment reset and verification, policy improvement, rollout across single or parallel robots, and an evolution stage where coding agents read logs, consult the literature, and rewrite the training code. Each station pairs two bimanual arms with an RTX 5090; agents reached 99 percent success on dexterous tasks including inserting a GPU into a motherboard, with GPT-5.5 in Codex and Opus 4.7 in Claude Code trading the lead while Kimi-2.6 lagged. A second item flags Tencent's ARGUS tracing system running on a production cluster exceeding ten thousand GPUs for over six months, including a 12,960-GPU mixture-of-experts training job read as a technosignature of Tencent's training maturity. Clark pairs the technical notes with an essay on automation and labor.

Overland AI secured a 20-million-dollar production contract to supply autonomous ground vehicles that will resupply the Marine Air Defense Integrated System, the Marine Corps' counter-drone air-defense platform. The deal, structured as an Other Transaction agreement under the Pentagon's APFIT initiative, covers the vehicle hardware plus the company's OverDrive autonomy stack and OverWatch command-and-control, with initial deliveries beginning roughly nine months after award and more than a dozen vehicles planned. Chief executive Byron Boots said the fully autonomous vehicles, which already integrate around thirty payloads within size, weight, and power limits, are less susceptible to contested communications because they can keep operating after losing connectivity, and that a single operator can task several at once. The vehicles supplement rather than replace the system's crewed Joint Light Tactical Vehicles.

Memora is a scalable agentic-memory framework that resolves the abstraction-versus-specificity tension that traps prior systems. The trick is to decouple what is stored from how it is found: each entry carries a six-to-eight-word primary abstraction, the only part embedded for similarity search, alongside a rich memory value that is never retrieved by its own content, so new information about an evolving topic merges into one entry instead of fragmenting the way RAG and Mem0 do. Short cue anchors provide alternative access paths, and a policy-guided retriever treats retrieval as active reasoning, refining queries, expanding through cue anchors, and deciding when to stop. The reported results are a new state of the art of 86.3 percent LLM-judge accuracy on LoCoMo and 87.4 percent on LongMemEval, with the largest gains on multi-hop reasoning, while using about half as many entries as Mem0 and up to 98 percent fewer tokens than full context.

Ai2's DiScoFormer is a transformer that, given a finite sample, estimates both the density and the score, the gradient of the log-density, of the underlying distribution in one forward pass with no per-distribution retraining. It is best understood as a learned generalization of kernel density estimation: the authors show a single attention head's weights approximate a Gaussian kernel over the data, so one cross-attention block already reproduces KDE, except attention learns several bandwidths at once instead of one fixed scale. A label-free consistency loss, the residual between the score head and the gradient of the log-density head, also enables test-time adaptation on out-of-distribution inputs. Trained on freshly sampled Gaussian mixtures per batch, in one hundred dimensions it cuts score error roughly six-and-a-half-fold and density error more than thirty-seven-fold against the best hand-tuned KDE, and keeps improving with more samples. The team pitches it as a drop-in for diffusion models, Bayesian sampling, and scientific simulation.

OSWorld 2.0 targets a gap in computer-use evaluation: existing benchmarks do not capture the realism, complexity, and long-horizon demands of actual desktop work, which limits what they can reveal about frontier agents. The benchmark assembles 108 long-horizon computer-use workflows spanning everyday and professional tasks, designed so that completing them requires sustained multi-step operation across real applications rather than short, scripted interactions. By stressing duration and interdependence between steps, it is built to surface the failure modes, error recovery, state tracking, and drift over long sessions, that shorter computer-use suites tend to hide.

California and Anthropic agreed a deal giving every state agency and local government access to Claude at a half-price discounted rate, bundled with training and support, aimed at helping public employees draft documents and analyze information. Governor Gavin Newsom said the technology should help government workers move faster and solve problems rather than replace them, and the agreement follows his March executive order directing faster, safety-conscious government adoption of AI. The arrangement runs alongside a separate federal dynamic: Anthropic and the Defense Department did not reach terms on a contract that would have allowed broad deployment of Claude, with Anthropic seeking limits around domestic surveillance and autonomous-weapons use, after which the department signed with OpenAI and designated Anthropic a supply-chain risk. California's technology office said that designation did not factor into its own negotiations.

Base44, the vibe-coding platform Wix acquired for 80 million dollars about a year ago, is rolling out its own model, a first iteration named Base1, trained on a dataset generated from tens of millions of real user interactions. Founder Maor Shlomo frames the move as a bet on owning the stack: controlling the model unlocks latency, cost, and efficiency optimizations, with the goal of eventually being faster and cheaper for customers than calling frontier models like Opus. Investors quoted in the piece treat proprietary data as one of three defensibility ingredients alongside distribution and tech stack, but caution against underestimating frontier labs, noting that legal-tech startup Harvey abandoned plans to train its own model. The competitive backdrop is frontier labs encroaching on the application layer, with Cursor and xAI now both inside SpaceX and Claude Code itself becoming a vibe-coding contender. Base44 says owning the model should produce a structurally stronger margin profile; it recently passed 100 million dollars in annual recurring revenue.

Modern LLMs lean on RL during post-training to push individual capabilities, but integrating several capabilities into one model remains hard, and existing approaches like off-policy fine-tuning and mixed-RL are either inefficient or lose performance. MOPD proposes multi-teacher on-policy distillation: train specialist teachers per capability, then distill them into a single student on the student's own rollouts so the transfer stays on-policy. The framing matters because it is one of several on-policy-distillation papers landing the same day, part of a visible push to compress specialist behavior into deployable generalists without the systems cost of serving every teacher at inference time.

Cursor launched Cursor Mobile, letting developers prompt and oversee coding agents from a phone, spin up new agents, or take over agents started on the desktop client. It extends the Cursor 2.0 shift toward independent agents and follows comparable mobile apps from Anthropic and OpenAI, reflecting a broader move from writing code to supervising code-writing agents; Anthropic's Claude Code lead, Boris Cherny, recently said most of his coding now happens on his phone. TechCrunch notes the release comes despite the 60-billion-dollar SpaceX acquisition of Cursor, placing yet another coding tool under the same corporate roof as xAI.

Orca is presented as an initial instantiation of a general world foundation model. Rather than optimizing isolated next-token, next-frame, or next-action prediction, it learns a unified world latent space from multimodal world signals and exposes it through multimodal readout interfaces, organizing the whole system around next-state prediction. The bet is that a single shared latent over heterogeneous world signals, with task-specific readouts layered on top, generalizes better across perception and action than stacking modality-specific predictors. It is early-stage, but it stakes out a position in the increasingly crowded world-model space by making state, not any one output modality, the central training target.

Getting LLMs to self-improve on hard reasoning without external expert supervision is bottlenecked by the lack of mechanisms that track per-problem learning progress and adapt accordingly. DRIFT adds difficulty routing to self-distillation, gating exploration by a rhythm signal and maintaining a success buffer so the optimizer spends effort where the model is actually making headway rather than uniformly across problems. It sits in the same self-improvement and on-policy-distillation cluster as several other papers today, and its contribution is the explicit, difficulty-aware scheduling of when to explore versus consolidate.

Foundation models in language and multimodality get strong generalization by aligning heterogeneous data under a unified formulation and training at scale, and Qwen-RobotManip investigates whether that recipe transfers to robotic manipulation. The challenge is that, unlike text, manipulation data is fragmented across embodiments, sensors, and action spaces. The report's claim is that alignment, putting heterogeneous manipulation data into a unified formulation, is what unlocks the scaling behavior, letting a single model approach genuine cross-task generalization rather than overfitting to narrow demonstration sets. It is one of two Qwen robotics technical reports circulating today, alongside a navigation-focused counterpart.

Distillation and RL fine-tuning are the two pillars of diffusion post-training, but they are usually studied in isolation, leaving the interaction between them, and how fine-tuning affects the generative quality of an already-distilled model, poorly understood. This paper introduces Rewarded Moment Matching Distillation to treat the two phases jointly, using a reward signal inside a moment-matching distillation objective so that the distilled student inherits both the speed of distillation and the preference alignment of RL fine-tuning. The contribution is a cleaner account of how to order and combine the phases without the quality regressions that arise when they are bolted together naively.

As harness frameworks mature, terminal-operating agents are increasingly capable of general computer-use tasks well beyond writing code, yet existing benchmarks do not adequately measure general-purpose terminal-use agents. TUA-Bench targets that gap, evaluating agents on a broad range of terminal tasks rather than the coding-centric workloads most existing suites assume. The point is to separate competence at general command-line operation, file manipulation, system tasks, multi-step shell workflows, from the narrower code-generation skill that dominates current agent evaluation.

Hybrid attention models improve long-context efficiency by keeping only a subset of full-attention layers and replacing the rest with linear attention, but the effectiveness of this Transformer-to-hybrid conversion hinges critically on which layers retain full attention, and existing layer-selection heuristics are crude. This work studies the selection problem directly, proposing a more principled way to decide which layers to preserve when morphing a pretrained Transformer into a hybrid, so the converted model keeps the quality that depends on full attention while capturing the efficiency of linear attention elsewhere.

Frontier software-engineering benchmarks typically hand an agent complete requirements upfront and grade autonomous implementation, which misses how real development works. SWE-Interact places coding agents in a multi-turn, interactive, user-driven workflow where requirements arrive incrementally and the agent must ask, adjust, and iterate across a session. The testbed is built to expose the conversational and clarification skills that one-shot SWE benchmarks ignore, and to measure whether agents that score well on autonomous implementation actually hold up when a human is steering the task turn by turn.

Coding agents are becoming a major application of agentic LLMs, but serving them efficiently is hard and the data needed to study real workload patterns is largely missing, since public traces and benchmarks do not capture day-to-day coding-agent usage. TraceLab characterizes those workloads directly, providing traces and analysis of how coding agents actually hit a serving stack, request shapes, context growth, tool-call patterns, latency-sensitive segments, so inference systems can be tuned to the real distribution rather than to synthetic benchmarks. It is infrastructure-flavored work aimed at the serving side of the agent boom.

Reinforcement learning with verifiable rewards is bottlenecked by hard prompts whose correct trajectories have such low probability that sampling misses them within a practical budget, leaving the policy update with little signal. LatentRevise frames these zero-hit prompts as RLVR's sampling frontier, the place where genuinely new reasoning behavior is most needed, and proposes learning from them via latent revision rather than discarding them. The method aims to extract usable gradient signal from prompts that standard rollouts simply never solve, which is exactly the regime where capability gains on hard reasoning tasks tend to stall.

On-policy distillation trains a student on its own rollouts under teacher feedback and is becoming central to LLM post-training, but like RL it faces an on-policy systems bottleneck because rollout generation can dominate training time for reasoning workloads. AsyncOPD studies asynchronous on-policy distillation, deliberately letting the student's rollouts go stale relative to the current parameters, and characterizes how much staleness the method tolerates before quality degrades. The practical payoff is throughput: if moderate staleness is harmless, rollouts and updates can be decoupled and pipelined, easing the bottleneck that makes on-policy methods expensive.

Streaming video editing is held back by two coupled problems: keeping backgrounds and non-edited regions stable over time, and hitting the low latency that interactive use requires, while recent streaming-generation methods were built for synthesis rather than editing. LiveEdit targets real-time diffusion-based streaming video editing directly, aiming to preserve untouched regions frame to frame while applying edits at interactive latency. The work sits at the practical edge of video diffusion, where the research challenge is less raw quality than temporal stability and speed under a streaming constraint.

Proception, founded by former Tesla Optimus technical lead Jay Li, settled the trade-secret lawsuit Tesla filed against it last year, and Tesla dismissed the case this month. The company announced an 11-million-dollar seed round led by First Round Capital, with Y Combinator and BoxGroup, and is shipping its first high-dexterity robot hand, 22 degrees of freedom with multiple joints per finger, to researchers and robotics firms. Its differentiator is data collection: instead of VR-teleoperation rigs that lack tactile feedback and are bottlenecked by robot availability, Proception uses a sensor-laden glove to capture human hand-interaction data without a robot in the loop, and the same sensor skin covers the hand itself. Dexterous manipulation is widely described as one of robotics' hardest unsolved problems.

NVIDIA describes a new Palantir intelligent engine that runs open-weight NVIDIA Nemotron models inside closed, secured environments for US government agencies, pitched as a way to get open-model flexibility without exposing data or model weights outside an accredited boundary. The framing, open models in closed environments, is the operative idea: agencies get to use and adapt open-weight models on sensitive workloads while keeping deployment inside controlled infrastructure. It extends the pattern of defense-and-intelligence AI being packaged with governance and isolation guarantees rather than offered as a general cloud endpoint.

TIDAL introduced a policy that bars fully AI-generated music from being monetized and will use automated tools to remove AI tracks that impersonate a specific artist or group. Tracks judged 100 percent AI get an AI badge and cannot collect royalties, be monetized, or be sold directly to fans. TIDAL joins Spotify, Apple Music, Deezer, and Qobuz in labeling or filtering AI music; Deezer, which has gone furthest, says 44 percent of all new music uploaded daily is now AI-generated and offers its detection technology to rivals. The move marks streaming platforms converging on disclosure-and-demonetization as the default response to generative audio flooding catalogs.

Google is extending Gemini's personalized AI image generation to eligible free users in the United States, letting the assistant create images informed by a user's interests and data drawn from connected Google apps. The personalization hook, generating imagery conditioned on signals pulled from a user's own Google account, is the notable part, both as a product differentiator and as a data-use question. The expansion to the free tier is a straightforward distribution play in the increasingly commoditized consumer image-generation market, where the competitive edge is shifting from raw image quality toward personalization and integration.

Many modern generative methods, diffusion models, normalizing flows, and flow matching, estimate transport maps or plans between distributions without explicitly targeting an optimal-transport map, since in generative modeling the transport cost itself is irrelevant. This paper asks what the fundamental statistical limits of valid transport-map estimation are in that setting, characterizing how accurately a transport map can be recovered from finite samples when validity, rather than optimality, is the target. The result is theoretical grounding for a class of methods that are usually justified empirically, clarifying what is and is not achievable when estimating maps between distributions.

This work argues that long LLM reasoning traces can be segmented into recurring cognitive episodes, identifiable units of reasoning behavior, and that doing so enables interpretable analysis of how a model actually reasons. Rather than treating a chain of thought as an undifferentiated token stream, the authors identify episode boundaries and types, giving a structured handle on which reasoning moves a model deploys, when, and how they compose into a solution. The framing turns opaque reasoning traces into something that can be inspected and diagnosed at the level of cognitive operations rather than raw tokens.

MuonSSM brings orthogonalization, in the spirit of the Muon optimizer's orthogonal updates, to state space models, aiming for more stable and better-conditioned sequence modeling. State space models can suffer from conditioning and stability issues in their recurrent dynamics; imposing orthogonality constraints is meant to keep the state transitions well-behaved over long sequences. The contribution is a concrete recipe for orthogonalizing the SSM parameterization, positioned as a way to improve training stability and long-range behavior without abandoning the linear-time advantages that make state space models attractive.

LeVo 2 is a song-generation model that uses a hierarchical representation to improve stability and musicality, targeting the failure modes, drift, incoherence, and inconsistent structure, that plague long-form generated songs. By modeling music hierarchically rather than as a flat sequence, the system aims to hold melody and structure together over a full song while keeping output melodious. It lands in the active text-to-music space, where the research frontier has moved from short clips toward coherent, full-length compositions.

Qwen-RobotNav is the navigation-focused companion to the day's Qwen-RobotManip report, applying the same unify-and-scale philosophy to robot navigation. The claim is a scalable navigation model that generalizes across environments by aligning heterogeneous navigation data under a common formulation and training at scale, rather than fitting narrow, environment-specific policies. Together the two Qwen robotics reports signal a coordinated push to bring language-and-multimodal foundation-model recipes, unified data formulations and large-scale training, into both manipulation and navigation.

This paper evaluates DNA language models with a focus on what pretraining buys for downstream fine-tuning, asking whether large-scale self-supervised pretraining on genomic sequence delivers the gains that the language-model analogy promises. By systematically assessing pretrained versus less-pretrained genomic models across fine-tuning tasks, it probes how much of the performance is attributable to pretraining itself rather than to architecture or task-specific data. The result is a more sober accounting of where the genomic-foundation-model paradigm helps and where its benefits are thinner than the hype suggests.

NVIDIA reports that Firefly Aerospace operated an NVIDIA Jetson edge-compute module in lunar orbit for the first time, putting GPU-class onboard inference into a cislunar mission profile. The milestone is incremental but concrete: edge AI accelerators rated for space and run in lunar orbit point toward spacecraft doing more perception and autonomy onboard rather than round-tripping everything to Earth. It is a small data point in the broader push to move capable inference hardware to the literal edge, where latency and link constraints make local compute valuable.