16 games join GeForce NOW this May, including AAA titles like Forza Horizon 6 and 007 First Light on launch day. Ultimate members get expanded RTX 5080-class performance across more titles, and Firaxis' 30th anniversary brings classics to the cloud.

Starting today, agents can now be Cloudflare customers. They can create a Cloudflare account, start a paid subscription, register a domain, and get back an API token to deploy code right away. Humans can be in the loop to grant permission, but there’s no need to go to the dashboard, copy and paste API tokens, or enter credit card details.

TurboQuant has recently been launched by Google as a novel algorithmic suite and library for applying advanced quantization and compression to large language models (LLMs) and vector search engines — an indispensable element of RAG systems.

This article explains the differences between APIs (for software applications) and MCPs (Model Context Protocol for large language models), including how each works, why MCPs are not just API wrappers, when to use each, and the role of gateways in security.

This article explores the limitations of AI code review, showing that structural analysis catches only about 50% of security defects, while the other half are design flaws invisible without understanding intent. The author shares a personal anecdote (a bus tracker app that showed the wrong direction) and a real bug found in Google's Gson library to illustrate the importance of requirements. The open-source Quality Playbook method is introduced, which derives requirements from code and documentation to catch hidden bugs.

An AI agent that replaces a thousand SaaS tools.

The article argues that the next evolution of agentic SaaS is not more complex tool infrastructure but providing a complete command-line interface (CLI). The author points out that current attempts to constrain LLMs with carefully scoped toolsets are inefficient because LLMs are already fluent in shell and command-line environments. Going forward, every SaaS should ship a parallel command-line surface for its non-human users, i.e., AI agents.

Australia's financial regulator warns that AI agent governance is poorly managed as banks and super funds expand AI use. APRA's review identifies gaps in risk management, model monitoring, and cybersecurity. Regulators call for better board oversight and controls for agentic workflows.

Welcome to our monthly digest, where you can catch up with any AIhub stories you may have missed, peruse the latest news, recap recent events, and more. This month, we meet PhD students and early-career researchers, find out how machine learning is used for particle physics discoveries, cast an eye over the latest AI Index Report, and watch a robot beating elite players at table tennis.

AstaBench’s latest update adds new frontier-model results, including GPT-5.5, and highlights growing adoption from groups including the UK AISI, General Reasoning, Elicit, SciSpace, Distyl AI, and EvoScientist.

A summary and discussion of last week's big AI news, including OpenAI's new image model, Alibaba's Qwen 3.6 Max, SpaceX-Cursor deal, and more.

For decades, psychologists have debated whether the human mind can be explained by one unified theory or must be broken into separate parts like memory and attention. A recent AI model called Centaur seemed to offer a breakthrough, claiming it could mimic human thinking across 160 different cognitive tasks. But new research is challenging that bold claim, suggesting the model isn’t truly “thinking” at all—it’s just memorizing patterns.

MIT President Sally Kornbluth discusses the importance of basic research, funding challenges, and the role of universities in a live podcast conversation.

Draft is a privacy-first local workspace that captures AI conversations from ChatGPT, Claude, Gemini, and more, turning them into editable, searchable notes. It offers text-to-speech, offline access, and sharing. Currently in free beta.

PollyReach gives your AI agent a real phone number and voice to make and receive calls, enabling autonomous phone communication for business tasks.

A quiet day allows reflection on the growing implications of the inference age, with industry leaders highlighting the shift from training to inference compute and the rise of coding agents as platforms.

This week OpenAI faced a triple blow: Musk's lawsuit threatens its nonprofit-to-profit conversion, revenue miss triggers market selloffs, and AWS deal ends Microsoft exclusivity. Meanwhile, DeepSeek price war, Big Tech layoffs, and White House plans to bypass Anthropic's safety flags signal shifting pricing power and regulatory landscape.

Together AI and Adaption partner to bring Together Fine-Tuning natively into Adaptive Data, helping teams optimize datasets, run fine-tuning, evaluate results, and deploy stronger open models.

Together AI details their rapid response to the Linux kernel vulnerability Copy Fail (CVE-2026-31431), which allowed local unprivileged users a precise 4-byte write primitive via the AF_ALG interface, leading to privilege escalation. The team mitigated by unloading the vulnerable kernel module, rolling out patches, and enhancing detection, ensuring AI infrastructure security.

Apple is presenting new research at the annual ICASSP, taking place in Barcelona, Spain, from May 4 to 8, 2026. The company is sponsoring the conference and will showcase three papers covering audio-visual speech models, spatial audio generation, and speculative decoding.

Researchers at Apple and Gallaudet University developed a pseudo-annotation pipeline to address the scarcity of high-quality annotated sign language data. Their approach uses a fingerspelling recognizer, isolated sign recognizer, and a K-Shot LLM to generate likely annotations from signed video and English input. They achieved state-of-the-art results on FSBoard (6.7% CER) and ASL Citizen (74% top-1 accuracy) and are releasing nearly 500 human-annotated videos and over 300 hours of pseudo-annotations.

Apple ML Research presents STARFlow-V, a normalizing flow-based video generator offering end-to-end learning, robust causal prediction, and native likelihood estimation. It uses a global-local architecture in spatiotemporal latent space, flow-score matching, and video-aware Jacobi iteration to achieve strong visual fidelity and temporal consistency, demonstrating the first evidence that normalizing flows can produce high-quality autoregressive video generation.

A new debiasing technique called WRING avoids creating or amplifying biases that can occur with existing debiasing approaches.

Since introducing Empirical Research Assistance in fall 2025, Google Research scientists have applied it to real-world problems in epidemiology, cosmology, atmospheric monitoring, and neuroscience, demonstrating AI's transformative potential to accelerate scientific discovery.

Liminary turns everything you’ve saved into working memory for AI, surfacing relevant context across writing, meetings, and research. Founded by a former Dropbox ML engineering lead, it helps knowledge workers reuse their best thinking with traceable citations.

How goblin outputs spread in AI models: timeline, root cause, and fixes behind personality-driven quirks in GPT-5 behavior.

AWS developer advocate Morgan Willis demonstrates how intent-driven tools and semantic search reduce token consumption from 52,000 to 2,000 and further halve it in Strands Agents.

The article explores the rising cost of AI evaluation, especially for agent benchmarks, showing that evaluation has become a new compute bottleneck. Static benchmarks can be compressed 100-200x, but agent and training-in-the-loop benchmarks resist compression. Reliability demands repeated runs, multiplying costs. High evaluation costs risk concentrating validation power in well-funded labs.

IBM's Granite 4.1 family of dense decoder-only LLMs (3B, 8B, 30B) trained on ~15T tokens with multi-stage pretraining, long-context extension up to 512K, supervised fine-tuning on ~4.1M curated samples, and multi-stage RL with on-policy GRPO+DAPO. The 8B instruct model matches/surpasses previous 32B MoE model. All models under Apache 2.0.

OpenAI scales Stargate to build the compute infrastructure powering AGI, adding new data center capacity to meet growing AI demand.