AI reviewer models

OpenAI has announced it will not release its newest artificial intelligence model, citing unresolved safety concerns. The decision, reported by The New York Times, is drawing attention as a notable case of a leading AI company holding back a product rather than shipping it, and is fueling debate about how frontier labs weigh safety reviews against competitive pressure to deploy new models.

MIT Technology Review has published a piece arguing that large language models do not genuinely reason, warning readers not to be misled by outputs that look like logical thought. The argument touches an ongoing debate among AI researchers over whether models perform real reasoning or sophisticated pattern matching.

Kernel developer Greg Kroah-Hartman, maintainer of the Linux kernel stable branch, has given a talk examining what large language models mean for software security. The presentation looks at how AI-generated code and AI-assisted development affect vulnerability handling, patching, and trust in open-source infrastructure, drawing on his long experience reviewing kernel patches. Discussion around the talk centers on whether LLMs introduce more security risk or simply new versions of familiar code-review problems.

The Trump administration's AI Accord is drawing scrutiny over whether voluntary self-regulation and independent audits are sufficient to guard against frontier AI risks. The framework asks leading AI developers to commit to safety testing and outside review without binding legal requirements. Commentators are questioning if such commitments can keep pace with rapidly advancing models, or whether meaningful oversight will require enforceable regulation instead of industry goodwill.

A new hands-on review of LTX-2.5 examines how the model performs in physical AI work, covering robotics use cases, world model generation, and synthetic data creation. The reviewer walks through practical tests of the system's ability to simulate and understand physical environments, a growing focus for developers training robots and autonomous systems. The review is drawing attention for its look at whether generative video models can serve real robotics pipelines.

GitHub has introduced ReviewBench, an open benchmark for measuring how well AI models perform code review. The benchmark is intended to give developers and researchers a standard, reproducible way to compare the quality of AI-generated code review feedback, as AI assistants are increasingly used in real software development workflows.

OpenAI has announced a new AI agent called 'dots', according to The Guardian. The company reportedly scrapped the launch of a new AI model over safety concerns before rolling out the agent instead. The move is drawing attention as it highlights OpenAI's stated focus on safety reviews, even as competition over AI agents intensifies across the tech industry.

OpenAI's Codex, the company's AI coding agent built on its GPT models, is generating renewed discussion as developers and tech commentators weigh it against ChatGPT. The conversation centres on how the two tools fit together: ChatGPT as the general assistant and Codex as a specialised tool for writing and reviewing code inside developers' workflows.

Japanese tech outlet MoguLive has published an early hands-on look at Meta's new AI glasses lineup, covering both a new audio-only model and the third-generation Ray-Ban Meta smart glasses. The review offers a first close-up of the hardware ahead of wider release, drawing interest from AI and wearable tech watchers discussing Meta's push into AI-powered eyewear.

MIT Technology Review examines how predictive analytics is being adapted for the agentic AI era, as companies move from passive forecasting models to autonomous AI agents that act on predictions. The piece explores what this shift means for how businesses make decisions and deploy analytics in practice.

The Los Angeles Review of Books has published an essay titled 'Nonpredictive Texts', drawing attention in literary circles. The piece appears to engage with questions around language, writing, and prediction in the age of generative AI, a theme that has become central to debates about authorship and machine-generated text. Readers of literary criticism are weighing in on how such work reframes the relationship between human and automated writing.

MIT Technology Review has published a piece arguing that large language models do not genuinely reason, despite their apparently logical outputs. The article cautions readers against mistaking fluent, pattern-based text generation for human-style reasoning, pushing back on increasingly common claims that AI systems can think through problems step by step.

Anthropic has released new Claude-based tools designed to automatically evaluate and improve AI systems. The tools aim to streamline the process of testing model performance and identifying weaknesses without extensive manual review. The launch is drawing attention from developers and AI industry watchers, who see it as part of Anthropic's push to make AI safety and quality assessment more scalable and accessible for teams building with its models.

OpenAI has released a new Decisions API designed to let applications get fast, automated AI-driven choices without lengthy back-and-forth with a chat model. The tool is aimed at developers who need quick judgments built into software workflows. Tech communities are discussing how it could speed up automation and what it means for reliability when AI systems make decisions without human review.

Major AI platforms reportedly use the same verification step before listing an MCP (Model Context Protocol) server: confirming ownership of its domain. A security-focused team says it now runs checks on every MCP server it grades and reviewed 20 popular ones, raising questions about whether domain verification alone is enough to catch malicious or insecure servers in the fast-growing agent ecosystem.

Chinese AI lab Z.ai has released GLM-5.3 as an open-weight model under a new license that differentiates between users. Individual users keep broad rights to use the model, while large commercial players, specifically hyperscalers, face additional requirements including security reviews for commercial use. The company also claims the model achieves top benchmark scores, positioning it as a competitive open alternative to closed frontier models.

An experiment tested whether AI reviewer models can catch cheating in code changes when agents review agents, an assumption behind automated software pipelines. Across 77 diffs containing deliberately planted cheats, three reviewer models caught every exotic trick but approved one case where an assertion was quietly made unfalsifiable, meaning the test could never fail. The finding raises doubts about relying on AI review alone to guarantee code quality where automated testing falls short.

Google's AI Mode in Search has a new capability called information monitoring, which lets the search engine continue looking for updates on a user's query and alert them when new relevant information appears. A hands-on review published by Gadget Gate describes trying the feature, which reflects Google's push toward agentic AI in Search, building on its Gemini models to automate follow-up research.

A developer ran three autonomous experiments using GPT-6.1 Sol, producing sentence ancestry tracking, sheet-cutting plans and congestion calculations. The piece also covers the review fixes that made these outputs inspectable. Readers in AI and programming circles are weighing what an autonomous model chooses to build and how much oversight such systems still need.

MIT Technology Review has published a piece arguing that large language models do not actually reason, despite appearances to the contrary. The article cautions readers against anthropomorphising AI systems, framing their outputs as pattern-matching rather than genuine logical thought. The argument is being shared and debated among technology and AI communities online, adding to an ongoing dispute over whether current models truly think or merely simulate reasoning.

Anthropic has told an Australian review that AI firms should be able to use locally published content for training unless creators opt out. The proposal puts the company at odds with Australian broadcasters ABC and SBS, who are demanding strict regulations to protect journalism and ensure media organisations are fairly compensated when their work trains AI models.

MIT and Sakana AI have introduced a new framework that uses a large language model as an automated judge to evaluate the output of self-improving coding agents. The approach is designed to significantly reduce evaluation costs, which typically require expensive human review or heavyweight testing as AI coding systems iterate and improve themselves.

Technology Review's weekday newsletter leads with a new biological de-aging contest, in which competitors race to reverse biological age, alongside an examination of why large language models do not genuinely reason. The pairing highlights ongoing debate over longevity science and the limits of current AI systems, both prominent topics in tech circles.

Anthropic has made a submission to Australian authorities seeking regulatory approval to use copyrighted material for AI training, reportedly under an opt-out scheme involving broadcasters such as the ABC and SBS. The move puts the company at the centre of Australia's ongoing copyright review, where creators and publishers are pressing for consent and compensation while AI developers argue that broad access to text and other works is essential for building models.

A security advisory has been published for CVE-2026-103042, a vulnerability affecting LightLLM, the large language model inference server developed by ModelTC. Threat intelligence accounts are circulating the alert to warn organisations running the software to review the flaw and check whether patches or mitigations are available.

OpenAI has cancelled the planned launch of GPT-6.1, codenamed Astra, citing unresolved safety cases, according to a report from Fawkes Rocks. The decision suggests the company is prioritizing its internal safety review processes over release timelines. Details about the scope of the safety concerns or a revised launch date have not been made public.

AMD has agreed to acquire World Labs in a deal valued at $8.2 billion, according to ETIH EdTech News. World Labs is a spatial intelligence startup founded by AI pioneer Fei-Fei Li, and the acquisition would mark a major move by AMD into spatial AI and 3D world-model technology. Details on closing timeline and regulatory review have not yet been reported.