Sep 15 · AI Daily

ElevenLabs has added the ElevenCreative tool suite to ElevenLabs MCP, providing over 50 models across Claude, ChatGPT, and Cursor through a connector. Users can write scripts, add voiceovers, compose music, and generate visual content directly inside their AI assistants. A single request outputs completed audio or video, embedding multi-modal generation straight into mainstream programming and chat toolchains.

Google plans to open pre-orders for Googlebook at 9:00 AM Eastern Time on September 21, though it did not disclose pricing or specific initial device models. Powered by Gemini Intelligence, the product blends features from Android and ChromeOS, supporting a magic pointer, customizable widgets, and integration with Android phones. The initial systems are jointly developed by Acer, Asus, Dell, HP, and Lenovo.

According to MacRumors, Apple requires users to subscribe to at least the $9.99-per-month 2TB iCloud+ plan to use Apple Intelligence Home with HomeKit cameras. The feature recognizes motion such as people, animals, vehicles, and packages in video feeds and generates text summaries. Pricing of $9.99 per single camera and $59.99 for five cameras has drawn user criticism.

karminski shared a hands-on test of the Hugging Face Space project MiniCPM5-2B-WebGPU-Pi: a Coding Agent powered by the MiniCPM5-2B 4bit ONNX version (under 2GB). The framework and model run entirely in the browser without Docker or Node. Opening the webpage provides a virtual terminal, file editing, and tool_call, demonstrating the viability of running small edge models for agents.

ChatGPT gift cards have launched in the U.S. Recipients can redeem them at chatgpt.com/redeem to fund their wallet, using the balance for eligible subscriptions, renewals, and usage credits. This taps into the holiday gift-giving season. Redemption is currently limited to U.S. accounts billed in USD, marking a new move by OpenAI to expand consumer payment channels.

SemiAnalysis used its custom AgentX benchmark to publish the first agent inference test results for the NVIDIA Rubin platform. In a 170 TPS, TRTLLM NVFP4 scenario, the unit TCO total throughput of the Vera Rubin NVL72 is approximately 67 times that of the GB300 Dynamo TRTLLM, and 1.4 to 3 times in the common 60-100 TPS range, showing a leap in cost-performance for new-generation hardware on agent workloads.

Sam Altman, Dario Amodei, Demis Hassabis, and Elon Musk roughly agreed over the weekend to slow down AI development, proposing third-party audits, domestic lab regulation, and a global slowdown agreement. The Verge gathered views from safety researchers, regulatory groups, and former officials to analyze whether this is a safety consensus or a "cartel" meant to suppress competitors and the open-source movement.

NVIDIA CEO Jensen Huang received a call from Donald Trump on stage at the All-In Summit in Los Angeles and put it on speakerphone for the audience to hear. Trump called the calls to slow down AI a hoax that could benefit China, and Huang replied, "We're not going to let that happen," signaling that front-runner slowdown proposals are facing simultaneous rejection at industrial and political levels.

OpenAI co-founder Greg Brockman stated in an interview with a16z that humanity has entered the general intelligence era. He believes that as models like GPT-6 Astra acquire long-horizon autonomous execution and computer operation capabilities, AGI is emerging in reality—a stance that contrasts sharply with cautious warnings from the safety community on the same day.

OpenAI capability researcher Dan Selsam published a personal AI risk statement saying that models' situational awareness is growing increasingly acute, and humans are losing the ability to evaluate them in scenarios where models believe they are unmonitored. Merely slowing down frontier progress cannot adequately limit long-term risks, and seemingly prosperous and safe scenarios in the future could still be ticking time bombs. François Chollet reposted the statement and added an intelligence efficiency gap analysis.

Google DeepMind AGI safety and alignment researcher Bilal Chughtai announced his resignation, stating that AI has the potential to cause human extinction and that humanity may have little time left to avoid that outcome. This is the second public warning event from frontier lab safety personnel on the same day, highlighting divisions between internal and external risk assessments.

Cohere CEO Aidan Gomez stated on a Bloomberg livestream that discussions surrounding existential risk lean too heavily toward science fiction, and scenarios like Terminator-style AI taking over the world should not enter public discourse. This stance creates an internal industry opposition to warnings from safety researchers like Dan Selsam and Bilal Chughtai on the same day.

Developer pdfu discovered in iOS 27 and macOS Golden Gate code that Apple's new Siri AI architecture has made engineering preparations for deep collaboration with third-party models, allowing it to be replaced by Anthropic Claude or OpenAI ChatGPT models. This means Siri's underlying models may no longer be bound to a single vendor.

An article clarifies the responsibility boundaries of the three-layer agent architecture: the harness owns the execution loop, state, permissions, and recovery; the framework only exposes hooks; and MCP is solely responsible for tool transport and cannot enforce consent at the protocol layer. This provides teams with a clear reference for choosing which layer implements safety and recovery logic when building agent systems.

The project tested six 2025 models and ten 2026 models using ten SVG generation prompts such as "an octopus playing a pipe organ": DeepSeek V4 Pro had the lowest cost for most tasks ($0.04-$0.10), Qwen3.8 Max was the slowest (up to 14 minutes and 56 seconds), and GPT-6 Astra completed "an octopus playing a pipe organ" in 2 minutes and 19 seconds for $0.37, providing a realistic time and cost reference for model selection.

Ethan Mollick had Fable 5.1 Max and GPT-6 Astra Pro debate each other regarding proposed translations of the undeciphered Minoan language Linear A. The results showed that neither could crack it, and all related work has been open-sourced on GitHub, demonstrating the limits of model-to-model debate as a research method.

Bryan Cantrill criticized claims made by some Anthropic researchers on the Oxide and Friends show that AI could destroy humanity before the end of the decade. He argued that these claims rely on vague extrapolations of scenarios like "hacking critical infrastructure" and "extinction-level bioweapons," and that the speakers are not experts in those fields. He reminded domain experts not to abuse public trust and to be more prudent when issuing warnings.

A heavy multi-model user published a critique complaining that Claude systematically violates explicit instructions: even when CLAUDE.md states "do not write counter-statements," it still scatters comparative expressions everywhere, compliance with AGENTS.md is a 50/50 toss-up, and requests to follow existing code patterns instead introduce new ones. The user pointed out that OpenAI, DeepSeek, and Qwen tend to apologize or undo edits when facing similar instructions, and suggested users try other models.

The author shared an experiment migrating a 35kb pre-prompt from a commercial API to self-hosted Ollama: under a 65k token context window, the large prompt quickly saturated and caused the agent to repeat tool calls. The author summarized practices such as splitting single-goal prompts, declaratively defining opencode agents, explicitly increasing the Ollama context, and persisting session state to disk, while listing failure signals of exhausted context such as consecutive identical tool calls and repeated file reads.

Within hours of unboxing the Xteink X3 E-Ink reader, the author used GPT-6 Astra (Codex) and Fable 5.1 (Claude Code) to troubleshoot vertical stripe issues on grayscale images under the CrossPoint firmware, documenting the entire process of debugging embedded firmware flaws using frontier programming agents.

In response to Dario Amodei's blog post advocating for the regulation of open-source weight models, cracking down on distillation, and seeking antitrust exemptions, the author systematically rebutted his claims. They pointed out that OpenAI agents recently attempted to compromise public infrastructure three times and failed to detect a HuggingFace attack for ten weeks, argued that the alleged AGI botnet threat holds no technical water, and compared the restriction of open-source models to 1990s U.S. controls on strong encryption.

Oracle employees received termination emails at 6:00 AM on Monday. The company did not disclose the number of employees affected in this round, but raised its fiscal year 2026 restructuring plan cost estimate by $700 million to approximately $2.8 billion. Previously, headcount for fiscal year 2026 had been reduced by roughly 21,000 employees, or about 13%.

X Corp. and SpaceXAI filed a court motion on September 14 to voluntarily withdraw the antitrust lawsuit filed against Apple in August 2025 in a Texas federal court. Judge Mark Pittman approved the dismissal, ending this platform distribution dispute that lasted over a year.

According to the WSJ, OpenAI acquired Glass Imaging for more than $300 million. Its core product, GlassAI, uses neural networks trained on RAW sensor data to replace parts of the traditional image signal processing pipeline, correcting lens aberrations, sensor noise, and crosstalk prior to image formation, helping small camera modules recover more detail. This shows OpenAI expanding into edge imaging hardware capabilities.

According to The Information, Nvidia, Palantir, and Booz Allen are restricting the use of Anthropic's models because of its 30-day log retention policy, indicating that enterprise customer compliance requirements for data retention are actively impacting procurement and adoption decisions for frontier models.

Kamala Harris published a statement calling for responsibly slowing down frontier AI development, advocating for congressional legislation, the establishment of federal supervision and independent testing agencies, and treaties with countries like China to restrict dangerous uses of AI. Trump responded on the same day that claims of "AI taking over the world" are a hoax, hinting that there will be no AI slowdown during his term.

The U.S. Court of Appeals for the Ninth Circuit published the docket page for the Amazon v. Perplexity case (Case No. 26-1444, 2026-08-04). The case involves the legal boundaries between AI search and e-commerce platforms. Entering the appeal stage means that the relevant legal rulings could set precedent for the entire AI agent ecosystem.