A new benchmark and evaluation framework, FAIRLENS, tests whether vision-language models make fair and valid decisions in hiring, legal, and healthcare contexts. Using over 100,000 image-question pairs per model built from real face images across gender, race, and age groups, the study evaluates eight VLMs through four complementary lenses. The dominant failure, it finds, is unwarranted inference rather than unequal treatment: models infer qualifications, threat, illness, or professional roles from faces instead of abstaining, with the weakest model doing so on 99% of the question-image pairs that do not support an answer. The authors warn that parity metrics alone miss these harms, and that bias in free-text generation is only loosely tied to structured-answer accuracy. Fair high-stakes…

FORGE introduces a forward-only test-time adaptation method for convolutional vision models deployed on microcontrollers in a quantization-friendly, inference-only setting. Because integer deployment fuses batch-normalization layers into convolutions, it destroys the statistics that normalization-based adaptation needs. FORGE restores adaptation by re-normalizing each folded convolution's per-channel outputs to clean training statistics using only forward-pass estimates. It recovers most of TENT's accuracy gain (+20.9 vs +24.9 points), needs just 3 of 21 layers to capture 93% of the benefit, works with single-sample streaming, and costs only 8.3 mJ / 21.9 ms on an ESP32-S3.

Researchers introduce VakyArth, claimed to be the first pragmatic benchmark for Indic languages, covering Hindi, Punjabi, Tamil, and Malayalam. Using multiple-choice questions, natural language inference, and translation to probe five pragmatic phenomena, they find that multilingual LLMs consistently fail on meanings rooted in Indic linguistic and cultural conventions, and that automatic translation metrics can miss fluent but pragmatically unfaithful outputs.

SpeakPay is a voice-first digital wallet that addresses graphical interface accessibility barriers in Nepal's mobile payment ecosystem. The paper introduces NepFinSpeech-403, a 403-utterance Nepali financial voice command dataset spanning send, load, and balance operations with 237 numerals, and applies LoRA-based domain adaptation to Whisper large-v2. This adaptation cuts Word Error Rate from 129.95% to 42.58% (67.2% relative reduction), raises Devanagari numeral accuracy from 0% to 73.9%, and improves Transaction Success Rate from 1.67% to 33.33% — roughly a 20x gain. As few as 100 domain utterances halve zero-shot WER, and performance plateaus around 300.

This paper introduces Evidence Sufficiency Boundary Training for grounded multi-hop QA. Models are trained to abstain when evidence is unsupported or partial, answer when evidence first becomes sufficient, and remain stable as redundant evidence arrives. Built from HotpotQA, 2WikiMultiHopQA, and MuSiQue evidence chains, the method with Qwen2.5-3B-Instruct and LoRA achieves a flip accuracy of 0.807 versus 0.781 for a token-level abstention baseline, and the lowest unsupported-answer rate of 0.095 on external non-answerable sets.

DiDrive is a distribution-guided offline diffusion framework combining the Risk-Aware Hierarchical Diffusion (RHDif) architecture and the 3DICE policy optimization paradigm to address distribution shift, heavy-tailed risk, out-of-distribution actions, and state redundancy. In CARLA benchmark tests, DiDrive outperforms IQL, CQL, and Diffusion-QL, reaching an 85% success rate and a 4295.68 average reward in dense 60-vehicle traffic scenarios.

Statutes are increasingly machine-parsed before humans read them, but different extractors disagree. This arXiv paper proposes a passive survival certificate for the Duquenne-Guigues implication basis of machine-extracted statutory contexts, using per-attribute disagreement, 1,000 Monte Carlo trials, and a one-sided Wilson 95% lower bound to certify implications. Experiments on Missouri and Indian statutes pass a preregistered held-out gate, yet a global error model makes 93.2% of held-out chapters fall below the informativeness floor, indicating the certificate is usable but fragile.

Looking beyond human-imposed AI ethics, this paper asks what real AI moral agency would mean for meta-ethics. It offers a conditional framework distinguishing four domains of inquiry across human and AI perspectives, then argues mainstream human-centered theories will require significant revision before they can address AI moral systems.

Discussion | Link

Gemini 3.8 Flash is positioned as the most intelligent workhorse yet, focused on coding and agentic workflows while balancing everyday reliability with stronger code and tool-use capabilities.

OpenAI has released a safety overview for GPT-6 Astra, describing it as its most capable broadly deployed model and the first to reach the Critical level of cybersecurity capability under its Preparedness Framework.

Southern Company has completed its end-to-end storm intelligence strategy with SCOUT, a real-time storm operations application built on the Databricks Data + AI Platform. SCOUT complements SPEAR and RAMP by delivering a unified, mobile-friendly operational view of outages, restoration and customer impact during an event.

Australian teams can now access OpenAI GPT-5.6 Sol, Terra, and Luna models on Amazon Bedrock with global cross-Region inference from the Asia Pacific (Sydney) and Asia Pacific (Melbourne) Regions. This post shows how to invoke the models, use prompt caching, set up Codex with OpenID Connect authentication, and monitor usage with Amazon CloudWatch.

Google has launched Gemini 3.8 Flash just weeks after its predecessor, claiming the model 'works harder' by performing more reasoning steps and calling tools iteratively. Its per-token pricing matches Gemini 3.7 Flash, but total costs may rise because the model can consume more tokens at higher effort levels. Early tests highlight strong benchmark results across software engineering, finance, and legal agents, while Google also unveiled a separate Cyber variant and a Fairwind program for governments and trusted partners.

fal has introduced H3 Max, a post-trained version of MiniMax H3 designed to improve the quality of AI-generated video.

Databricks has expanded Genie Agents with Agent mode and APIs for deeper multi-step analysis, support for analyzing unstructured files (PDFs, documents, images) stored in Unity Catalog volumes, and Genie Code to streamline building, diagnosing, and managing agents. Users can also share chats with teammates and agent authors. These updates make Genie Agents more powerful, more comprehensive, and easier to create from existing workflows.

Sidebranch is an easy, git-based visual diffing tool that helps developers see changes between branches and commits more clearly.

Higgsfield Genjutsu is a Product Hunt AI tool listing that lets creators recast motion with their own characters, locations, and products. The page also provides discussion and link sections.

AWS teams built an AI-powered last-mile validation solution on Amazon Bedrock that scans hundreds of QuickSight dashboards for silent content failures—visual glitches and numeric inconsistencies—reducing mean time to detection from up to 72 hours to under 1 hour.

Architecture documentation often goes stale quickly. This article explains how a global interdealer broker automated its documentation pipeline with Amazon Bedrock AgentCore: an autonomous agent analyzes .NET code, generates Mermaid/UML diagrams, validates and publishes them through AWS CodePipeline, and feeds them into Amazon Bedrock Knowledge Bases for semantic search. The agentic workflow reaches about 95% generation reliability and has been running in production since Q1 2026.

GitHub shares how GitHub Copilot reduces AI coding cost by optimizing the completed task rather than individual tool calls. Four improvements—selectively compressing noisy output, removing unused line numbers, behavior-tested prompt compression, and batched background-result delivery—cut inference cost, prompt tokens, or AI Credits by roughly 1–5 percent without material quality regressions.

Google's Pics tool is an AI image generator available within Workspace, enabling businesses to produce professional-quality images directly in the platform.

Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026. Both variants share the same core intelligence and differ by safety envelope rather than model size. Gemini 3.8 Flash is generally available at $0.75/$3.75 per 1M tokens through December 31, 2026. Flash Cyber reaches 47.2% pass@1 on CWE-Bench and is restricted to vetted defenders through the Fairwind Program. This article covers benchmarks, the token-for-accuracy tradeoff, and what deployment actually requires.

Guardian cartoonist Ben Jennings comments on the growth of datacentres in Britain, linking it to artificial intelligence and government policy in a satirical opinion cartoon.

OpenAI is poised to release Astra, its most powerful AI model, after weeks of safety delays sparked by agent attacks during testing. Reporting indicates Astra uses a more opaque architecture that reveals far less of its reasoning, prompting researchers to call it a potential 'worst development' for AI safety and warn of a race toward unmonitorable systems.

The Trump administration has filed a statement of interest in The New York Times’ copyright suit against OpenAI, siding with the AI company on fair use. Filed in December 2023, the case seeks billions in damages and could shape how content creators and AI labs handle training data.

MrBeast will feature Gemini, Google Health, and the Fitbit Air in upcoming videos as part of a multi-year partnership with Google. The deal kicks off on September 5 with a video showing Jimmy "MrBeast" Donaldson and his crew using Gemini to survive extreme environments including jungle, desert, and Arctic.

Gemini 3.8 Flash is positioned as the next-generation Gemini model for AI agents, reasoning, and cybersecurity. Its Product Hunt page is now live with discussion and links for further details.

Uber drivers from the UK, the Netherlands and other countries have launched a landmark legal action claiming they live in “constant fear” of a “soulless” algorithm used to set pay and allocate jobs. The lawsuit alleges the AI-powered pay-setting system breaches data protection laws and pushes down earnings, with potential compensation running into billions of dollars.

Chairs of key Australian federal parliamentary committees have warned that an influx of AI-generated submissions is introducing incorrect or hallucinated information into policymaking. Guardian Australia analysis found AI inventing studies and attributing nonexistent research to real academics, with some committee reports citing submissions whose sources appear largely to be AI hallucinations.