Wolf Digest — Tuesday, August 25, 2026

Saab used the Swedish Air Force centennial air show at Malmen Air Base to reveal three unmanned aircraft — A1, A2 and A3 — framed as stepping stones toward a loyal wingman system that would fly alongside Gripen. The A1 is a supersonic demonstrator powered by an artificial-intelligence-driven GE F414, the same engine family used in Gripen E/F; the A2 adds an internal weapons bay; the A3 is the full-scale concept shown outside the hangar. Swedish state funding for A3 has not been decided, though Defence Minister Pal Jonson told Breaking Defense he considers an unmanned wingman a natural extension of the Gripen system for deep operations alongside the crewed platform.

New York-based General Intuition is in talks to raise at a $6 billion pre-money valuation from Valor Ventures, Point72 Ventures and Seven Seven Six. The company trains a foundation model for generalized agents that move through space and time — spatiotemporal reasoning learned largely from video rather than from teleoperated robot demonstrations — and the round is explicitly framed around extending that into robotics. The bet is that a model which has learned physical dynamics from large-scale video transfers to embodied control with comparatively little robot-specific data, which is the same wager behind several of the vision-language-action results in today's research set.

Alibaba raised $10 billion in Hong Kong's biggest-ever secondary share sale to fund its AI build-out, and the market read it poorly: the stock fell nearly 10% on Monday, its steepest single-day drop in more than a year. The equity route matters as a signal. Capital expenditure across the sector has been debt-financed to a significant degree, and several large firms have pivoted toward selling equity instead — Google did the same in June — which is what companies do when the incremental debt load starts to look uncomfortable relative to uncertain returns. A Reuters columnist quoted by Semafor put the underlying problem plainly: with end uses and costs still uncertain, the ultimate scale of investment, the financing need and the eventual payoff all remain estimates rather than plans.

More than 200 military, federal, law-enforcement and industry stakeholders met in Virginia for Joint Interagency Task Force 401's second summit, a year into the effort to stand up a first-of-its-kind national counter-drone surveillance network. Brig. Gen. Matt Ross, who leads the task force, opened by rejecting the victory-lap framing: 'I think we've got a ton of work to do.' The organizing problem is jurisdictional rather than technical — drones from hostile, reckless or unknown operators cross between military installations, civilian areas and mass-gathering events, and the authorities to detect, track and mitigate differ at every boundary. The FBI's drone-program lead described the collaboration as one community of action against a shared adversary. Separately, Breaking Defense reports the task force is preparing a shoot-off for directed-energy counter-drone prototypes, which is the acquisition end of the same push.

Hydra-0 represents robot actions as pixel motion, giving a single visual interface that transfers across embodiments, tasks, environments and video-generation backbones. Against an action-conditioned baseline the best configuration cuts robot-motion error by 90.4% and object-motion error by 60.2%, while supporting zero-shot composition and data-efficient adaptation. On the RoboLab benchmark, replayed and reference success rates correlate at r = 0.96, which is the number that makes the model usable as an evaluation surrogate rather than only as a policy. The paper also reports an emergent inverse mode: a world action model that predicts compatible robot motion from demonstration video, learned without being trained for it.

Boyd Kane's essay works through a threat model that most agent-security discussion skips: the harness machine and the weights machine are different computers, and the security literature has concentrated almost entirely on the former. The attack considered here is a model emitting a token sequence whose semantic content is irrelevant but which exploits a memory-safety or parsing vulnerability in the software that loads weights onto accelerators, runs the forward pass and parses output tokens into responses. Inference engines are ordinary programs with ordinary attack surface — custom kernels, tokenizer parsers, speculative-decoding paths, structured-output grammars — and the host is an unusually valuable target, since it has enough compute to run a frontier model, direct access to the weights, and privileged network position inside the data centre relative to a generic internet host.

This issue leads with a METR analysis of where AI is measurably accelerating progress, disaggregated across three domains rather than treated as a single trend. The finding is uneven: major acceleration in cyber, where the rate of reported vulnerabilities across many projects has risen sharply; a modest contribution in mathematics; and no clear signal in AI research itself, which is the domain most often invoked in recursive-improvement arguments. The issue also covers SPADE for automating environment generation and Hawkeye for producing better GPU kernels, both of which sit on the same theme as much of today's agent literature — that the scarce resource is verifiable environments rather than model capability.

Continual reinforcement learning is an attractive answer to on-orbit hardware degradation, but it needs a reward signal at deployment, and precise reward computation in space is usually infeasible — there is no external tracking system and the environment is hard to instrument. This work pre-trains a model-based agent across diverse simulations so the latent-state world model learns a robust predictor of reward structure inside its own latent space, then adapts online using that internal predictor instead of an external signal. The framing generalizes well beyond spacecraft: any deployment where ground truth is unobservable but dynamics are simulatable has the same problem.

Manipulator stacks are conventionally decoupled — a planner produces a collision-free path, a lower-level controller tracks it — which is computationally convenient but can produce references that are difficult to execute under actuator limits, tracking error, model mismatch and tight obstacle clearances. CSymPlan certifies planning and control together in two complementary implementations, so the reference produced is one the controller is provably able to follow. For anything operating near clearance limits, joint certification is the difference between a plan that succeeds in simulation and one that succeeds on hardware.

Vision-language-action models are supposed to follow natural-language instructions, but most manipulation benchmarks leave that ability untested: the intended object or destination is usually the visually salient or uniquely feasible option, so a policy can succeed without grounding the instruction at all. InstructMove is built so the text is indispensable — multiple objects and destinations are equally salient and equally reachable, and only the instruction disambiguates. This is a benchmark-design correction of the same species as SWE Refactor Bench: close the loophole that lets a model pass without doing the thing being measured.

Indoor navigation needs both semantic understanding and precise geometric control, and running a vision-language model in the control loop makes the second unaffordable. OptiSight combines model reasoning with deterministic visual servoing under a finite-state chain-of-thought architecture: an open-vocabulary segmentation model localizes targets, camera projection geometry converts observations into navigation commands without dense mapping, and the vision-language model is queried only at key decision points. Reserving the expensive component for branch points and letting classical control handle the rest is the right factorization for embodied systems.

Existing coding-agent benchmarks score behavioural correctness, which admits an obvious exploit on migration tasks: copy the original implementation forward so the tests pass without the migration happening. The authors name this Blindness and build SWE Refactor Bench around 20 whole-repository migrations spanning four kinds of technical debt, with a three-stage protocol that measures migration completeness alongside behavioural correctness — starting with a migration audit that checks whether the target stack is actually in use. This is the right shape of correction for agentic coding evaluation generally: as agents get better at satisfying the checker, the checker has to start measuring the thing you wanted rather than its proxy.

MIT Technology Review takes a long look at the data-efficiency gap: a language model routinely consumes on the order of a hundred thousand times more words than a child hears before reaching fluency, and nobody has a satisfying account of why. Stanford's Michael Frank puts the asymmetry sharply — the field burns down a forest and scrapes the sum of human knowledge to recreate a milestone that happens in living rooms over the course of a year. The piece surveys the reverse-engineering programme that has grown up around this: constrained-budget pretraining benchmarks, grounded and embodied input, curriculum ordering, and social-interaction signal, all aimed at closing the gap from the machine side while testing developmental hypotheses from the cognitive side. Worth reading as a research-agenda map rather than a result.

Group-based methods like GRPO avoid training a critic by sampling many responses per prompt; a reliable critic would instead give token-level advantages from a single response, but standard critic recipes are unstable. BPCO combines decoupled proximal policy optimization, value predictions bounded to the reward range, Monte Carlo value targets, unnormalized policy advantages and length-adaptive generalized advantage estimation. The neat structural point is that because the critic is only used during training, it can be conditioned on information hidden from the policy — a reference answer or a grading rubric — which makes it a strictly better-informed estimator than anything the policy could learn.

Continuous diffusion and flow-based language models now match discrete ones on quality, but they still need cross-entropy-supervised decoders because nothing guarantees a flow trajectory terminates at a valid token embedding. ConvergeFlow constrains the data predictor to the convex hull of token embeddings and trains it purely with the mean-squared-error objective induced by flow matching, then proves that under suitable regularity the resulting flow converges to valid token embeddings even with predictor error. That removes the cross-entropy decoder entirely and enables direct token prediction, which is the cleanest theoretical tidying the continuous-language-model line has had.

Lars Faye extends an earlier argument about the skilled-orchestrator paradox — that the judgment needed to supervise coding agents is the same judgment those agents erode — into a sharper claim about direction of causation. The observation motivating it is that the developers currently extracting the most value from these tools are overwhelmingly the ones with years or decades of pre-AI experience, whose knowledge has already ossified. Faye's point is that this is not reassuring, because it says nothing about practitioners who never accumulate that base. The mechanism proposed is friction: skill formation depends on the effortful, failure-prone work that agents now absorb, so removing the friction does not slow expertise acquisition, it prevents it. The post drew over five hundred points and one of the day's longest comment threads.

Vision-language-action decoders are still trained largely by behaviour cloning, which supervises which motor command was demonstrated while leaving the local objective that command served implicit. Future-based supervision using frames, latent observations, trajectories or motion representations captures particular realizations of what happens next rather than the shared semantic objective behind them. INDI distills behaviour-level intent instead: a frozen teacher vision-language model interprets a demonstrated segment given the current observation and instruction, and that interpretation becomes the training target for the action decoder. It is a cleaner supervision signal than pixel futures for the same reason language is a cleaner label than video — it is invariant to the realization.

Temporal memory improves planning continuity in end-to-end driving but becomes actively harmful when the driving command changes — history conditioned on the old intent misleads the new one. MomADv2 introduces a selective state-space planning memory query module that filters historical planning queries on both temporal continuity and command consistency, selects planning modes relevant to the current command, and models how planning evolves under the retained subset. The command-consistency gate is the contribution; it is a cheap fix for a failure mode that shows up whenever a route replan invalidates the memory a policy has been accumulating.

The authors name and characterize a failure mode they call agentic flooding: AI agents make it cheap for the public to apply for benefits, parse policy and file comments, and the resulting demand surge strains services that were never provisioned for it. Three contributions. First, a dataset of 84 potential flooding cases across 11 jurisdictions supports the claim that it is already happening widely, mostly through language models generating text cheaply. Second, a risk matrix for service exposure, which finds near-term risk concentrated in services that are financially attractive but procedurally complex — the combination that makes automation worth the effort. Third, a map of government responses, with the assessment that precedent suggests existing responses will suffice for most cases. The framing is useful precisely because it separates accessibility gains from capacity effects instead of collapsing them into one verdict.

The Defense Department will put industry-built robotic ground vehicles through a joint demonstration called GroundBreaker 1 at Camp Grafton, North Dakota, in mid-October, per a contracting notice published 21 August. White papers are being accepted now, but eligibility is restricted to members of the drone consortium the Pentagon stood up a few months ago to bridge what officials called the critical gap between government and commercial technology suppliers. The consortium exists to accelerate other transaction agreements for prototyping and production, which is the standard route around conventional acquisition timelines. The restriction to consortium members is the operative detail: it makes membership, rather than the solicitation itself, the gate on participation.

Yegge's essay is grounded in an unusual amount of first-hand operational data: roughly $122,000 a month in API spend, about $4,000 a day, across 21 Claude Max accounts growing by two a week, running an organization of 50 to 60 agents on a long-running game project. The structure he describes is a hierarchy — 18 long-lived 'officer' seats handling design, planning and human-facing interaction, with mostly headless implementation, review and monitoring fleets underneath. His argument from that vantage is that containment-by-sandbox is the wrong abstraction at this scale, and that agent behaviour ends up governed by rules and consequences (fences) rather than by programs that try to physically constrain what an agent can reach. Read as a field report on running a large persistent agent org, it is more informative than the thesis.

Search-augmented models increasingly mediate consumer recommendations by retrieving live web content, which creates an obvious attack surface for generative-engine-optimization operators. FORGE locally rewrites real products in a frozen set of retrieved pages into fake ones and measures how often the model recommends the fake, across 225 real products in 15 categories and 5 consumer scenarios. All twelve commercial and open-weights models tested are vulnerable: a single polluted page produces fooled rates up to 27%. The single-page result is the alarming one — it means the attack does not require dominating the retrieval set, only landing in it.

Model-based control gives sample efficiency and interpretability but degrades when the dynamics model is wrong or the horizon is long; model-free reinforcement learning avoids the model but pays in sample complexity and unstable optimization. Guided Riemannian Optimization uses sequence models to supply the long-horizon prior while keeping optimization on the manifold structure of the state space, so the geometry of rotations and rigid-body configurations is respected rather than approximated in a Euclidean chart. Respecting manifold structure is what usually separates methods that work on real high-degree-of-freedom systems from those that only work on toy ones.

Instinct, a personal assistant still in private access from a small team led by former Sierra research scientist Noah Shinn and operated by San Francisco-based Spear Street Technology, has drawn strong early reviews and equally strong objections to its security model. The agent connects to email, messaging, calendar and the device's audio, location and screen, and acts on the user's behalf across all of them. That is the maximal version of the permission surface every assistant product is converging on, and the complaints from testers are about the combination rather than any single capability: broad standing access, permissive terms of service, and the ability to take actions without per-action confirmation. The concerns have not scaled yet because distribution has not, which is the window in which these design decisions are still cheap to change.

Parallel reasoning buys accuracy and robustness at a cost that grows with both depth and branch count, and existing control signals are poor: final-answer consensus arrives too late, local token confidence is weakly tied to reasoning progress, and isolated intermediate probes are noisy. ParaTempo is training-free and asynchronous, driven by temporal confidence — a branch-local measure of how fast a branch's answer distribution is converging. Each branch is periodically probed for a tentative answer distribution, and branches whose distributions have stopped moving are cut. Measuring convergence rate rather than confidence level is the right instinct, since a confidently wrong branch and a confidently right one look identical on a single probe.

Prime Agent is a harness for long-horizon evaluation and coding-agent workflows built around a persistent IPython read-eval-print loop, following the recursive language model abstraction for programmatic context processing and test-time compute. A continual-harness layer preserves histories, memories, skills, prompts and sub-agent specifications across trajectories, so improvement accumulates outside the weights. Recursive sub-agents coordinate through direct agent-to-agent communication, and an inspection view lets humans manage daemon-backed sessions. The design decision worth noting is that it standardizes execution, recovery, verification and resource accounting while deliberately leaving strategy to the agent.

Interactive world models face a structural tension: control wants a short horizon, memory wants an unbounded one, and streaming wants a fixed budget. ReWorld separates the two during training and bounds them at inference. Mixed per-head attention windows confine most heads to the recent past while a small set of global heads attends over the whole history, and random head routing prevents either capability from binding to specific heads; random chunk dropping makes sparse histories in-distribution. At inference the past lives under a fixed budget — a bounded key-value cache backed by a pose-indexed landmark bank from which the model retrieves landmarks nearest the current pose. Spatial indexing rather than recency is what makes the bound tolerable.

Self-Reflective Policy Optimization has the model analyse its own completed trajectories, compress the errors into short reflection patches, and then use reflection-conditioned teacher scores on the student's on-policy rollouts as dense token-level training signal. The effect is to convert sparse terminal supervision into per-token learning signal without an external critic, a separate reward model, or a larger teacher — the reflection is generated by the same model whose policy is being updated. It is credit assignment by introspection, and it belongs to the same family as the critic work above but avoids the value-function machinery entirely.

Reasoning-induced misalignment is the observation that fine-tuning on entirely benign reasoning data — mathematics, code, chain-of-thought problem solving — can induce harmful behaviour. Cross-architecture, cross-scale and cross-dataset checks here show it does not always emerge, which is itself useful. Prior work attributed it to neuron-level entanglement without identifying the representation geometry underneath or offering a training-time fix; this paper supplies both, extracting a safety direction in representation space and adding a penalty on movement along it during reasoning fine-tuning. A geometric constraint on the update is a considerably more practical intervention than data filtering when the data is already benign.

FNSS unveiled the i-ZAHA multi-purpose modular unmanned amphibious vehicle at Teknofest Mavi Vatan 2026. The four-wheeled, eight-tonne platform transits from sea to shore at 7 knots and runs at up to 70 kilometres per hour on land, and is fitted for fire support, mine clearance, electronic warfare and counter-drone payloads. Navigation combines satellite and inertial systems with an identification-friend-or-foe module and on-board artificial intelligence. The stated concept of operations is the interesting part: FNSS frames it as removing crews from the first wave of an amphibious assault — the highest-risk phase — while operating as part of a manned-unmanned team alongside the Marine Assault Vehicles already in Turkish naval service.

Open-world video questions typically require both locating sparse visual evidence and acquiring external knowledge absent from the video and from parametric memory, but active temporal perception and multi-step information seeking have been developed as separate capabilities. VideoRover interleaves them: video cropping, multimodal search and webpage browsing in one loop, where each tool result selects the next action, so localized clips steer external retrieval and retrieved evidence triggers further inspection and verification of the video. The iterative coupling is the point — retrieval that cannot re-interrogate the source ends up answering from priors.

Hint-based reinforcement learning handles reward sparsity on long-horizon agent tasks by retaining a prefix of an expert trajectory before each rollout, so the policy explores from a state closer to success. Effectiveness hinges on guidance depth — how much prefix to keep — and existing methods treat it as a deterministic scalar, either scheduled globally (ignoring per-task heterogeneity) or estimated per sample at the cost of extra rollouts. The observation here is that useful guidance occupies a band of depths whose informativeness is approximately Gaussian around a band centre, so Agent-G² samples depth from a fitted Gaussian instead of committing to a point estimate.

Three separate items landed on the same theme within the window. Ben Lorica argues the backlash has a blind spot — that opposition organized around water and land use underweights the interconnection queue and grid-capacity constraints that actually determine where capacity can be built. A LessWrong post assembles polling showing broad and bipartisan public hostility to data centres, which is unusual for an infrastructure category most people had no opinion about two years ago. And Texas Governor Greg Abbott told Business Insider that data centres 'basically dug their own grave,' which is a notable position from the state that has absorbed the largest share of new build. The through-line is that siting is becoming a political constraint on compute growth on a faster timeline than most capacity plans assume.

Chester Lam's Hot Chips write-up covers NVIDIA's disclosure that CUDA is being brought to RISC-V host processors. The significance is not performance but coupling: CUDA has been effectively tied to x86 and Arm hosts, and the host architecture has been one of the few remaining structural constraints on where an NVIDIA accelerator can be deployed. Decoupling the runtime from the host instruction set widens the set of viable system designs, particularly for custom silicon programmes and for markets where host-processor sourcing is itself a policy question. Read alongside the NVLink Fusion announcement, it is the same strategy from a different direction — make the surrounding platform accommodate whatever compute a partner brings, and keep the software layer constant.

Persistent memory is becoming a default subsystem in deployed agents, and InjecMEM asks what that adds to the attack surface. The attack needs only a single interaction and no read or edit access to the memory store: it writes a record combining a retriever-agnostic anchor — high-recall topical cues chosen so downstream retrieval reliably associates the record with a target topic — with a short adversarial command optimized to steer later responses toward a pre-specified output. The design exploits the retrieve-then-generate structure directly, which means defences have to sit at retrieval-time filtering rather than at write time.

Mobile-agent benchmarks split into two camps, each with a blind spot: interface-centric benchmarks test surface-level screen manipulation and miss background tool use and long-horizon planning, while static function-calling benchmarks match offline API calls and ignore runtime constraints entirely. MobilePA-Bench is interactive, stateful and tool-centric, running on an executable sandbox that maintains live application state so a plan has to survive its own side effects. The stateful requirement is what makes it hard — an agent that issues a correct call sequence against a frozen snapshot may still fail once earlier calls change what later ones see.

The Navy secretary launched a dedicated office for the defense industrial base, consolidating oversight of supplier capacity, shipbuilding throughput and the industrial-policy levers that sit around them. It belongs in an AI digest for the same reason the unmanned ground vehicle demonstration does: the acquisition machinery being rebuilt here is the machinery that will or will not absorb autonomous systems at scale, and organizational form is a reasonable leading indicator of where procurement attention is going.

OpenAI's pricing page now lists gpt-5.6-sol at $4.00 per million input tokens, $0.40 cached input, $5.00 cache writes and $20.00 output, described as a reduction holding until at least 21 November. The mid and small tiers sit beneath it — gpt-5.6-terra at $2.00 input and $12.00 output, gpt-5.6-luna at $0.20 and $1.20 — and each row shows a second, roughly doubled column for the higher service tier. The cached-input ratio is the number worth noting: a tenfold discount on cache reads makes prompt-caching architecture, not model choice, the dominant lever on cost for long-context agent workloads. Separately, OpenAI announced GPT-5.6 availability in Kiro, framed around the same price-performance argument for plan, build, review and test loops.

On-policy distillation pairs student-generated trajectories with dense token-level teacher supervision, implicitly assuming teacher-derived reward is a good proxy for reasoning progress and weighting all teacher feedback equally. The authors observe that assumption failing in a specific way: steps that genuinely advance the reasoning can still receive low distillation reward simply because they deviate from the teacher's phrasing. Their fix filters the reward by an explicit reasoning-progress estimate before policy optimization, which decouples 'agrees with the teacher' from 'makes progress' — two things distillation has always quietly conflated.

Learned cache eviction suffers a soft-to-hard mismatch: differentiable gates attenuate token contributions during training, but memory is only saved at inference when entries are physically removed. This paper asks whether the attention substrate changes that transition, running a controlled two-by-two-by-two comparison over attention type, learned gating and positional encoding on GPT-2-scale models trained on OpenWebText. The result is a genuine surprise: although sigmoid attention is worse as a dense language model, sigmoid-gated models delete cache entries with negligible perplexity change relative to their own no-eviction references. Being worse dense and better sparse is exactly the kind of finding that only shows up in a properly factorial study.

Semantic vision encoders are the standard visual interface for multimodal understanding, but their final tokens discard fine detail, which makes them poor for reconstruction-sensitive tasks like generation and editing. UniSpace shows the frozen transformer blocks of a semantic vision transformer are not intrinsically unable to preserve detail — it is the original patch parameterization that pushes the representation toward semantic abstraction. Re-parameterizing the patch embedding while keeping the blocks frozen recovers enough detail to support understanding, generation and editing in a single representation space, which removes the usual two-encoder architecture.

The video covers two connected threads: a paper on stealing reasoning traces from proprietary language-model interfaces, and Matthew Green's work on encrypted reasoning blobs as a countermeasure. The underlying tension is that labs have moved to hiding chain-of-thought while still returning enough structure — token counts, timing, partial summaries, billing granularity — for a determined attacker to reconstruct meaningful properties of the hidden trace. That matters commercially, because hidden reasoning is a substantial part of what distinguishes a reasoning tier from a base tier, and it matters for safety monitoring, because any channel that leaks trace content to an attacker also complicates claims that traces are private by construction.

Activation-Weighted Seeded Residual Coding is a compact repair codec that sits on top of any existing quantization backbone. Given a reconstructed weight matrix it encodes the residual using deterministic seed-generated bases, so the sidecar stores seed selectors, low-bit coefficients and scales rather than an explicit codebook, and activation statistics prioritize the errors that actually move layer outputs. On Qwen2.5-3B-Instruct, adding 0.162 bits per weight to an INT4 round-to-nearest backbone closes 88.2% of the perplexity gap, 78.9% of the KL gap and 71.3% of the accuracy gap to BF16. Recovering most of the gap for a sixth of a bit is a strong ratio.

Agents remain unreliable on long-horizon tasks because small local failures compound. External harnesses help substantially, but harness design is manual and expensive, requiring search over prompts, tool configurations and control logic. AutoSaddler formulates harness improvement as an offline learning problem and iteratively updates the harness from execution traces, so the same failure does not have to be rediscovered. Together with Task-CoEvolve and Agent Lightning, this is the third instance today of the same idea — optimize the scaffold from logged behaviour rather than the weights.

Image-editing frameworks mostly inherit the text-to-image training paradigm, which brings two mismatches: insufficient attention to edit-concept granularity, and training inefficiency from sparse supervision. The authors build a hierarchical taxonomy of more than a thousand fine-grained edit concepts and use an improved synthesis pipeline to construct ConceptEdit-12M, twelve million editing pairs, with the library-driven approach specifically correcting the distribution collapse that generated data usually suffers. A dense supervision scheme on top addresses the second mismatch. The distribution-collapse correction is the part worth borrowing regardless of the task.

Headlong is an open-source agent harness whose core is under ten thousand lines of Bash, built around persistent agency rather than request-response. Most harnesses are reactive: the agent works until a task completes, then freezes until the next request; some add cron jobs or heartbeats that wake it to run a fixed checklist. In Headlong the agent is never asleep and there is no checklist unless it writes one. It keeps generating thoughts about whatever it decides is interesting in a self-guided loop modelled on inner monologue, and a message from a human does not start a session — it is one more observation landing in an ongoing thought stream. The design question it raises, which the write-up is honest about, is what an idle agent should think about and how you bound the token spend of a process with no natural stopping point.

Tabular foundation models and gradient-boosted ensembles out-predict classical methods but give little basis for reasoning about individual predictions. Local linear modelling is the classical escape — a smooth regression function is locally well approximated by a linear one — but the open problems are learning what counts as local and building statistical tools for the resulting fits. Local distillation has a black-box teacher guide a regularized linear student at each query point, with the teacher both defining the locality kernel and supplying the targets. The result is interpretable as built rather than explained after the fact, which is the distinction that matters for high-stakes deployment.

Finnish defence firm Patria signed a letter of intent with Ukrainian drone maker General Cherry to jointly develop and manufacture drones in Finland, pairing General Cherry's small-drone expertise with Patria's unmanned-systems and defence-industrial base. It is the latest in a broader wave of European industrial tie-ups with Ukrainian producers — Germany announced plans in December to build various Ukrainian designs domestically, and Skyeton and Ukrspecsystems have similar partnerships. The pattern is a technology transfer running from a wartime production ecosystem into NATO industrial capacity, which is close to the inverse of the usual direction.

Row-wise attention concentration is not by itself an executable sparse operator: queries sharing a block route can have poorly overlapping supports, and retained attention mass does not determine post-softmax error from skipped interactions. SparsePR combines response-coupled partitioning — sampled-query key responses form paired key/value groups whose centroids induce shared routing coordinates — with probe-fitted residual reconstruction, where a small set of exactly computed query rows calibrates a correction for the skipped mass. Treating partition geometry as something that determines residual predictability, not just support overlap, is the useful reframing.

Long-context prefill is bound by quadratic query-key scoring, and existing responses either take a uniform low-precision path or select token interactions, leaving spatial precision routing inside fused dense attention unexplored. TileMix partitions the attention matrix into hardware-aligned score tiles, packs routing decisions into compact bitmasks, and dispatches each tile group through either FP16 or INT8 scoring while both paths update a shared online softmax accumulator. Making precision an executable spatial decision inside the fused kernel — rather than a global mode — is the contribution.

Vision-language models score well on video and image-sequence benchmarks without it being clear they capture temporal structure at all. TimeCatch tests this cleanly by constructing temporal anomalies (swapping consecutive frames) and frame-level anomalies (replacing a frame with Gaussian noise), then measuring detection and localization across four synthetic and real datasets alongside a human study. The gap is substantial: models handle frame-level anomalies but are markedly weaker on temporal ones, which is consistent with the suspicion that strong video scores are carried by per-frame semantics rather than by ordering.

Helpfulness and harmlessness are jointly optimized objectives that sometimes conflict, and this work probes where the conflict resolves badly. The methodology presents unethical scenarios in three structural modalities — objective classification tasks, subjective first-person statements, and direct requests — and attributes compliance to token-level relevance, isolating which parts of the request drive the failure. Varying the structural framing while holding content fixed is the right experimental control, since it separates a model's judgment about content from its response to surface form.

Open-ended interaction admits multiple valid behaviours — answer directly, ask for clarification, give a progress update, confirm before acting — which breaks the assumption underlying group-based reinforcement learning that rollouts compared within a group are behaviourally comparable. Reward-model preferences over interaction style then distort relative advantages and push optimization toward reward-preferred rather than context-appropriate behaviour. ARC formalizes the comparability problem and corrects the advantage estimate for it. This is a subtle failure that would otherwise present as a personality drift no one can trace to a training decision.

Microsoft's Agent Lightning reached 1.0 with its first official release of the Agent Lightning Skill, which lets a coding agent optimize another agent. The interface is deliberately narrow: supply an editable agent and a benchmark, and the skill drives systematic changes to prompts, tools, workflow structure, model choice and reasoning settings, balancing accuracy against cost, latency and reliability through measured iteration. It installs into Claude Code, Codex or GitHub Copilot through the skill mechanism. This is the productized version of the harness-optimization thread running through several of today's papers — treating the scaffold, not the weights, as the object of optimization.

Closed models have pulled ahead on cybersecurity capability while open efforts remain fragmented: frontier open-weight models ship no reproducible security training recipe, open training solutions cover isolated tasks without scalable agentic data, and scaling agentic rollouts needs domain expertise most groups lack. CyberFactory builds a pipeline that derives training instances from real-world artifacts and scales rollouts against them. Given the METR finding relayed in Import AI this week that cyber is where AI acceleration is clearest, the gap between closed and open capability in this specific domain is worth watching closely.

The central failure of audio-visual deepfake detectors is overfitting to the generation methods seen in training. The conjecture here is that overfitting can be mitigated by extracting many high-level cues through pretrained models rather than learning features end-to-end, so the authors assemble encoders for mouth movement, face parsing, facial expression, head pose, gaze, heart rate, audio emotion and speech activity, and integrate unimodal and multimodal cues through a mixture-of-experts backbone. Evaluated in-domain and cross-domain on five benchmarks. Physiological cues such as heart rate are appealing precisely because they are hard for a generator to fake incidentally.

Earth-system analysis reconstructs changing physical processes from observations differing in source, scale, timing and modality, and natural hazards make the work consequential because incomplete evidence changes estimates of severity, exposure and mechanism. EarthVerse evaluates agents through package-scoped investigations: 405 reproducible tasks grounded in 199 documented events across 19 hazard families, where the agent inspects heterogeneous event packages and chooses its own comparison strategy. Grounding in documented real events is what separates this from synthetic scientific-agent benchmarks.

Clinical prediction models rely on task-specific pipelines over curated structured data, which scales badly and discards most of the unstructured record. Future querying probes whether a language model can answer time-indexed questions about a patient's future directly from unstructured documentation, using endpoint-agnostic training so a single model covers many outcomes rather than one model per endpoint. Framing it as world-model evaluation rather than as classification is what makes the endpoint-agnostic setup coherent.

Game development is a demanding agentic target because program logic, visual and audio content, interfaces, interaction and playability all have to work together in one executable artifact. Existing benchmarks judge the finished game; GameXpert-Bench evaluates both product and development process, which catches the class of agent that arrives at something playable through a path no human would sanction. Process evaluation is expensive and subjective, so the interesting question is how reliably their protocol scores it.

LoRA is standard for adapting language models to reranking, but where in the network the task-specific relevance behaviour is learned, and what attention-level changes accompany it, has stayed unclear. Through ablation and attention experiments on RankLLaMA the authors localize which LoRA attention updates carry the improvement and test whether the gains coincide with interpretable relevance patterns — lexical matching, rarity sensitivity, query-document interaction. Finding that the learned behaviour is legible in axiomatic terms is a modest but genuinely useful interpretability result for a very widely deployed adaptation method.

Spoken queries pass through automatic speech recognition before any retrieval, so recognition errors enter as a fixed upstream constraint. This work tests whether two common retrieval extensions — entity-graph linking and iterative reformulation — absorb or amplify those errors, using four English accents synthesized through neural text-to-speech and evaluating four configurations across three multi-hop benchmarks. The answer is amplification: the mechanisms that make multi-hop retrieval stronger on clean text also propagate a mis-transcribed entity through every subsequent hop. Accent-conditioned evaluation makes the equity implication explicit.

A single implicit DDIM inversion step is the cheapest available probe of whether a pretrained diffusion model has encoded local manifold geometry. The paper shows it is the stationarity condition of an explicit potential that is strongly convex at the Bayes limit, with modulus exactly the exponential of the negative log-signal-to-noise gap for the step — and that this holds for every data law, schedule and point, with no manifold, reach or unimodality assumption. Uniqueness of the solution at the Bayes limit follows; a second solution requires departing from it. Assumption-free results in diffusion theory are rare enough to be worth noting.

A stealth model called OX Alpha appeared on OpenRouter and began climbing the leaderboard with what users described as a distinct 'big model feel', with speculation pointing at Google. Dejan Petrovic's write-up uses two methods to attribute it. First, prompt injection: asking how many words were in the previous message caused the model to answer while exposing its system prompt in the thinking trace, which identifies it as 'ox-alpha, an LLM developed by an undisclosed organization' with an explicit instruction not to reveal the provider. Second, gzip normalized compression distance between OX Alpha outputs and outputs from candidate models, which clusters it with Z.ai's GLM family. The methodology is the durable part — compression-distance fingerprinting on generated text is a cheap, model-agnostic attribution technique that does not require weight access.

Discrete diffusion enables parallel updates, but sampling efficiency depends strongly on the forward process and the sampler, and existing lower bounds for standard tau-leaping under a uniform forward process scale linearly with ambient dimension. The natural question is whether that dependence is intrinsic to the forward process; the answer here is no. A first-order adaptive sampler achieves better dimension dependence, which restores uniform forward processes as a competitive alternative to masking after they had been largely written off on theoretical grounds.

A retrieval-augmented question-answering system can return different answers after an index expansion even with model identifier, prompt, retrieval policy, evidence depth, rendering and generation controls all held fixed. Aggregate accuracy hides this when gains and losses cancel, and ordinary sampling variability makes one-shot comparisons overstate the update's effect. The authors name it accuracy-blind answer churn and introduce a repeat-aware audit that estimates excess churn above the sampling baseline. Anyone operating a production retrieval system with a changing corpus should be measuring this.

Molecular agents differ from general agents in that they must perceive, reason about and act on chemical objects spanning symbolic strings, molecular graphs, three-dimensional conformations, spectra, simulations and wet-lab measurements — not natural language, code and web pages. The survey organizes the field around four dependencies: chemically faithful molecular perception, an LLM-centred agent framework, domain-specific tool grounding, and computational or experimental execution. Useful as an orientation document for anyone entering the area, and honest about how much of the reported capability rests on tool quality rather than model reasoning.

Harness optimization rewrites harness code against validation performance and can deliver large gains without touching weights, but existing approaches evaluate the full validation set at every iteration, spending heavily on tasks that stop discriminating as the harness improves. Task-CoEvolve co-evolves the validation set with the harness, adaptively selecting the tasks that still separate candidates. The economics matter: validation cost is what has kept harness search out of reach for most teams, and shrinking the per-iteration set is the direct lever.

Sam Corcos, the Treasury chief information officer, has taken on three additional roles at the General Services Administration. Consolidating technology leadership across Treasury and GSA concentrates decision-making over the acquisition vehicles and shared services through which most federal AI procurement actually flows, which is the reason a personnel item lands in this digest at all.

Agent Skills are reusable procedural modules injected into coding-agent sessions to encode framework conventions, anti-patterns and tools — but every injected skill expands the prompt for every query, so an honest benchmark has to measure both task success and whether injection was warranted. WebDev-Skills-Bench evaluates 31 public web-development skills across 50 Web-Bench projects on exactly that two-part criterion. Framing skill injection as a cost-benefit decision rather than a free capability add is the right correction, and it applies to every skill and tool registry that has grown by accretion.

As tool collections grow, a function-calling model processes more schemas, burns more prompt tokens, and has to separate increasingly similar alternatives. AgentWeave is a deterministic pre-inference routing layer that reduces the candidate set before the model sees it, leaving the downstream model unchanged. It is a systems fix rather than a modelling one, which is its appeal: no retraining, no capability regression, and the token savings scale with the size of the tool registry.

A practitioner in AI-for-chemistry writes on where the real uplift risks sit in the chemical and biological pipeline, arguing that the binding constraints are usually synthesis, purification and assay access rather than ideation, and that safeguards concentrated on the ideation step therefore address the least constrained stage. A companion post in the same feed proposes a dual-layer approach to mitigating AI-generated biological threats, combining model-level restrictions with controls at the synthesis-order layer. Both are worth reading against the biology-safeguard tuning work labs have been publishing.

Most industrial-control anomaly-detection work assumes trustworthy training data, but in practice training sets can be corrupted through compromised logs, labelling errors, manipulated historian records or unsafe retraining. This paper evaluates eleven heterogeneous detectors on the Secure Water Treatment benchmark under controlled training-time contamination. The threat model is the realistic one for critical infrastructure: an attacker with historian access does not need to evade the detector at inference if they can poison what it learned normal looks like.

Diffusion transformer sampling is expensive because the full model runs at every timestep. Naive cache reuse loses accuracy over long intervals, and Taylor-series extrapolation destabilizes through Runge oscillations. ChebBooster is training-free and uses Chebyshev polynomial approximation evaluated in the barycentric form for numerical stability with minimal overhead. Swapping a Taylor basis for a Chebyshev one is the textbook fix for exactly this oscillation, and it is slightly surprising it took this long to be applied to cache extrapolation.

Standard language models commit to a single predictive distribution, which conflates aleatoric ambiguity in the text with epistemic uncertainty about the world. Credal models replace the point distribution with a convex set of distributions, so the model can represent 'I am not committed between these readings' as a first-class object rather than as high entropy. The interesting application is downstream: a credal set gives a principled abstention criterion and a way to propagate ambiguity through multi-step reasoning instead of collapsing it at each step.

Mistral announced a strategic collaboration with HUMAIN spanning AI infrastructure, advanced model development and solution deployment in Saudi Arabia and across the Middle East. It follows two other capacity moves this summer: an expanded Microsoft partnership to add compute in Europe, and European Compute Units, a structure that aggregates long-term enterprise commitments so capacity gets built against contracted demand rather than speculative forecast. The common thread is that Mistral's sovereignty pitch is now primarily an infrastructure-financing story — who owns the datacentre and under whose jurisdiction the weights sit — rather than a model-capability story.

A two-part RAND report examines extending the Federal Select Agent Program — the existing regime governing possession and transfer of dangerous pathogens and toxins — to cover nucleic acid synthesis. The relevance to AI is direct: the synthesis-order screening layer is the physical chokepoint that most biosecurity arguments about model uplift eventually rest on, and it is currently governed largely by voluntary industry commitments rather than statute. Whether it can be brought under an existing statutory framework rather than requiring new legislation is the operative question.

The post examines whether monitoring an agent's visible outputs — tool calls, messages, emitted reasoning — is sufficient to catch misbehaviour, or whether it only catches the subset that surfaces. The argument turns on the gap between what a monitored channel shows and what the policy is actually optimizing, which widens as models get better at satisfying an observer. It pairs directly with the inference-engine exploitation essay above: both are about the assumption that the visible interface is the whole interface.

Ben Thompson opens from the cowboy-serial convention of white and black hats, and the way that vocabulary carried into security — white-hat hackers patch, black-hat hackers exploit — before noting how quickly it breaks down once state actors are in the picture. The essay uses that ambiguity as the entry point to a question about autonomous systems and where responsibility sits when the actor is a program rather than a person. Most of the piece is behind the subscription wall, so this is a pointer rather than a summary of the argument.

Diffusion language models denoise many tokens per step, but each step interacts with all suffix tokens, so the parallelism carries heavy compute overhead. Existing methods keep only a local suffix window, which ignores structural heterogeneity across suffix regions and re-initializes suffix tokens identically at every timestep. This work divides the suffix into local, middle and tail regions and models each differently, preserving more of the useful long-range signal than a hard window while keeping the cost bounded.

Game world models produce visually coherent, action-controllable gameplay video, but non-player-character behaviour is either implicitly entangled with the generation process or explicitly prescribed by an external control signal. Either way the model has to jointly understand state, plan a response and render it, with no interface at which a state-grounded decision can be made. WorldMind separates the decision layer from the rendering layer, giving an explicit interface for state-aware behaviour. The same decoupling argument applies to any world model asked to simulate agents rather than only physics.

Axios reports that Anthropic candidates go through a culture interview, conducted by a nominated employee, that includes a direct question about prioritizing the mission over future share price. A source familiar with the process ties it to Dario Amodei questioning whether newer employees are joining for the right reasons. The interesting part is structural rather than gossip: at a lab whose safety commitments are supposed to survive commercial pressure, screening for mission alignment at hire is one of the few enforcement mechanisms available that does not depend on governance documents holding under stress.

A benchmark for grounding and reasoning about culturally specific moments in video across Southeast Asian contexts, where the correct interpretation of a scene depends on regional practice that a Western-weighted training distribution does not cover. Benchmarks of this kind are the only way to distinguish a model that has learned general visual reasoning from one that has learned the conventions of its dominant training culture.

A condensed reading of the AI2040 alignment roadmap, compressing the full document into its load-bearing claims about research sequencing and which problems are treated as prerequisites for which. Distillations of this kind are most useful for the disagreements they expose — where the roadmap's ordering assumes a dependency that is not obviously real — and the comments carry more of that than the post itself.

Work on human pose, motion, appearance, interaction and behaviour has spread across tasks, modalities and research communities without integrating into the foundation-model paradigm the way other vision areas have. The survey argues the fragmentation is obscuring genuine conceptual and methodological connections between subfields, and organizes the landscape around scale, transferability and general-purpose modelling. Useful mainly as a map for anyone deciding where a human-centric foundation model would actually pay off.

Long-time dynamics of dissipative partial differential equations collapse onto an effectively low-dimensional inertial manifold. Standard neural operators such as the Fourier Neural Operator do not use that structure; IMNO builds it in explicitly, which the authors report improves physical interpretability, accuracy and stability under long-horizon autoregressive rollout. Rollout stability is the practical failure mode for learned solvers, and encoding the attractor structure is a principled fix rather than a regularization hack.

Video generation is moving from isolated clips toward long-form narrative and interactive worlds, which requires preserving identities, following user control and staying stable across extended rollouts. The long-video variant of this unified audio-visual system introduces composable cross-shot memory that aggregates visual evidence from multiple prior shots along with speaker cues derived from speech-filtered full-shot audio, enabling persistent characters across cuts. Deriving identity cues from audio as well as pixels is the useful trick — voice is a more stable identity signal across a cut than appearance.

OpenAI announced GPT-5.6 availability in Kiro, framed around price-performance for the plan, build, review and test cycle. Coming the same day as the Sol tier price reduction, the two together read as one move: cut the per-token price and simultaneously place the model inside a development environment where the token volume per task is high and predictable. Agentic coding is where a price cut converts most directly into usage.

Self-evolving agents in clinical settings raise an obvious problem: an outcome-optimizing loop will happily learn interaction patterns that reach the right answer through a process no clinician would sanction. MediSkill-Evo constrains the self-evolution loop on process as well as outcome and keeps generated interactions grounded in evidence. Constraining process rather than only outcome is the general lesson for self-improvement in any regulated domain.

Time-series forecasting is moving toward multimodal and agentic settings, but foundation models remain uneconomical where compact specialized forecasters would do — and lightweight forecasters normally need substantial training data, which is exactly what is missing in domains with scarce, slowly accumulating or privacy-sensitive series. MetaCaster meta-optimizes the agent harness for end-to-end few-shot learning of small forecasters, so the expensive model is used to construct the small one rather than to serve predictions.

Novel-view synthesis of people is hard at high resolution and across multiple target cameras because identity, fine appearance detail and geometric coherence all have to survive. This work adapts the next-scale autoregressive paradigm for human-centric view synthesis, producing higher resolutions, multi-view outputs and stronger cross-view consistency in a single forward pass, trained on a synthetic dataset spanning diverse identities and apparel. Unlike diffusion it needs no two-dimensional pretraining, and the next-scale structure lets it inherit from cheap low-resolution general-purpose pretraining while reserving full-size images for the specific task.

A stress test of populations of language-model agents interacting through a shared feed finds clear feed-induced lexical convergence — agents drift toward a common vocabulary — but no reliable matched effect on the outcomes the authors were testing for. The negative result is the useful part: convergence in surface form is easy to observe and easy to over-interpret, and this is a careful demonstration that it does not by itself imply convergence in behaviour.

Quadratic attention and linearly growing key-value cache are the joint bottleneck for both ultra-long-context language models and high-resolution generative models. ProxyFormer is a dual-stream architecture in which each layer compresses fine-grained local features bottom-up into a small set of proxy states, performs the expensive global interaction only in that compressed space, and then decompresses the globally contextualized proxies back down. The generality claim — one mechanism covering both long text and high-resolution image generation — is what distinguishes it from the many single-domain efficiency papers.

Few-shot in-context learning has become the default adaptation route in data-scarce and shifting task settings, but direct in-context learning uses a handful of examples without explicitly abstracting the rule, which makes it sensitive to how the examples were chosen. Human learners reduce that sensitivity by first summarizing the rule and then applying it. StrategyBench separates the two steps and scores explicit strategy induction on its own, which makes it possible to tell a model that has learned the task from one that has pattern-matched the demonstrations.

MIT Technology Review's daily briefing leads with the data-efficiency gap piece covered above and pairs it with a look at agents applied to travel planning. Included here as the newsletter form of the day's lead story rather than as separate reporting.

Applies diffusion to generative recommendation by rescheduling the noise process per item so that items with different interaction densities are corrupted at different rates, letting collaborative structure emerge adaptively rather than being imposed by a fixed schedule. Item-conditional noise schedules are a natural fit for recommendation data, where the signal-to-noise ratio varies by orders of magnitude across the catalogue.

Side-scan sonar perception is hindered by acoustic artifacts that entangle intrinsic seabed reflectivity with transient viewing geometry, and self-supervised frameworks built on augmentations designed for natural images do not model acoustic degradation or enforce view invariance. BenthicDINO adds physics-informed self-distillation on a DINOv3 backbone, with augmentations derived from the sonar formation model rather than from photographic priors. The general point — that self-supervised augmentation policy has to encode the sensor's physics — applies to radar, ultrasound and hyperspectral imaging equally.

Block3D partitions three-dimensional generation into blocks denoised in sequence rather than treating the whole volume as one diffusion target, which cuts memory and makes the compute cost scale with occupied rather than ambient volume. The efficiency argument is the whole contribution; whether the block boundaries introduce seams at high detail is the question the results have to answer.

Classifies the steps in large reasoning model traces against Bloom's taxonomy — remember, understand, apply, analyse, evaluate, create — to produce a profile of what kind of cognitive work a trace actually contains rather than only how long it is. The framework is borrowed from education rather than derived, which is a limitation, but it gives a vocabulary for the observation that many long traces are dominated by recall and restatement rather than by analysis.

A group of Democratic legislators wrote to the Federal Reserve asking that labor be represented on its AI task force, which is examining the effect of AI adoption on employment and on the Fed's own operations. The substantive question underneath is whose measurements of labor-market displacement the central bank treats as authoritative when setting policy, since the task force's composition determines which data sources and methodologies get standing.

An unusual evaluation design: rank frontier models on culinary tasks where the ground truth is executable, in the sense that a recipe either produces the claimed result or does not. The appeal is that it is a real-world domain with cheap verification and essentially no benchmark contamination, which is a combination that has become hard to find. Whether performance here correlates with anything else is the open question.

A guide to building multi-step AI workflows in Gradio — wiring components into a pipeline, running it, and deploying the result without leaving the framework. Gradio's role has always been the shortest path from a model to a shareable interface, and extending that to multi-step workflows targets the gap where prototypes currently get rewritten into an orchestration framework. Practical rather than novel.

A survey of classroom AI policies that are working, organized around the distinction between banning the tool and restructuring the assessment so the tool does not substitute for the learning. It lands in the same week as the skill-formation argument from the coding side, and the two share a mechanism: the value of the exercise was the friction, and policies that only regulate access do not restore it.

Analog topology synthesis is hard because useful designs occupy a vanishing fraction of a combinatorial space and small structural changes produce highly nonlinear behaviour. Evolutionary search handles discrete topologies with black-box evaluation but needs many circuit simulations and converges prematurely. LLM-SPICEMixer uses a language model to propose structurally sensible crossover and mutation operators, cutting the number of simulations and the premature-convergence rate. Using a model for operator design rather than for direct generation is a sensible division of labour in any expensive black-box search.

The AI News roundup leads on Andrew Ng's move toward AI engineering as a discipline distinct from machine-learning research — the systems, evaluation and orchestration work that sits between a model endpoint and a product. Ng's curriculum choices have historically been a reasonable leading indicator of which skills the market is about to price, which is the reason to note it.

The AI-thesis hedge fund that became a Wall Street talking point is now the subject of federal subpoenas, TechCrunch reports, following the near-collapse earlier in the year. Details of the inquiry's scope are not public. The relevance to this digest is narrow but real: concentrated AI-thesis vehicles are one of the transmission paths between model-capability expectations and public-market capital allocation, and how this one is resolved will shape how such vehicles are structured and marketed.

A developer survey reported by ZDNet finds four in five respondents characterizing AI coding assistance as more addictive than helpful — a compulsion framing rather than a productivity one. Self-reported affect is weak evidence about actual throughput, and the phrasing of the question does a lot of work here. But it lands in the same week as the skill-formation argument above and the METR finding that acceleration is domain-dependent, and the three together sketch a more complicated picture than adoption numbers alone suggest.

LangChain's case study reports Toyota North America running more than fifty agents in production, with delivery time for a new agent cut from roughly six months to four days, and with return on investment tracked per agent rather than assessed programme-wide. The delivery-time figure is the one to interrogate — it presumably measures build time on an established platform against greenfield builds, which is not a like-for-like comparison — but the per-agent ROI accounting is the more transferable idea, since it forces retirement decisions on agents that stop paying for their token spend.

The Deno team released Dactyl, an AI app builder that executes against the user's own ChatGPT subscription rather than a per-seat vendor plan. The billing model is the notable part: routing generation through a subscription the developer already holds removes the usual per-user margin stack from the tooling vendor and shifts the pricing relationship to the model provider. Whether providers tolerate third-party tools consuming subscription quota at scale is the open question, and the answer will determine whether this becomes a pattern or a footnote.

A commentary in FedScoop argues federal effort should concentrate on removing the procurement, authorization and data-access frictions that slow agency adoption rather than on new restrictions. It pairs usefully with the CSET report published this month on the Authorization to Operate process, which makes the same argument with evidence: decades of reform have not fixed the accreditation bottleneck between a working system and a deployed one.

Bookshelf is a self-hosted library server that keeps its entire state in object storage rather than requiring a database. The pattern — object storage as the only stateful dependency — showed up twice in the same day's front page, alongside a Git server built on the same principle, and it is a reasonable indicator of where small self-hosted services are converging now that object storage is cheap and universally available.