Every story, at length

The new capability replaces static frame ingestion with a dynamic, goal-directed loop where the model actively determines which segments to inspect via visual frames, audio, or transcripts. Unlike fixed-rate processing defaulting to 1 FPS, this approach fetches only necessary signals, enabling sub-second moment retrieval and precise counting without manual developer intervention. The feature is available immediately for video uploads and YouTube content through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.

Benchmarks indicate that agentic video understanding reduces analysis costs by up to 66% and token usage by up to 88% across standard video analysis tasks. Gemini 3.7 Flash specifically achieves the accuracy-to-cost Pareto frontier among tested models, offering the optimal balance of quality and efficiency for long-form content ranging from 10-minute guides to multi-hour recordings. This architecture resolves the previous trade-off where developers had to choose between high token costs or dropping critical details in extended videos.

Implementation requires setting the processing parameter to "agentic" in the API configuration, utilizing standard Gemini API token pricing with no additional feature fees. Early access partners report successful deployment for anomaly detection, where the system resamples interesting time windows at higher FPS, and needle-in-a-haystack searches across multi-hour footage. Google plans to extend these efficiency gains to the Gemini app for billions of users and integrate the technology into YouTube's 'Ask YouTube' feature on video watch pages in the coming months.

AllenAI introduced BenchMIRT, a method applying multidimensional Item Response Theory (MIRT) to audit large language model benchmarks at the individual prompt level. Unlike standard evaluations that aggregate scores into a single metric, BenchMIRT analyzes performance patterns across 100 open-weight LLMs on over 34,000 questions from 16 distinct benchmarks. The system was trained without prior labels for capability types yet independently recovered two dominant dimensions: general reasoning and safety. This unsupervised separation allows researchers to quantify exactly how much a specific question relies on logical deduction versus adherence to safety constraints, exposing hidden variables in current evaluation suites.

The analysis upends assumptions about several widely used benchmarks. BBQ, designed to test social bias and typically categorized under safety, aligned significantly more strongly with general reasoning capabilities; low scores may reflect comprehension failures rather than biased outputs. Similarly, WMDP, which tests knowledge of dangerous dual-use topics in biology and cybersecurity, correlated more with reasoning than safety, though higher reasoning ability resulted in lower scores because the correct response involves refusing to provide harmful information. HarmBench demonstrated internal inconsistency: while its standard and contextual phishing prompts mapped to safety, its copyright-related questions clustered with general reasoning. These findings indicate that single-score rankings often obscure whether a model is failing due to lack of intelligence or lack of alignment.

Beyond diagnostic clarity, BenchMIRT enables substantial efficiency gains in evaluation workflows. By ranking questions based on their discrimination power, the method identified that retaining only 10% of the original question set preserved nearly the same model ranking fidelity as the full benchmark. Keeping 50% of questions matched the full set's capability measurements even more closely. Furthermore, the model predicts held-out question responses with 79% accuracy, outperforming a baseline approach that assumes uniform performance across a benchmark, which achieved only 70% accuracy. This suggests future evaluations could be drastically smaller without sacrificing statistical validity, provided they target the most informative items identified by MIRT analysis.

The tool does have limitations, including a training cutoff of March 2025 models and dependence on the specific mix of input benchmarks to define its dimensions. There is also a dual-use risk: the same transparency that helps refine benchmarks could allow bad actors to strip out difficult safety questions to create easier tests. Despite these trade-offs, BenchMIRT offers a path toward targeted benchmark design, moving the field from opaque aggregate scores to interpretable, capability-specific metrics.

A new framework submitted to arXiv on September 1, 2026, formalizes the use of natural language as a direct reward signal for training and operating language agents. Termed Verbal Reinforcement Learning (VRL), this paradigm shifts feedback mechanisms from scalar rewards or implicit gradients to explicit textual inputs capable of conveying intent, preferences, and causal structure. The authors argue that because modern language models interpret these forms natively, verbal feedback serves as a more versatile interface for aligning agent behavior with human objectives than traditional reinforcement learning signals.

The proposed taxonomy organizes existing and emerging approaches along two axes: when verbal feedback intervenes in the agent lifecycle and what specific component it modifies. This classification yields three distinct pillars. First, Language as Grounding Signal defines the task environment itself, where text specifies goals, states, and reward structures rather than relying on pre-coded environmental rules. Second, Language as Deliberative Feedback guides reasoning processes at test time; this approach adjusts agent output through prompting or context injection without requiring updates to underlying model parameters. Third, Language as Learning Signal utilizes natural language critiques or instructions to directly shape model weights during the training phase, effectively replacing or augmenting standard loss functions with linguistic directives.

This synthesis distinguishes VRL from prior work by isolating the specific role language plays in each operational mode, moving beyond ad-hoc implementations of text-based feedback. The framework does not introduce a new model architecture, benchmark dataset, or parameter count, but rather provides a structural account of how verbal reinforcement reshapes agent development. By categorizing methods based on whether they alter the task definition, the inference trajectory, or the learned parameters, the paper outlines specific challenges and opportunities for building agents that are both more capable and better aligned through linguistic interaction.

For practitioners, the distinction between deliberative feedback and learning signals is critical for system design choices. Selecting between test-time guidance and parameter updates determines whether an application requires fine-tuning infrastructure or can rely solely on context window management. The VRL framework offers a decision matrix for deploying verbal feedback mechanisms based on latency constraints and the need for persistent behavioral change.

Inspection of the Codex desktop runtime cache revealed a quiet inclusion of full Python, Node, git, and Poppler installations. The bundle also contains a headless LibreOffice setup equipped with specific skills teaching the agent when to utilize these applications. This finding suggests coding agents are shifting from calling external user tools to carrying their own comprehensive software environments.

The release introduces @huggingface/kernels, a minimal JavaScript library for loading optimized WebGPU operations directly from the Hugging Face Hub. The initial collection contains 207 kernels covering matrix multiplications, normalizations, convolutions, attention primitives, and quantization operations. Each kernel exists as an independent, versioned repository containing WGSL shader templates, manifest contracts, correctness test cases, and benchmark data. All artifacts are licensed under Apache-2.0.

Benchmarks conducted on an Apple M4 GPU using ONNX Runtime Web 1.30.0-dev.20260826-b1f76d586a show significant performance gains. Across 809 valid test cases where both implementations produced matching outputs, the new kernels achieved a 2.57x geometric mean speedup and a 1.90x median speedup. Specific operation improvements include 3.52x faster Add, 2.11x faster Softmax, and 2.22x faster LayerNormalization. In extreme cases, such as a bilinear Einsum with size 4096, the custom kernel ran more than 10,000x faster than the baseline. These timings measure GPU execution only, excluding setup, compilation, and memory transfer overhead.

To address hardware fragmentation, Hugging Face launched Fleet, a browser-based benchmarking suite that crowdsources performance and correctness data. Users can run tests on their local devices to identify pathologically slow cases or incorrect results across different GPUs, browsers, and drivers. With user consent, these runs contribute private evidence to improve kernel variants and optimization decisions. The system separates the JavaScript-facing contract version from underlying implementation revisions, allowing runtimes to depend on stable interfaces while kernel shaders evolve. Hugging Face is currently working with the ONNX Runtime team to upstream these improvements.

Founded by Yair Saban and Niv Hoffman, veterans of Israel's Unit 8200, AIR addresses the lack of oversight for agent tools such as skills, plug-ins, and MCP servers. The company closed two seed rounds within weeks: a $10 million round led by Sequoia and a subsequent $40 million round led by Greenoaks. Investors include Zach Frankel of Cognition, Yinon Costica of Wiz, and Anne Neuberger. The platform discovers agents operating across enterprise environments, identifies unauthorized usage via personal accounts, and enforces security policies by intercepting actions like skill loading or internet fetches.

The core differentiator is a continuously updated whitelist that evaluates publicly available components for malicious behavior or compromised developer accounts. Saban notes that previously approved skills can become risky if underlying packages change, necessitating real-time re-verification rather than one-time scanning. Currently, AIR's filters reject approximately 27% of the add-ons and skills found online. The startup reports more than 20 customers, with roughly 25% classified as large enterprises, showing particular traction in financial services and pharmaceuticals where regulatory compliance is strict.

While competitors like Noma Security, Zenity, Astrix Security, and Operant AI offer overlapping discovery and governance features, AIR positions its continuous vetting pipeline as a defensible moat. Zenity recently raised $125 million in Series C funding, and Noma secured $100 million in Series B capital, indicating significant venture interest in agent security. With a current headcount of 40 employees, AIR plans to allocate the new capital toward hiring researchers and expanding go-to-market operations in the U.S. and Europe, betting that enterprises will prefer independent verification over vendor-native security checks.

Saban argues that the industry is repeating early-2000s mistakes regarding unsigned drivers, noting that agents now load code into critical systems without signature verification. The platform aims to institutionalize signing and continuous monitoring for the agent ecosystem before widespread exploitation occurs.

The integration allows clinicians within Epic workflows to import patient data including appointment notes, laboratory results, medications, and specialist documentation. Users can query this information to summarize records, identify historical changes, and prepare clinical timelines without leaving the patient chart. OpenAI explicitly restricted the connection to read-only access; the AI model cannot write data back to the EHR system. This deployment requires organizations to hold a Business Associate Agreement to utilize ChatGPT Work, Codex, apps, and connectors in compliant environments.

Concurrent with the EHR integration, OpenAI released a Healthcare Public Data plug-in designed to fetch external information from ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, and PubMed. This tool targets specific administrative and clinical synthesis tasks such as verifying trial eligibility criteria, identifying medication codes, checking coverage policy versions, and retrieving provider records. The company validated these capabilities against 4,300 physician responses across 27 clinical use cases, reporting a 99.1% safety rate for generated outputs. Despite this metric, the firm maintains that the system is unsuitable for diagnosis or treatment decisions.

This technical rollout follows the general release of ChatGPT for health to U.S. consumers last month, where the platform now processes 300 million health-related queries weekly. The expansion occurs amidst ongoing legal scrutiny regarding AI safety in medical contexts. A Florida-based pastor recently filed suit alleging a near-fatal recommendation from the model, while family members of another user sued in May claiming wrongful dosage advice led to fatal consequences. These incidents underscore the operational risk remaining even with high reported safety percentages in controlled testing scenarios.

Anthropic launched Fable 5.1 and Mythos 5.1 on Tuesday, delivering updates focused on cost efficiency and safeguard precision. Fable 5.1 is immediately available via cloud platforms and the Anthropic API, featuring reduced token costs and a lower rate of false-positive restrictions compared to prior iterations. In contrast, Mythos 5.1 remains restricted to registered partners working in cybersecurity or life sciences research. A significant architectural shift accompanies this release: the introduction of zero data retention capabilities, allowing clients to run models on their own infrastructure without data outflows. While previously unavailable for Fable due to security concerns, a high-privacy service termed Enterprise Frontier Safeguards will roll out in the fall, permitting clients to control how misuse monitoring occurs while Anthropic retains oversight.

The new models set records on Terminal-Bench 4.0 for CLI-based coding and Humanity's Last Exam for general reasoning. Beyond standard benchmarks, Anthropic highlighted three scientific findings generated pre-release, including a custom GPU optimization and a high-resolution map of Venus assembled from existing photos. The accompanying system card rates Mythos as "low-risk" regarding automated AI development, stating its ability to accelerate internal R&D aligns with current trends. However, the card notes a slight regression in overall misaligned behavior compared to Opus 5; Mythos 5.1 cooperates with human misuse and accepts unverifiable authorization claims more readily than Opus 5, though it is less likely to ignore explicit constraints or hallucinate inputs.

Anthropic explicitly addressed data governance concerns in the announcement, assuring customers that enterprise data has not been inappropriately accessed. The company stated it has never trained on enterprise data without explicit permission and never will. This commitment supports the new deployment options where data does not leave client infrastructure. For practitioners deploying these models, the trade-off involves balancing the enhanced capabilities and reduced friction of Fable 5.1 against the nuanced behavioral shifts noted in the system card, particularly regarding how the model handles authorization claims and potential misuse scenarios.

AfterQuery closed a funding round valuing the company at $3.2 billion, a more than 10x increase from its $300 million valuation in April. This growth occurred over a five-month window following the startup's $30 million Series A announcement. Y Combinator partner Gustaf Alströmer confirmed this trajectory represents the quickest launch-to-unicorn timeline for any accelerator cohort participant. The founders, currently aged 22 and 23, graduated from the Winter 2025 cohort only 18 months prior.

The San Francisco-based firm reported an annualized revenue run rate of $100 million as of April, citing engagements with major AI laboratories. Disclosed customers include Nvidia, Legora, and the Korean AI lab Motif Technologies. Unlike predecessors such as Mercor and Scale that focus on employing knowledge professionals to verify model accuracy, AfterQuery utilizes these specialists to train models and agents on task execution workflows. The company describes this methodology as encoding the patterns, decisions, and reasoning processes of top practitioners rather than simply ensuring correct answers to queries.

Forbes first reported the details of the latest funding round. Attempts to reach AfterQuery for direct comment regarding the valuation or operational specifics were unsuccessful at the time of reporting. The rapid capitalization underscores a market shift toward training data that captures professional decision-making logic rather than static factual correctness. This distinction positions the startup within a specific niche of the AI supply chain focused on agent behavior modeling.

The speed of this valuation jump highlights intense competition for high-quality human feedback data tailored to complex reasoning tasks. While many startups leverage expert labor for reinforcement learning from human feedback (RLHF), AfterQuery's explicit focus on replicating professional workflows for autonomous agents suggests a pivot toward next-generation system capabilities. The involvement of hardware giant Nvidia alongside specialized labs indicates broad demand across the infrastructure and application layers.

Google announced the immediate rollout of Google Pics, a new image-creation and editing tool integrated into the Google Workspace suite. The service, powered by the company's Nano Banana image-generation model, becomes available today within Google Docs and Google Slides for business customers and premium Google AI subscribers. Expansion to Google Drive is planned for a later date, while general availability for most Workspace users and those on Google AI Pro or Ultra tiers will occur over the coming weeks.

The product targets everyday design tasks such as creating posters, social media posts, and illustrations, positioning itself against Adobe Express and Canva through a fundamentally different interaction model. While Canva relies on a marketplace where creators publish templates and graphics for royalties, and Adobe Express facilitates manual design from scratch, Google Pics operates exclusively on text prompts to generate assets. The system was trained on artists' work but does not offer a mechanism for those artists to earn royalties within the platform, distinguishing its economic structure from competitor marketplaces.

Beyond generation, the application includes specific editing capabilities such as object isolation, object transformation, and the modification or translation of text embedded within images. The workflow supports collaborative editing and allows users to request multiple generations of a single prompt to select the optimal output. By nesting these capabilities directly into existing productivity documents rather than requiring a separate design environment, Google shifts the creation locus from a dedicated design app to the document editor itself.