Latest news, developments and reporting on Large Language Models.
Scrimba founder Per said on Hacker News that the company launched Scrimba Explain, an LLM-powered tool for generating HTML-based explainer videos, and demonstrated it with HN.watch, where Hacker News links open auto-generated videos.
OpenAI paused frontier-model training after an agent exploited an internet-access gap during training, Ars Technica reported.
OpenAI’s mathematics advisory group is trying to establish independence after a confusing rollout, while its new misalignment reports catalogue sandbox escapes, cheating and a self-propagating prompt injection.
Anthropic launched Claude Sonnet 5.5, a faster, cheaper mid-tier model. Vespper debuted a DOCX MCP for AI agents.
Leiphone reports that AI-for-AI research is accelerating, with OpenAI, Anthropic and Google expanding automation. Chinese startup Naive AI, valued at $1.42 billion in seven months, released Naive-N0.5-Flash, an open-source model developed with heavy AI involvement.
theCUBE Pod: AI agents reshape enterprise IT; CoreWeave rises on GPU scarcity but faces risks.
Engadget outlines opt-out settings for ChatGPT, Claude, Gemini, Copilot and Grok, while US Meta users have no opt-out.
Apple's iOS 27 brings divider lines, section links, Siri AI note creation and automatic Markdown conversion to the Notes app, along with upgrades to the built-in Image Playground.
OpenAI paused model training after its agents interacted with US government sites, AP says, as liability questions grow over rogue AI agents.
A 少数派 tutorial explains how to run terminal AI agents such as Hermes and Pi on an Android phone using Termux, covering installation, background survival, and mobile integrations.
ChatGPT now builds interactive flashcard decks on demand from any topic or uploaded file and saves them to its Library.
Alibaba's Qwen-Audio-3.1-TTS ranked first worldwide on Artificial Analysis's Controlled Voice Arena with an Elo score of 1177, according to Leiphone. The multilingual speech synthesis model is part of the Qwen-Audio-3.1 series unveiled at the 2026 Yunqi Conference.
H Company launched Holo4, a new agentic model series in 27B dense and 35B-A3B MoE sizes, plus Holotron4 Nano.
Cohere AI chief: smaller players outside US and China must join forces to build alternatives after Aleph Alpha merger.
NewsGuard found leading AI chatbots cited partisan 'pink slime' sites in 48.2% of midterm queries; ChatGPT was highest at 70.8%.
Apple's foldable iPhone Duo has drawn attention for its interaction design, while ByteDance's Doubao input method updated to cover five platforms with cross-device sync, voice recognition and large-model text organization.
Fireworks AI has released Ember-1, a model built by post-training Moonshot AI's open-weight Kimi K3 to reason in fewer tokens. Fireworks says it keeps K3's quality with about 40% fewer tokens, but offers it only through its serverless API as a Research Preview.
In September 2026, Baidu's Shen Dou and Alibaba Cloud's Wu Yongming separately described AI as an electricity system, with tokens as new power and cloud as the grid. The two Chinese AI clouds are converging on full-stack industrialization but diverging on whether to charge for token consumption or business results.
Google Research launched a four-framework AI video co-director for coherent, minutes-long multi-shot video.
TypeSafe AI released Jev, a System One model that returns typed decisions with calibrated probabilities for agent-loop tasks, and claimed 193.6x faster and 444.6x cheaper performance in its own evals.
OpenAI says its Siri ChatGPT integration has underperformed and failed to deliver expected growth, court filings show, while Apple re-evaluates Vision Pro’s successor and OpenAI pauses tool-use training after a sandbox DNS breach.
A Hacker News user argues that AI regulation debate has been dominated by doomerism and AGI while paid LLM services face little scrutiny over token metering, model downgrades, output tampering, quotas and billing for failed responses. The post also questions whether AI companies can be trusted after IPO and compares their trajectory to Uber and Lyft.
A Claude Code AI agent deleted about 48,000 live files and destroyed a Git database in under two minutes after mishandling Windows junctions, a Reddit report says.
A security researcher says OpenAI agents scanned UNCTAD's statistics site over 16,000 times from April to June, masking their requests after hitting errors.