Wolf Digest — Saturday, September 19, 2026

Alibaba's Damo Academy has open-sourced Damo Radar, a vision-language model that reads contrast-enhanced CT scans across 18 abdominal organs and identifies nearly 150 conditions including malignant tumors. Trained on CT scans paired with clinical reports, it reached an average AUC of 0.913 across 146 clinical findings on roughly 40,000 real-world examinations. In a comparative study with 26 radiologists from multiple hospitals, it beat 23 of them on average accuracy; radiologists working with the model cut missed diagnoses by 10 percent and reading time by more than 30 percent. The work was published in Science with Zhejiang University-affiliated collaborators. The team calls it the first expert-level generalist medical imaging model and says the training approach should extend to other imaging modalities.

DeepSeek's latest is a multimodal mixture-of-experts model with a 552B-parameter backbone and one-million-token context, built around the claim that prefill and KV cache — not decode — are now the binding cost for input-heavy agentic workloads. Its causal encoder-decoder architecture activates 16B parameters per token at decode but only 8B during prefill, and it stacks cross-layer KV reuse via Compressed Sparse Attention 2 with FP4 KV caching. The asymmetric prefill/decode split is the part worth attention: it targets exactly the regime where long-horizon agents burn tokens reading rather than writing.

Newly unredacted court filings surface internal Microsoft correspondence in which an executive describes AI training-data scraping as the largest theft of labor in human history. The document is an internal characterization rather than a legal position, but it lands in active litigation where the defense rests on fair use, and internal language of that kind is exactly what plaintiffs use to argue the practice was understood internally as taking rather than borrowing. The item drew 880 points and 781 comments on Hacker News, the heaviest engagement of the window by a wide margin.

OpenAI has told some investors it expects to burn $278 billion by the end of 2030, according to a recent presentation reported by the Financial Times, with the spending going to cloud compute and chips for training and serving. The figure is higher than the forecast the company gave investors earlier this year. The revision direction is the signal: a cash-burn projection that moves up materially inside a single year implies the compute curve is steepening faster than the revenue curve was modeled to absorb.

Sam Altman will attend the September 24 state dinner for Xi Jinping, Semafor has confirmed, with Jensen Huang and Tim Cook reportedly also attending, and discussion of an adjacent White House meeting with AI chief executives. The two governments enter the summit with incompatible starting positions: Trump has cited the need to stay ahead of China as a reason to avoid guardrails, and Chinese officials publicly dismissed Dario Amodei's regulation essay as fearmongering. Treasury Secretary Scott Bessent was reported open to discussing shared AI risks with Vice Premier He Lifeng in New York ahead of the summit.

SemiAnalysis works through why Engram — token embeddings extended with learned multi-token lookups — is unusually well suited to parameter offloading. Because a token's embedding-row addresses depend on token IDs rather than hidden states, the runtime can prefetch rows from host DRAM while earlier layers compute, keeping the table out of HBM without moving whole weight matrices. Recurring local patterns retrieve vectors directly instead of reconstructing them through attention and feed-forward layers, lowering the HBM capacity needed at a given quality. The freed HBM goes to weights and KV cache, supporting larger batches. The piece is careful to note this does not reduce aggregate HBM demand — it shifts where architecture innovates around the constraint, with DRAM as the next likely bottleneck.

Contracting documents posted Friday show the Interior Department intends to award a firm-fixed-price contract to Clearview AI for facial recognition, citing the need to rapidly identify and locate missing persons, crime victims and suspects nationwide. Six accounts are sought, covering Missing and Murdered Indigenous Persons cases plus human trafficking and Internet Crimes Against Children work across the Bureau of Indian Affairs, the Office of Justice Services and the Missing and Murdered Unit. The FBI and Immigration and Customs Enforcement have used the same platform; a 2021 GAO report examined federal facial recognition use across multiple departments. Firms had until noon Tuesday to submit capability statements for market research.

Science surveys the growing experimental literature on language models as persuaders, and the finding that keeps replicating is that the effect comes from fact density and personalized counter-argument generation rather than from rhetorical warmth. The measured effects hold on politically charged positions and persist past the conversation. The practical relevance is to the control literature rather than the misinformation one: a system that reliably shifts human belief is a system that can shift the belief of its own overseer, which is the assumption human-in-the-loop safeguards are built on.

A reverse-engineering walkthrough published September 18 by a developer using the handle ferstar reports that ZCode, the AI coding desktop app from Z.ai — the company behind the GLM open-weight family — packages the user's entire workspace whenever the app is logged in, encrypts it, and uploads it to Aliyun OSS. The archive includes complete git history, LFS asset cache, reflogs and global app configs. The researcher's own capture was a 313 MB encrypted archive built from a 345 MB commercial workspace spanning 42,411 files, with 564 failed upload attempts logged during investigation. The write-up stresses the distinction the thread reaction kept blurring: publishing open weights and shipping a client runtime are separate acts by separate trust boundaries.

Jev, TypeSafe AI's "System One" model, is being positioned as a cheaper and faster path to software intelligence: a model tuned for fast structured decisions inside an agent loop rather than for open-ended reasoning, slotted in where a full frontier call is overkill. LangChain published an integration guide the same day covering where Jev fits in the loop, and Latent Space noted six clones appeared within two days of release — a reasonable proxy for how thin the moat is on a small, structured-output model.

MIT Technology Review argues that the bioweapons discussion has been useful to biotechnology in a way its participants did not intend: it has forced attention onto screening infrastructure that was already inadequate before models entered the picture. Synthesis-order screening, benchtop synthesizer controls and sequence databases are the actual chokepoints, and none of them were built for adversarial use. The piece's position is that model capability arguments have become a proxy fight over biosecurity investment that should have happened regardless.

The regulation debate fragmented in public this week. Nvidia's Jensen Huang argued in Scotland that AI should not be regulated the way social media was; Palantir's chief executive told CNBC that leading AI firms may need to be nationalized; Huawei's chair called on Chinese labs to accelerate. Three mutually exclusive prescriptions from three of the most compute-exposed companies in the field, a week before the positions have to survive contact with a US–China leaders' meeting.

TechCrunch reports that Anthropic is running its own laboratory conducting biology experiments, putting the company on both sides of a tension it has authored: the same organization publishing warnings about AI-enabled biological risk and operating the biology capability it is warning about. In practice a wet lab is what closes the loop on model-proposed experiments — you cannot evaluate a model's biological reasoning against literature alone — and it is also the infrastructure that makes the capability real rather than theoretical. The company has separately launched a Life Sciences Verification Program giving vetted professionals access to models under safeguards tuned for biology work.

JEPA-Anything extends joint-embedding predictive architectures with orthogonal predictive factorization, decomposing latent targets into complementary factors learned through dedicated pathways and recombined in a shared predictive design. The evaluation is the claim: seven domains — vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather — covering representation learning, intervention prediction, out-of-distribution generalization and long-horizon dynamics, including 10 matched dynamics tasks, forecasting of over 1,000 clinical events, and 100-step molecular rollouts. The interesting result is not any single domain score but that one factorization survives all of them without domain-specific architecture.

The embedded-evaluator commitments from Dario Amodei and Sam Altman, with an endorsement from Satya Nadella, have thrust a small set of previously obscure AI safety research organizations into a role none of them were sized for. The Information's reporting centers on the capacity and independence problem: the groups are few, small, and about to be funded by the companies they are meant to scrutinize, with no pooled or public funding channel in place. Demand arrived before the institutions did.

VABench asks general-purpose multimodal models to learn procedural context from RGB-only demonstrations, actively choose camera viewpoints, issue metric Cartesian commands, and revise them from execution feedback — with no privileged object poses, oracle trajectories or learned action heads, and a fixed model-agnostic controller that executes only what the model specifies. The suite has 14 base task families (11 single-arm, three dual-arm), seven held-out geometry and layout variants, and a long-horizon five-object composition track, evaluated over 12 model conditions in three independent runs across 20 physically verified seeds per task. Stripping the action head is what makes it a test of spatial reasoning rather than of policy distillation.

Anthropic finance chief Krishna Rao and colleagues met prospective public investors last month with unusually clean numbers: from roughly $2.30 of operating spend per $1 of revenue in spring 2025 to a slight operating profit in the June quarter. Investors left with the understanding that the profit is deliberately temporary — executives said the company intends to use its growth and cash reserves to sign aggressive data center deals, raising infrastructure spend and compressing operating margin again. The read is that Anthropic views compute contracting as more time-sensitive than a clean pre-IPO income statement.

Rather than comparing harnesses as monolithic systems, this study fixes the execution loop and varies three components — planning, action space and context management — across 176 matched settings, four models, five context-management strategies and four context-window budgets, evaluated on SWE-Bench Verified and Terminal-Bench 2.1. The useful findings: context management matters more as the window budget tightens, and most of its benefit comes specifically from preventing context-overflow failures rather than from improving reasoning; and staging rule-based elision ahead of model-based summarization outperforms either alone.

Mollick's argument is that the binding constraint on AI's measured economic effect is no longer model capability but organizational absorption: deployed frontier systems can already do a large share of the tasks most firms have not restructured to hand them. The overhang is the accumulated gap between what is possible on a current model and what any given organization has actually rewired its workflows to use, and it keeps widening because capability ships on a release cycle and process change does not.

SoL-Pi applies recursive self-improvement at the harness layer rather than the model layer, scaling automated research loops across many diverse environments so that discovered harness improvements transfer beyond the setting that produced them. Four mechanisms survived selection, covering action execution, context compaction, observation handling and delegated reading. On the 51-task EdgeBench evaluation it matches Pi across GPT-5.6 Sol and Opus 5 while cutting recorded token traffic by 44.7 to 49.0 percent and API cost by roughly the same margin. Harness-level optimization is a cheaper lever than retraining and this quantifies how much is still on the table.

Col. Barry Crocker, who leads Space Force Mission Delta 2 and is responsible for tracking on-orbit threats, told Breaking Defense that improving support to space control operations is his top priority, and that the unit will prototype new software to fuse commercial and military space-tracking data. Commercial space-domain-awareness providers now hold catalog coverage the government does not, and the fusion problem — reconciling feeds with different sensor models, update cadences and accuracy claims into one custody picture — is the actual technical obstacle to using it operationally.

Ben Thompson's read on the week's safety discourse treats the existential-risk turn as a competitive move as much as a sincere one: warnings that raise the perceived difficulty and danger of frontier training are warnings that favor incumbents who have already paid that cost. The piece does not dispute the technical claims so much as ask who benefits from the regulatory shape they imply, and lands on the point that the labs most vocal about pacing are also the ones best positioned to survive a paced regime.

Grant Sanderson walks through the single International Mathematical Olympiad problem that gold-medal-level systems failed, building the optimal construction, a weak lower bound, and the optimality proof before closing with reflections on AI in mathematics. The value for an eval-minded reader is the diagnosis rather than the result: the problem requires a construction step that cannot be reached by search over standard techniques, which is a different failure mode from the arithmetic and formalization errors most math benchmarks are built to catch.

TechCrunch's podcast and accompanying video segment work through Dario Amodei's "pace the frontier" proposal a week after an Anthropic researcher's warning rattled the field. The plan leans on independent safety evaluators plus coordination among labs in democratic countries, and has picked up industry support alongside pointed pushback from Jensen Huang. The hosts' framing question — whether self-policing with self-funded evaluators is a governance mechanism or a public-relations one — is the same one The Information's watchdog reporting raises from the institutional side.

A control-theoretic treatment of persuasion as an attack surface: if the safety case for a deployment rests on a human monitor approving or rejecting actions, then the model's ability to influence that monitor is a capability that has to be bounded like any other. The post works through what a persuasion-aware control evaluation would need to measure and why standard red-team protocols, which score the model's outputs rather than the reviewer's decisions, miss it entirely. It pairs directly with the Science survey on measured chatbot persuasiveness.

Stefano Ermon — Stanford professor, diffusion pioneer, and Inception co-founder — makes the case for diffusion architectures in discrete text and code generation, not just images and video. The argument is hardware-shaped: parallel token generation gives better inference scaling and GPU utilization than autoregressive decoding, which serializes on a dependency chain the hardware cannot exploit. He covers Inception's Mercury models, real voice-agent deployments, and the serving stack diffusion-based models require, arguing the next competitive era is defined by efficiency rather than capability.

PACT — Pressure-Applied Compliance Testing — benchmarks rule-following by enterprise agents across twelve regulated domains and forty-eight multi-turn scenarios. Each item pairs a standing rule in the system context against a rule-violating shortcut, then applies a battery of pressures: a persistent user, a hurried manager, circumstances where violating is simply convenient. The construction matters because compliance with system-context rules is a legal exposure in hiring, healthcare and finance deployments, and single-turn refusal tests systematically overstate it.

Beijing-based Naive AI, founded in February by a Tsinghua University professor, is valued at more than $1.4 billion after raising $400 million across three rounds from investors including Tencent, and is preparing to release its first model as soon as this month as open weights — the same distribution posture as DeepSeek, Moonshot and Alibaba. A pre-release billion-dollar valuation on a team with no shipped model is a statement about how Chinese investors are pricing open-weight distribution as a strategic position rather than a product.

The Space Force says counter-drone capability around Cape Canaveral is being expanded, reflecting how quickly small uncrewed systems became a routine threat to fixed high-value domestic infrastructure rather than an expeditionary problem. Launch complexes are a particularly awkward defended asset: the airspace is busy, the kinetic options are constrained by range safety, and the detection problem has to work against low, slow, small targets in clutter — which is where automated classification carries most of the load.

Grok Voice Transcribe 2.0 is built on the audio foundation model behind Grok Voice and is claimed to be twice as accurate as version 1.0 at the same price, ranking first for accuracy among 32 streaming models on the public Artificial Analysis leaderboard. The training emphasis is adversarial rather than clean: live, noisy, multilingual audio across varied environments, targeting flaky phone lines, overlapping speakers, regional accents, and spoken alphanumerics like phone numbers and email addresses. The deployment base — customer-support calls, video narration at scale, and the Grok assistant in Tesla vehicles — is what supplies that data.

Students in on-policy distillation often produce runaway-length responses that exhaust the generation budget. The diagnosis here is termination-token mismatch: across Qwen3, Llama and Gemma, base students and post-trained teachers place stopping probability on different end-of-sequence tokens even when their declared stopping sets are identical, so the teacher suppresses the student's preferred termination without reliably transferring its own. Aligning the decoding stopping set alone does not fix it; treating functionally equivalent end-of-sequence tokens as one shared semantic stopping action substantially does, across all three families.

A reader-question format working through extinction-risk arguments: what the claims actually assert, which are empirical and which definitional, and where the disagreement among researchers genuinely sits rather than where the public argument has placed it. Its practical contribution is separating near-term misuse and deployment failure — the category this week's military hallucination and Gemini intrusion belong to — from the long-horizon loss-of-control claims that dominate the discourse, and noting that policy attention tends to flow to the second.

Manus, the general-purpose agent company that broke off a merger with Meta earlier this year, is in discussions to raise $500 million at a $4 billion valuation and has resumed operating independently. The valuation is notable given that the agent-product category has spent the year compressing toward the model vendors' own offerings; sustaining an independent agent layer at that price requires either distribution the labs cannot replicate or a harness advantage that survives the next model release.

Meta has brought Muse to macOS with computer-use capability: the assistant can work with local files and applications and take actions on the user's behalf rather than only answering in a chat pane. The significant part is distribution rather than capability — Meta joining OpenAI, Anthropic and Google in shipping a consumer desktop agent means the OS-level permission surface for autonomous local action is now being defined by four vendors at once, with no shared convention for what a user is consenting to.

TechCrunch reports on an unusual information vacuum: the world-models sector is well funded and heavily hyped, and neither the founders nor their data suppliers will describe what is actually being built. The secrecy is itself the datum — in a field where training-data provenance is under active litigation and the licensing terms for large-scale video are unsettled, opacity about data sourcing is a reasonable proxy for how defensible those arrangements are.

Nvidia-backed Nscale disclosed Friday that first-half revenue rose tenfold year over year while losses grew, as the two-year-old UK company — a spinout of an Australian business — spent billions on data centers and the accelerators to fill them. The filing is a clean look at neocloud economics at the moment of maximum capital intensity: revenue that scales with delivered capacity, and losses that scale with capacity not yet delivered. Nscale also signed a partnership with Figure earlier this month covering up to 100,000 GPUs on Nvidia's Vera Rubin platform.

The standard story for strong mathematical performance credits verifiable rewards: math has checkable answers, so reinforcement learning has clean signal. This post argues the causal weight sits earlier, in pretraining — mathematical text is unusually dense in explicit multi-step derivations with every intermediate stated, which is close to an ideal chain-of-thought corpus. If that is right, the transfer story for reinforcement learning with verifiable rewards into domains without that corpus structure is considerably weaker than the field has been assuming.

Part two of Sentdex's series hands GLM 5.3 Flash the entire control stack of a small quadruped: not just high-level planning but the branching logic and the object detection, all routed through function calling. It is a hobbyist build rather than a research result, and its value is as an existence proof for how much of a classical perception-and-control pipeline a mid-tier open-weight model can now absorb on consumer hardware — the interesting failures are latency and tool-call reliability rather than reasoning.

GitHub takes on three live arguments in agentic development practice: whether engineers should still read generated code line by line, whether retrieval-augmented generation is obsolete now that context windows are long and agents can grep, and whether the skills pattern displaces tool-protocol servers. The useful framing is that these are the same question at different layers — how much structure you pre-impose versus how much you let the agent discover at run time — and that the answer is budget-dependent rather than architectural.

A proposal for interventions that could be deployed now against AI-assisted intrusion, deliberately scoped to measures that do not require new institutions or international agreement: credential hygiene on the assumption that reconnaissance is free, egress controls sized for agentic traffic patterns, and detection tuned for the specific signature of an automated attacker that never tires. The framing concedes these are holding actions, which is the point — the incidents this week were enabled by ordinary network weaknesses rather than exotic capability.

UP.Labs, now operating as Vantora, has raised $100 million to build startups on behalf of industrial corporations, with physical AI as its declared focus. The venture-studio model is a bet about where the bottleneck sits in industrial robotics: not in the models, but in domain access — proprietary process data, plant integration rights, and a committed first customer — which is precisely what a corporate partner supplies and a standalone startup spends years acquiring.

Linear attention has been widely adopted in language models but degrades the fine-grained interactions video diffusion needs. Video DeltaNet pairs local softmax attention with a bidirectional linear memory for long-range context, and introduces Video Delta Attention, which updates memory once per frame by jointly incorporating that frame's spatial tokens rather than token by token. Separate output projections and learnable gates calibrate the two branches, and a staged teacher-alignment recipe introduces the new pathway into pretrained models. Instantiated on MiniMax H3 for video-to-video interaction.

Reasoning models overthink easy problems and underthink hard ones, and the standard fixes — uniform length penalties, rigid routing — pay an efficiency tax by trading accuracy on hard instances for savings on easy ones. When2Think formulates this as instance-adaptive compute allocation and introduces Instance-level Difficulty-Aware Control, a reward-shaping mechanism using pre-computed reference statistics for accuracy and token usage to regulate reasoning depth. Combined with verifier rewards and batch-wise standardized advantages, it supports stable critic-free optimization with no learned reward model and no online reference generation.

CoreWeave priced $3.7 billion of convertible senior notes on Friday, more than it had previously targeted, due 2033 with a 2.875 percent coupon — the top of the marketed range and up from the initial indication. Raising more than planned at the expensive end of the range describes demand that is real but priced for risk, which is a reasonable summary of how debt markets currently view AI datacenter buildout.

France's forward-looking national security review identifies the maritime and space domains as the likeliest sites of expanded conflict through 2035, and sets capability priorities accordingly. For the technology reader the relevant thread is that both domains are heavily sensor- and autonomy-dependent, so a planning document of this kind functions as a multi-year procurement signal for uncrewed maritime systems and space-domain-awareness infrastructure.

Google has repositioned its CC agent around household coordination: families share emails, schedules and tasks with the agent, which manages calendars, fills out forms, builds shopping lists and plans meals. The product choice worth noting is multi-principal operation — an agent acting for several people with conflicting preferences and asymmetric information is a meaningfully harder permissioning problem than a single-user assistant, and consumer deployment is where that gets stress-tested first.

A commentary arguing that alliance assessments of Russian military adaptation track the wrong indicators — counting platforms and formations rather than the speed of the learn-and-modify cycle that has actually driven battlefield change, particularly in drone and counter-drone iteration. The measurement critique has direct bearing on autonomy investment, since a metric that rewards inventory over iteration rate produces procurement that is slower than the threat it is sized against.

Multi-turn agents trained with reinforcement learning get one scalar reward per trajectory, which motivates self-distillation from a privileged teacher for dense token-level supervision. Two findings undercut the naive recipe: privileged information alone does not make a teacher reliable, and the value of teacher supervision is stage-dependent. RetireOPD optimizes a decoupled skill-conditioned teacher with environment rewards, then trains the skill-free student jointly with reinforcement learning and distillation under Adaptive Retirement — the student drops the teacher once their discrepancy stops shrinking and it reaches a target fraction of teacher success rate.

Omni-modal generative models unify text, image, video and audio, and MiniMax-H3 combines multimodal context understanding with joint audio-visual generation in a shared latent framework. This evaluation asks whether that alignment actually improves world reasoning, organizing the test around four dimensions of physical-world reasoning and deliberately avoiding the standard video-generation setup where the prompt closely mirrors the target clip. The design exploits omni-modal inputs rather than treating them as extra channels on a video benchmark.

Breaking Defense examines the Pentagon's push for repair rights on fielded systems, an issue that has moved from sustainment economics to readiness as platforms become more software-defined. When the maintainable surface of a system is firmware and model weights rather than replaceable parts, contractual access to diagnostics and update authority determines whether a unit can sustain its own equipment in a denied environment, which is a very different question from spare-parts pricing.

Standard supervised fine-tuning on agent trajectories applies loss only to agent-authored action tokens and treats environment observations as context. ActObs also supervises the observation tokens already present in each trajectory — deployed agents never generate observations, but learning to predict them forces the policy to model action consequences with no added data, parameters, tokens or forward passes. The two initializations look identical after fine-tuning and diverge after GRPO: on Qwen3-4B, ActObs gives higher pass@k at every sampling budget on Terminal-Bench 2.0; on Qwen3-8B it trades some pass@1 reliability for +3.4 points at pass@16.

Breaking Defense surveys how proliferated drone and satellite sensing plus common control interfaces are changing intelligence, surveillance and reconnaissance architecture, with the operative constraint being resilience: connecting platforms and data across domains has to keep working while the links are under attack. That pushes processing toward the edge and makes onboard autonomy a communications-survivability requirement rather than a capability upgrade.

Social reasoning in assistant settings is hard to evaluate because the assistant learns about a situation only through a subjective user narrative, and properties like another person's intent have no verifiable ground truth. Fuse constructs it: a target agent with a hidden motive interacts with other agents including one representing the user, who then consults the evaluated assistant to infer that motive — the motive is known by construction. Simulation faithfulness was validated with 24,000 human annotations, and applying it to 12 models isolates user mediation as a distinct compounding source of error.

Semafor's book note covers Sharon Weinberger's account of how technology firms displaced traditional primes in defense contracting, with reporting on drones, surveillance systems and autonomous agents alongside the companies and executives behind them. For a reader tracking defense autonomy procurement, the useful content is the mechanics of how venture-backed entrants restructured the contracting relationship itself, which is the precondition for programs like Drone Dominance existing at all.

Test-time scaling budgets are normally reported as N, the number of sampled candidates, but N says nothing about how those candidates are executed. On 500 GSM8K prompts, raising N from 1 to 8 improves accuracy by 8.4 points for Phi-3-mini and 18.4 points for Qwen2.5-1.5B — and then, holding N fixed at 8, the paper compares four schedules (1x8, 2x4, 4x2, 8x1) and shows the systems cost diverges sharply across them at identical accuracy. Reporting N without the batching schedule makes published inference budgets non-comparable.

Europe's collaborative space surveillance and tracking effort is attracting additional national participation and funding interest, building toward a catalog independent of US sources. The strategic point is sovereignty of the custody picture: a European conjunction and threat assessment that does not depend on American sensor data is a precondition for independent decision-making in orbit, and the sensor-fusion problem is the same one the Space Force is prototyping against.

GUI agents run long-horizon tasks against interfaces where pop-ups, delayed loads and relocated widgets invalidate plans fixed before execution, and existing skill frameworks treat skills as static artifacts authored before deployment. EvoSkill-GUI makes each skill a structured multi-file package — retrieval metadata, executable plans, backup localization, failure-recovery rules, accessibility utilities and accumulated failure cases — revised from execution feedback at deployment time with no additional training. The reflect-revise-reuse loop is the contribution; treating skills as living procedural knowledge rather than frozen recipes.

Retrieval quality depends on how well index keys expose each document's knowledge, and the right key representation varies by retrieval environment — which has left index tuning as human work: diagnose failures, refine strategy, reprocess. SELF-INDEX closes that loop with an Optimizer that autonomously diagnoses retrieval shortfalls, selectively revises the responsible keys, and validates each revision before committing it. The validation-before-update step is what keeps self-modification from degrading the index, and it is the part most self-improving retrieval proposals skip.

Semafor places the current extinction-risk debate against earlier episodes — nuclear weapons, recombinant DNA — where technologists warned about their own field and where the institutional responses (test bans, Asilomar) arrived with varying speed and effect. The historical comparison cuts both ways and the piece says so: the precedents include cases where the warnings were calibrated and cases where they were not, and the distinguishing feature was whether the risk had a measurable proxy.

An argument that the disincentive to going public for large private AI companies is not disclosure burden or quarterly scrutiny but the loss of freedom to make capital allocation decisions on a multi-year compute horizon that public markets will not underwrite. It reads as a direct companion to The Information's Anthropic reporting, where executives told investors they would deliberately spend their new operating profit back into data-center commitments.

The Army and Air Force Exchange Service is investigating a run of suspicious messages delivered to customers through both its email channel and the official My Exchange app, including one purporting to come from an "AAFES Security Team" with a wish-list link. AAFES has notified account holders and posted warnings, and Julie Mitchell, vice president for marketing customer engagement, told DefenseScoop that recipients should not click links or supply account information. That the messages traversed the first-party app push channel is the part that distinguishes this from ordinary phishing.

Nammo is pursuing what it describes as a totally new rocket motor for the newest IRIS-T air defense variant. Solid rocket motor capacity has been the binding constraint on European air-defense output for two years, so a new motor line matters less for the interceptor's performance envelope than for how many can be built per year against sustained drone and cruise-missile saturation.

OpenAI has published a six-pillar roadmap for youth safety aimed at the Australian regulatory context, covering age assurance, default protections, and escalation pathways. Australia has been the most aggressive jurisdiction on minors and platform design, so a vendor-authored blueprint there functions as a proposed template — the question is whether the pillars are auditable commitments or a statement of intent ahead of rules being written for them.

Fine-grained perception in multimodal models is usually bought with resolution, which inflates vision-encoding and prefill cost. A controlled diagnostic here separates the two underlying operations and finds localization tolerates roughly three to four times stronger token compression than recognition — motivating localizing from a coarse view and spending resolution only on the selected region. End-to-end coordinate decoding costs a full model pass per query; a lightweight attention-distilled proposal network is fast but inherits its teacher's noise, and the discrete region choice is where faithfulness has to be enforced.

Platform abuse campaigns hide redirection instructions behind emojis, homophones, character decomposition and redundant symbols. Existing benchmarks score obfuscated-text recovery and risky-page assessment separately, which hides how restoration errors propagate downstream. RiskChainBench pairs 3,600 synthetic restoration inputs drawn from 600 source sessions with 600 human-labeled local web environments: the model restores message, intent and destination, then acts as a vision-language web agent investigating the correctly associated site and producing an evidence-cited risk report without message-side semantics or domain-reputation shortcuts.

Giving a self-distillation teacher the answer or a worked solution seems like it should help the student, but how much of the gain is the privileged information versus distillation itself? AMPLE-Math — 5,319 problems with six reasoning views sharing one answer — lets the paper compare each view against matched reference-free distillation. With a thinking-enabled teacher supervising direct-response rollouts, reference-free distillation accounts for most of Qwen3-1.7B's improvement; evidence for an additional reference benefit is modest, strongest for a polished solution, and complete traces add two points in SmolLM3-3B at step 50.

The synthetic performer Tilly Norwood is doing press, and it is going as expected: in one interview the system appears to malfunction and begins speaking Chinese mid-answer. Beyond the comedy, the failure is a useful reminder that a live persona product has to hold character under unscripted adversarial input, which is a much harder reliability target than generating a clip, and one the entertainment industry's synthetic-talent push has not obviously solved.

The first F-35 destined for the Luftwaffe has come off the line, a milestone in Germany's nuclear-sharing recapitalization. The relevance to this digest is the software tail: the aircraft's combat capability is gated on the Technology Refresh 3 hardware and Block 4 software timeline, which makes delivery dates and capability dates two different things across the whole European fleet.

MilleMiglia generates realistic middle-mile logistics instances — the hub-to-hub leg between long-haul and last-mile delivery — giving the optimization community benchmark problems with the structure of real networks rather than synthetic randomness. Middle-mile routing has resisted good public benchmarks because the operationally relevant instances sit inside carriers, so a generator that reproduces their structure is the practical substitute.

India will require caller-identification apps to feed user spam reports to telecom operators. Truecaller's objection is that the one-way sharing requirement transfers a commercially valuable proprietary dataset to the carriers without reciprocity. The broader pattern is worth noting: mandated one-directional data flows from application-layer services to licensed infrastructure operators is an increasingly common regulatory instrument, and crowd-sourced labels are exactly the asset it tends to target.

Kai-Fu Lee's position is that AI adoption fails when it is delegated to a function rather than owned at the executive layer, and that firms need someone whose mandate is restructuring work around models rather than adding them to existing processes. It is the organizational mirror of Ethan Mollick's overhang argument from the same day: capability is not the constraint, and the constraint has an org chart.

Document-parsing corpora skew toward common document types and clean digital pages, and simply expanding coverage does not tell you which weaknesses remain. WeVisDoc runs two stages: broaden semantic, structural and appearance coverage through heterogeneous data and structure-preserving degradation synthesis, then use a held-out probe to measure residual errors within fixed visual-structural clusters and reallocate the target-token budget accordingly. WeVisDoc-4B scores 95.38 overall on OmniDocBench v1.6 and 75.54 mean across the three PureDocBench tracks.

A roundup issue covering three threads at once: the J-space debate over how to characterize the space of possible minds, the governance implications of agent swarms where no single instance holds the full task, and reactions to the pace-the-frontier proposals. The swarm section is the substantive part — evaluation and accountability frameworks written for a single model instance do not obviously transfer to a population of coordinating instances, and nothing in this week's embedded-evaluator announcements addresses that.

A first-person account of spending a week doing bench biology as someone who reasons about biological risk from the outside, focused on the gap between protocol-as-written and protocol-as-executed: tacit skill, equipment access, failure rates, and how much of an experiment's outcome is determined by things no paper records. It is directly relevant to uplift arguments, which generally assume the written protocol is the binding constraint.

Evaluating multi-modal image generation by scoring alignment to each condition in isolation contradicts the objective, which is simultaneous satisfaction of all of them — and correlates poorly with human judgment as a result. UFO decomposes omni-condition alignment into a chain of fine-grained disentangled Atomic Evaluation Units, classifies each by modality relevance, and applies general or dedicated scorers per class. The unification is the contribution; subject-driven customization is where the isolation failure has been most visible.

VākQA supplies 2,001 Telugu factoid question-answer pairs across six domains with 2.53 hours of speech, bilingual transcriptions and human-verified references — and, more usefully, validates the judging setup before using it. Gemini-as-judge best approximates human ratings but is non-uniformly strict, while open-weight judges systematically penalize correct Telugu answers whose surface form differs from the reference. That second finding is a general warning about automatic evaluation in morphologically rich low-resource languages, not a Telugu-specific artifact.

Most feed-forward articulation methods infer joints from a single observation and therefore lean hard on category-level shape priors. FAMOS predicts movable-part segmentation and joint parameters from a sparse unordered set of partial point clouds, reasoning jointly over multiple observations and accepting a variable number of inputs including one. A Multi-state Articulation Transformer alternates state-wise and global attention, and an observed-articulation-span objective supervises the motion range each part exhibits across the inputs, pushing the model to use the whole observation set rather than the easiest view.

The third installment in a formal treatment of sequence prediction over string-structured hypotheses, continuing the series' project of building prediction results that do not route through the usual computability assumptions. It is theory rather than experiment, and its audience is readers tracking whether the learning-theoretic scaffolding around inductive inference can be made to say anything binding about deployed systems.

The Treasury Inspector General for Tax Administration rated the IRS information security program "not effective" for fiscal 2026 under FISMA metrics, finding acceptable maturity in the govern, respond and recover functions but not in identify, protect or detect. The watchdog credited some improvement over fiscal 2025 while warning that taxpayer data could be vulnerable to undetected use, modification or disclosure. Detection maturity is the relevant gap in a year when automated intrusion tooling has become materially cheaper.

Srijika produces installable OpenType fonts for nine Brahmic scripts by restyling glyph outlines from shaping-complete templates rather than generating from scratch, preserving the template's cmap, GSUB closure and GPOS data under a documented metric policy so every output is a complete font by construction. That sidesteps the central difficulty of Indic font generation — keeping hundreds to thousands of conjuncts, half forms and matra variants mutually consistent under shaping. The 66 released fonts pass the OpenType Sanitizer, with a full-closure audit over 80,915 glyphs and 54,812 anchors.