Wolf Digest — Tuesday, June 23, 2026

Reflection AI, an open-source AI lab, will pay SpaceX 150 million dollars a month from July 1, 2026 through 2029 for immediate access to Nvidia's latest GB300 accelerators and supporting hardware in SpaceX's Colossus 2 data center near Memphis, Tennessee. The arrangement underscores how SpaceX's Colossus build-out is being monetized as a neocloud, selling frontier GPU capacity to outside labs, and how an open-weights lab is willing to commit roughly 5.4 billion dollars over the term to secure GB300 supply rather than wait in line for scarce capacity.

MIT Technology Review recaps the state of Anthropic's dispute with the federal government. By its account, Anthropic said in April it had built a model called Mythos that was strong enough at working with code to pose a global cybersecurity threat, and gave access to a small group of cybersecurity experts under restricted terms, after which the government moved to control the technology. The piece lays out the points to watch as the situation develops, including how access is governed, how export controls on frontier security models are scoped, and how other labs respond. Reported here as the policy backdrop to OpenAI's Daybreak launch and the day's red-teaming discussion.

PhoneBuddy is a training recipe and open-model line for agents that operate real phones. The core difficulty is that the deployment environment, real devices running real apps, is slow, stateful, side-effectful, and hard to reset or verify, while scalable mock environments only approximate it. The authors pair a real-app environment with PhoneWorld, a mock environment that reconstructs runnable mock apps from real GUI usage structure, build a shared supervised fine-tuning stage from both, then compare real-app RL against mixed RL across both. Across a 150-task human evaluation on real phones spanning apps, mini-apps, and cross-app workflows, success rates improve materially over the supervised baseline.

PlanBench-XL is an interactive benchmark of 327 retail tasks over 1,665 tools that tests whether agents can iteratively retrieve usable tools, invoke them to uncover intermediate evidence, and chain those results toward a goal under retrieval-limited tool visibility. An optional blocking mechanism injects missing, failing, or distracting tool functions, forcing agents to detect disrupted paths and adapt at runtime. Across ten leading models, massive-tool planning remains hard: GPT-5.4 achieves 51.9 percent, indicating that discovering and composing the right tools from a large ecosystem, rather than reasoning over a fixed toolset, is a live bottleneck.

Long agent traces of chain-of-thought and tool calls accumulate stale content that anchors later generations and eventually overruns the context window. Standard scaffolds compact at fixed token thresholds, which ignores trajectory structure and risks discarding partial results mid-derivation. SelfCompact instead lets the model decide when and how to compact, pairing a compaction tool the model invokes to summarize accumulated context with a lightweight rubric specifying when to fire (a sub-task resolved, the trajectory converging) and when to suppress (mid-derivation, or when stuck). The authors show both elements are needed: the tool alone is used unevenly across open-weight models, often at unhelpful moments, until the rubric guides its timing.

AI chipmaker Groq confirmed a 650 million dollar raise and is hiring new executives as it leans into its neocloud business, selling inference capacity built on its LPU inference accelerators. The move follows an Nvidia 20 billion dollar arrangement that TechCrunch characterizes as a not-acqui-hire, and positions Groq to re-staff and compete on low-latency inference serving rather than purely on chip sales.

This paper tests whether a model can recognize that its own prior response was elicited by an adversarial prefill attack. Across ten open-weight instruction-tuned models from 3B to 70B and four safety benchmarks, no model reliably recognizes its own compromised outputs, claiming genuine intent on prefilled responses at an average rate of 27.3 percent. The introspective signal stems largely from safety and refusal-related reasoning: orthogonalizing model weights against the refusal direction collapses the gap between claiming rates on prefilled versus natural outputs to near zero, though that direction is not the unique mediator. The signal is also probe-dependent, with internal-intention versus external-tampering framings eliciting qualitatively different answers.

Google DeepMind and the studio A24 are teaming up, with a reported 75 million dollar commitment, to build AI tools for filmmaking. The partnership pairs DeepMind's generative video and image models with a studio known for auteur-driven work, signaling a push to move generative media from demos into production pipelines and to shape how AI tooling is adopted by working filmmakers.

Group-based RL has improved LLM agents, and recent frameworks shifted from trajectory-level to step-level updates for finer credit assignment. But long-horizon agentic RL still suffers from sparse, delayed reward, and existing step-level methods treat exploration as isolated linear trajectories, ignoring the graph structure of state transitions and producing high-variance value estimates and myopic credit assignment. Group-Graph Policy Optimization (G2PO) models the shared graph of states across a group of rollouts, aggregating value signal over converging paths to deliver lower-variance estimates and less localized credit assignment for multi-turn agents.

Multi-turn tool-using agents must coordinate long tool sequences while tracking dialogue state and policy constraints, yet most approaches separate inference-time orchestration from parameter-level learning, leaving tool selection weakly structured and preference updates vulnerable to train-deployment prompt mismatch. ToolGraph combines schema-derived topology, transition weights estimated from successful rollouts, and history-aware controls for write prerequisites and repeated-search loops, then trains on divergence-point preferences, the steps where successful and unsuccessful trajectories first split, to sharpen tool selection for within-benchmark self-improvement.

World Action Models (WAMs) are embodied predictive-action models that make a forecast of the future available to action. Some recent WAMs repurpose large video-generation models; a parallel line uses language or vision-language backbones with no video core. This survey argues the rapid expansion has blurred the boundaries among broad world models, video-generation models, action-grounded video world models, vision-language-action policies, and WAMs, and gives the field a common account. It organizes existing work along two axes, what each method must generate (rendered futures, latent futures, or video-generation-free action reasoning) and a decomposition by predictive substrate, backbone, action coupling, and deployment regime.

Humanoid loco-manipulation is usually simplified into a stop-and-go process, walking to an object, stopping to manipulate, then resuming, and commonly relies on low-degree-of-freedom end effectors that act like an open-close grasp. CoorDex is a learning pipeline that converts high-dimensional body and dexterous-hand control into coordinated latent residual control, enabling high-DoF dexterous loco-manipulation while the robot is moving. Starting from simulated whole-body and hand demonstrations, it trains a coordinated controller that couples body and hand priors rather than treating them as separate stages.

Following the o3-style paradigm of interleaving code execution with reasoning, most multimodal work has focused on tool use within vision-perception tasks, relying on predefined heuristics for visual manipulation and remaining unable to handle numerical computation. AIR gives multimodal LLMs adaptive interleaved reasoning that decides when to call code versus reason in language, extending interleaved code reasoning beyond visual operations to numerical and mixed problems, and trains the policy so the model learns to invoke computation only when it helps.

Linear mode connectivity offers a foundation for understanding and merging independently trained networks, but existing methods optimize the interpolation path from a single endpoint, limiting scalability to large pretrained transformers. This work applies functionality-preserving weight transformations to align functionally equivalent solutions and introduces a dual procedure in which both models jointly learn their transformations toward a shared interpolation path. The bidirectional optimization substantially reduces interpolation barriers, enabling more reliable merging across billion-parameter architectures than one-sided alignment allows.

End-to-end OCR models such as DeepSeek-OCR put OCR back in the spotlight by using a language model as the decoder to leverage the prior distribution of language. The downside is that as output sequences lengthen, the accumulated KV cache drives up memory and progressively slows generation, unlike humans, whose reading efficiency does not decay with length. Unlimited OCR Works targets that decline directly, proposing mechanisms to keep memory and throughput roughly constant as documents grow so that long-document OCR stays efficient rather than degrading.

Multimodal agents repeatedly re-examine the same video frames, UI screenshots, and rendered artifacts as the context window slides and reasoning iterates, yet every look-back re-encodes from scratch because prefix caches only serve reuse at a fixed leading position. Kamera shows this recompute is avoidable and identifies exactly what naive KV reuse loses: the cross-chunk conditioning a chunk absorbs from its neighbors. That loss is asymmetric, and the direct readout of a cached chunk is recovered exactly and for free by a standard state-merge, enabling training-free reuse of multimodal KV state regardless of position.

Nathan Lambert revisits Z.ai's GLM-5.2, the MIT-licensed mixture-of-experts model released June 13, to argue its real significance is agentic rather than just leaderboard standing. His thesis is that GLM-5.2 is the first open-weights model to deliver a step change in reliable agentic and tool-use behavior, narrowing the gap to closed frontier systems for builders who need self-hostable agents. He situates it against the current open-versus-closed dynamics and ties it to his own companion release, the Tmax open RL recipe for terminal agents. Reported here as analysis of an already-covered model release rather than new launch news.

Meshes are among the most common 3D representations but are hard to generate directly because they carry strong symmetries, including permutation invariance over faces and over vertices within a face. MeshFlow generates triangle meshes directly as triangle soups, avoiding serialization into long autoregressive sequences, and uses equivariant optimal-transport flow matching that respects those permutation symmetries. The result is a generator aligned to the structure of mesh data rather than fighting it through sequence modeling.

AdamW is the de facto optimizer for training large language models, yet its theory lives mostly in finite-variance regimes, even though empirical evidence indicates stochastic gradient noise in LLM pretraining is typically heavy-tailed. Recent results show sign-based optimizers like Lion and Muon achieve sharp heavy-tailed rates and that AdaGrad can converge under heavy tails, but no rigorous convergence theory exists for AdamW in this regime. The paper poses this as an open problem and surveys what would be required to settle whether AdamW remains effective when gradient noise lacks finite variance.

Flow matching has become a standard for behavior cloning in robotic manipulation, but state-of-the-art flow-matching policies rely on a globally fixed isotropic source distribution that mismatches the fragmented, heteroscedastic structure of robotic action spaces, forcing the model to learn entangled vector fields and bottlenecking training. This work introduces latent-action guidance for the source distribution, shaping the starting noise by inferred latent actions so the learned flow is less entangled and the policy trains more efficiently and performs better on manipulation tasks.

Deep-research agents are LLM-based systems for autonomous multi-step scientific reasoning, with clear potential to accelerate the physical sciences, yet rigorous evaluation in this domain is lacking. PhySciBench introduces 200 expert-curated questions balanced between physics and chemistry across six task categories that reflect real research workflows, plus a multi-agent framework, giving the field a benchmark to measure how well autonomous research agents handle genuine physical-science problems rather than generic reasoning puzzles.

After months of testing between Anduril and Lockheed Martin, the US Army has selected Anduril to lead the common data layer baseline for its Next Generation Command and Control (NGC2) initiative, the effort to make the service's disparate systems interoperate. The decision boosts Anduril's role as a key software integrator on a flagship Army modernization program centered on moving data fluidly across the battlefield network.

Shield AI completed its acquisition of Aechelon Technology, a maker of high-fidelity, physics-based simulation and image generation used in flight training and synthetic environments. The deal, tied to a previously announced raise at a 12.7 billion dollar valuation, pairs Aechelon's simulation stack with Shield AI's autonomy software for AI pilots and uncrewed aircraft, strengthening the training-and-simulation pipeline behind its autonomous systems.

The Defense Intelligence Agency, with a workforce exceeding 16,000, issued a request for information on June 17 to inform a potential AI prototype project aimed at overcoming inefficiencies in its procurement enterprise. The RFI is an early step toward a possible launch, signaling the intelligence agency's interest in applying AI to back-office acquisition workflows rather than only to mission analysis.

In a June 17 request for information, the US Army said it is interested in a robust, ruggedized autonomous ground vehicle to recover broken or disabled platforms from contested environments, continuing its exploration of robots for dangerous battlefield tasks. The notice frames vehicle recovery as a candidate mission for ground autonomy, where removing crews from exposed recovery operations is the explicit motivation.

At the ISC conference in Hamburg, NVIDIA introduced software to speed AI for science spanning chemistry, materials discovery, and the search for dark matter. The releases include the DAQIRI library, new ALCHEMI NIM microservices for materials and chemistry workflows, and the cuPhoton reference code for photonics and experimental physics, extending NVIDIA's accelerated-computing stack from model training into domain-specific scientific simulation.

Three new Los Alamos National Laboratory supercomputers, named Mission, Vision, and Veritas and built with HPE and NVIDIA, will use NVIDIA Vera CPUs to accelerate scientific discovery, including agentic scientific-AI workloads. The systems extend the trend of national labs standing up large CPU-plus-accelerator machines aimed at coupling simulation with autonomous AI research agents rather than batch HPC alone.

Ben Lorica argues that the economics of AI data centers look increasingly weak and names them his leading candidate for what could pop the AI bubble within six to twelve months. The concern is not that AI stops improving or that demand vanishes, but that capital spending has raced far ahead of proven revenue while the financed assets, chiefly rapidly depreciating accelerators, lose value quickly. The piece is a financial-sustainability counterweight to the day's compute build-out news.

Jack Clark's Import AI 462 examines superpersuasion, the prospect of models that systematically out-argue humans, alongside discussion of self-sustaining AI systems and candidate technical paths toward artificial superintelligence. The newsletter synthesizes recent research and frames the safety and governance questions these trajectories raise for the field.

Recent attempts to combine LLMs with causal discovery have them infer pairwise directions, propose graph structures, or inject model outputs as priors, but these can obscure whether a causal claim is supported by data and assumptions or merely by textual associations, prompt artifacts, and hallucinated mechanisms. The paper argues for a different role: agents should inspect data, retrieve context, and explain and clarify method assumptions, acting as transparent assistants to principled causal inference rather than as oracles that emit causal directions.

PaddlePaddle released PP-OCRv6 on Hugging Face, a family of OCR models spanning 1.5 million to 34.5 million parameters and covering 50 languages. The lineup targets practical document recognition at small footprints, offering a lightweight open alternative to LLM-decoder OCR systems for multilingual text extraction on modest hardware.