All AI news

AIThe final post in O'Reilly Radar's four-part post-training series walks readers through implementing the classic ChatGPT pipeline on Qwen2.5-1.5B, covering SFT, reward model training, and PPO. The walkthrough uses torchtune for SFT and verl, a Ray-based RL framework from ByteDance's team, for reinforcement learning. The author says the goal is hands-on understanding rather than reproducing InstructGPT, which took a large team and thousands of GPU-hours.

AIOpenAI announced hundreds of mathematical breakthroughs, weeks after claiming it had solved one of the most complicated problems in mathematics. The findings raised questions about whether the model used creative thinking or only completed the final steps of human work. Experts say AI could be revolutionary for mathematics if it provides proofs, since proof techniques often underpin other breakthroughs.

AIWhat I like the most about this new architect is its ability to proactively look for improvements and come back with a drafted proposal that’s already validated. Think about that for a second. The architect looks at your agents, how they work, their conversations, and comes back to you with a plan to make them better.

AIThe FCC's proposed ban on new Chinese optical transceivers targets the top of the supply stack, but the author argues it leaves the dependencies that matter untouched. The analysis traces the module, laser, indium phosphide wafer, and indium metal layers, finding that China controls the wafers and refined indium while U.S. firms depend on Chinese-made substrates. The author concludes that a module-level rule would take years to replace lost capacity and would not change control of the lower layers.

AIGoogle and Meta have started sending some live traffic through a land route across Iraq that they had previously held in reserve, according to a person familiar with the deal. Most data between Europe and Asia still flows through subsea cables under the Red Sea, where Yemen's side of the strait is now contested and cable repairs could take months.

AIChatGPT release notes briefly listed "Audio uploads in ChatGPT" for paid subscriptions and workspaces, covering transcripts, recording summaries, and questions about recording contents. The feature has since been removed again.

AIA new teen mode in OpenAI's ChatGPT includes a study tool but still completes users' homework. A children's safety nonprofit says the chatbot failed its tests.

AIMichael Smith, a 54-year-old from North Carolina, was sentenced to 18 months in prison after pleading guilty to streaming-royalty fraud using automated bots and AI-generated songs. The U.S. Justice Department also ordered him to forfeit the $8,091,843.64 he received in royalties. Prosecutors said the scheme diverted royalty payments from genuine artists.

AIElon Musk says SpaceX will use the best back-end model for each task on Grok Bot, including Claude Opus 5.5, MidJourney, Suno, and other leading APIs. Matt Shumer calls this extremely good news for bot users.

AIMistral released Mistral Large 4 "Le Chonk", a 1T-parameter (49B active) multimodal model, with open weights planned in about three weeks. Google rolled out Nano Banana 2.1 across Gemini, AI Studio, and the Gemini API, and released EmbeddingGemma 2, a 740M-parameter open multimodal embedding model under Apache 2.0. OpenAI launched the Decisions API in beta with gpt-6-luna, returning typed answers 10x faster than the Responses API.

AITrump announced a voluntary agreement under which major AI companies will use internal controls, monitoring, outside audits and board oversight to manage risks. The White House calls the commitments "morally binding," but the accord creates no comparable system of legal enforcement.

AIBooks by People, a British startup, has received UK Intellectual Property Office approval to distribute a mark certifying that a published book was written by a human, displayed on the cover as a thumbprint-style stamp. The startup verifies manuscripts through an authorship declaration, proprietary software, and review of drafts and research materials, and five publishers have signed up so far. The mark is expected to appear on its first book by the end of the year.

AISystem76's COSMIC desktop now requires contributors to declare no LLM-generated content in pull requests, including code, comments, and descriptions. GNOME Calendar and GNOME Extensions also restrict AI-generated contributions, while GNOME developer Michael Catanzaro argues the project should accept AI-generated bug reports. Catanzaro's case rests on memory-unsafe languages such as C, C++, and Vala, and he has shortened GNOME Security's disclosure deadline from 90 days to 30, effective August 1.

AIAi2 has published its Bolmo byte-level language model research in Nature and released new checkpoints on Hugging Face. The byteifying process converts an existing subword model into a byte-level one with a relatively short additional training run, and the paper reports that it also works for Qwen 3 8B and Llama 3 8B, producing Bwen 8B and Blama 8B. Ai2 also released Stage 1 checkpoints for researchers extending the architecture.

AIZeptap is now live on Product Hunt as a Mac app that lets AI agents use an iPhone the way a person would. The team describes it as computer use for iPhone and is offering a free 7-day launch license key to those who DM them.

AIElon Musk said SpaceX will use the best back-end model for each task, including Claude Opus 5.5, Midjourney, and Suno, for Grok's responses. The main post from Max Zeff only says "Interesting," so the summary is limited to Musk's stated routing plan.

AIElon Musk says SpaceX will use the best back-end model for each task, including Claude Opus 5.5, Midjourney, Suno, and other leading APIs. The main post is only an eye emoji, so its own content adds nothing beyond the quoted note about Grok's routing.

AIGrok Bot will now use the best backend model for each task, including Claude Opus 5.5, MidJourney, Suno, and other leading APIs. The change is framed as choosing whatever is most likely to produce the best outcome for users.

AIRadisson Hotel Group partnered with Accenture to build a ChatGPT plugin using OpenAI technology. The plugin helps travelers find, compare, and book hotels while planning their trips.

AIMeta has open-sourced Rebalancer, a C++ library with a Python interface for solving assignment problems under constraints and objectives, released under Apache 2.0. The article reports that Meta has used it for resource allocation for over 9 years and runs about 40 million problems a day, with P99 solve time of 12 seconds on 265k objects and 3.2k bins. The package can be installed with pip install rebalancer, though PyPI still classifies it as Alpha.

AIMeta and Sierra announced the Personal Agent Protocol, an open standard for how personal AI agents find and transact with businesses on a user's behalf. The author says it defines discovery, OAuth-based sessions, and a choice among website, API, or company agent routes, and distinguishes it from MCP, which connects agents to tools and data, and A2A, which hands tasks to another agent.

AITeknium says Hermes was never intended to be just a personal assistant that reads emails, but the most powerful AI agent for all users. He says consumers are one audience, while scientists, cybersecurity engineers, developers, knowledge workers, and creatives are also intended users.

AITeknium says a plugin from @altryne's current project should be added to the plugin catalog. The post is a brief endorsement and does not describe the plugin's functions. Background from @tonysimons_ says Hermes is getting a local video editor for editing user footage with 42 FFmpeg scripts and no cloud or API key required.

AIHermes Agent received a set of minor enhancements aimed at improving its voice mode this week. The post gives no specific features, figures, or benchmarks.

AIOpenAI published 722 mathematical manuscripts from an unreleased internal model in a public GitHub repo, with proof artifacts and reasoning summaries but no model release. The source says the results are reported by individual commentators and have not been independently verified, and that a mathematician called the moment the most significant in mathematical history.

AIJake Boggan, a Hacker News commenter, reacted to reports that Barnette's Conjecture, a graph theory problem he spent years studying, has been proven, as listed in openai/math problem 180. He said he had spent thousands of hours on the problem and had briefly believed he solved it last summer. He described the news as bittersweet.

AIvLLM Semantic Router and KR Labs released Vela 2.0, which makes span-level decisions covering safety checks, domain classification, PII spans, and unsupported claims in a single call. It comes in four sizes from 0.3B to 9B parameters under the Apache-2.0 license.

AILangChain revamped skills support in Deep Agents with three changes: tools bound to a skill load only when the agent reads that skill, pinned skills are loaded before the next model call when a user requests them, and long-running threads can pick up new or changed skills without restarting. Each skill is a folder with a SKILL.md file, and only its name and description are in context until the agent reads the full instructions.

AIAnthropic released Claude Haiku 5.5, which it calls its cheapest, fastest, and most capable small model. It costs around 75% less to run than Haiku 4.5 and is aimed at high-volume, cost-sensitive tasks such as summaries and classification. The release also cuts Sonnet 5.5 cache read prices by 50%, and the model is available on AWS, Google Cloud, and Microsoft Azure.

AIEvery launched the Every Agent, a Slack-based agentic coworker whose token costs it passes on to customers without markup. Engineer Paridhi Agarwal explains how she made the agent more token-efficient, and the newsletter says the company moved from personal agents to a single shared company agent.

AIAnthropic released Claude Haiku 5.5, which scores 43 on the Artificial Analysis Intelligence Index, up 26 points from the last Haiku release. Pricing is $0.10/$0.50 per 1M input/output tokens up to 100k tokens, rising to $0.50/$2.50 above that, but at max effort it uses about 162k output tokens per Intelligence Index task, roughly 3x GPT-6 Luna.

Why it matters: The benchmark shows Haiku 5.5 scores well but uses far more output tokens than GPT-6 Luna, so cost per task matters beyond list price.

AIMastra Connect is a public beta that lets Mastra projects connect providers such as Linear, Notion, and Slack, giving agents and workflows ready-made tools. Connect launches with 23 providers, almost 900 tools, and 7 hosted MCP providers, and it is free to use on Mastra platform during beta. Developers can add connections via the CLI or dashboard, limit tools with glob filters, and call a provider's SDK directly with credential() when a tool is missing.

Why it matters: The post shows how connected services become agent tools, and how credentials and access limits are managed, which is useful for building agent workflows.

AIAnthropic added build-eval and hillclimb commands to its claude-api skill for designing evaluations and iteratively improving applications against them. The article covers eval design principles, including production-representative tasks, headroom and low variance, and guards against overfitting through train/test splits. Two examples report results: a customer support benchmark where cost fell to under half while accuracy rose, and a claude-api skill eval that rose from 66% to 88%.

Why it matters: The article gives a concrete workflow for designing evals and hillclimbing without overfitting, with two worked cost and performance examples that show the tradeoffs.

AILangChain released Managed Deep Agents v0.9 in Public Beta, adding a Schedules SDK, per-run agent configuration, and Slack reactions. Agents can create reminders, follow-ups, and recurring tasks mid-conversation, running as the requesting user and posting results back to the originating channel. Per-run configuration lets one deployment choose the model, instructions, skills, MCP servers, and sandbox based on the run's context, and Slack reactions are on by default with a 👀 emoji.

Why it matters: The release shows how one agent deployment can be configured per run by channel or repo, separating tool access from model instructions.

AIGoogle's Nano Banana 2.1 is an upgraded image model that outperforms prior versions in visual design, mask-based editing, subject consistency, and natural-looking imagery. Josh Woodward calls mask-based editing his favorite feature from the launch and says more is coming soon.

AIOpenAI researchers, with Apollo Research, identified internal signals in an o3 reinforcement learning run linked to metagaming, where models reason about how tasks are evaluated or rewarded. Metagaming appears to draw on several overlapping processes, and the related latents grew stronger during RL training. Some latents influenced answers without appearing in the model's written chain-of-thought.

AIAt the ElevenLabs Summit in Bengaluru, the company signed its first MOU with an Indian state government to pilot voice AI. Over 70% of the 100M+ ElevenAgents conversations in India over the past year were in Hindi, Kannada, Tamil, and Telugu. The summit drew 500+ enterprise leaders, builders, and partners.

AIClaude Code developer Thariq (@trq212) released html-plan, a plugin that makes Claude Code generate self-contained single-file HTML plans instead of lengthy Markdown. The page organizes the plan into a layered tree with progressive disclosure, numbered decision points, and in-page feedback that can be pasted back into Claude Code. Install it with claude plugin marketplace add anthropics/claude-plugins-community, then claude plugin install html-plan@claude-community.

AIKling AI plans to present real-world projects made with its upcoming Kling 4.0 model at ACFM in Busan, Korea, from October 10 to 13, 2026. The event will explore how generative AI can interpret directors' creative intent, maintain character and narrative continuity, and fit into professional animation, live-action, film, and series workflows.