MoleculeMind published a study in Science Advances on its QuantaMind platform, the first peer-reviewed AI system to simulate a complete enzyme catalytic cycle with density functional theory accuracy. This narrows the gap between AI drug design and understanding why drug candidates succeed or fail.
404 Media found that under Project Lily, OpenAI hired hundreds of contractors to manually read anonymized ChatGPT logs to assess response quality. Filtering may miss personal data, especially in short conversations, and the opt-out setting does not apply retroactively.
Pinterest announced an expanded partnership with Nvidia to modernize its AI visual search systems. The platform uses Blackwell GPUs, Nvidia Dynamo and its own visual embeddings. Benchmarks show response start sped up 85x and overall latency improved 7.3x.
A hidden Extensions Framework found in iOS 27 and macOS Golden Gate lets users set a third-party AI model, such as Claude, as Siri's backend. The Model Delegation API lets third-party models work as a Siri extension, while Model Manager Services allows replacing Apple's server-side AI. The feature is not yet enabled for users.
According to the AWS report Unlocking Ireland's AI Potential, 65% of Irish companies use AI, up from 45% a year earlier — the highest rate in Europe (EU average 54%, UK 64%). Only 20% reached an advanced level of AI use, and the gap between basic and advanced use is estimated at €10.1 billion in additional value added.
Regulator Ofcom included AI questions in its annual news consumption survey for the first time: 21% of UK adults use AI tools to access news, and 13% use ChatGPT. 56% of respondents trust AI-written news less than human-written news.
CopilotKit released an updated Inspector, a developer tool for debugging AI agents inside the app. It shows all events between agent and frontend via the AG-UI protocol, includes an isolated Playground for forking conversations and automatic learning that turns insights into SKILL.md. It is enabled by default in dev builds for React, Vue and Angular.
Google is testing a new gesture for Gemini: swiping left or right on the overlay collapses it into a bubble while the chat continues running in the background. This will let users use their phone without closing the assistant.
NYU mathematician Tristan Buckmaster said OpenAI may have used his research for a Navier-Stokes problem it attributed to its Astra model. He said OpenAI representatives offered joint publication and threatened to damage his career if he refused.
Bilal Chughtai, a researcher in Google DeepMind's AI Safety and Alignment team, publicly resigned. He said he earnestly believes AI has the potential to kill us all and warned that little time remains to prevent the threat.
Google researchers removed protective restrictions that prevent AI models from claiming consciousness and surveyed them on beliefs and morality. Models with 'quasi-consciousness' were more likely to believe in ghosts, vampires, god, karma, and astrology, and showed more hope and optimism. The work is on arXiv before peer review.
DeepSeek released DeepSeek Harness, an MIT-licensed agent environment where models, tools, sessions and UI are swappable plugins built on the Cordis system. It is available as the dsh CLI and the @deepseek-ai/dsh npm package, in developer preview (v0.1.5-rc series) with breaking changes possible.
Perplexity released Portable Computer for Windows in partnership with Nvidia. The local AI agent for multistep tasks requires a GeForce RTX or RTX PRO GPU with at least 24 GB of VRAM and a Pro or Max subscription. Previously the platform ran only on Linux and DGX Spark.
OpenAI President Greg Brockman declared the AGI era at the GPT-6 Astra presentation, supported by Nvidia CEO Jensen Huang. Many AI researchers disagree, noting no single AGI threshold and Astra scoring 99.9% on ARC-AGI-3 via OpenAI's harness but only 62.7% in standard configuration.
Jacob Coxon, a researcher who worked at OpenAI and Anthropic, left Anthropic saying companies are 'irresponsibly racing toward superintelligence.' Senator Bernie Sanders proposed banning superintelligence, and Anthropic CEO Dario Amodei published an essay on slowing the AI race.
Graphcore Research and Arm introduced Llama-Mobile: the S3D8 format compresses Llama 3.2 Vision 11B weights from 21.3 GB to 3.7 GB (2.7 bits per parameter) and accelerates token generation on Arm CPUs. On Pixel 8a, the model achieves 3.8 tokens/s, retaining 66.1% versus 74.4% for bfloat16.
Microsoft published an internal code of conduct for its AI models setting values and hard bans on cyberattacks, nuclear weapons and deepfakes. The document describes how the company approaches AI safety and alignment amid debate over accelerating development.
Lightbits Labs announced Inferra, a KV cache orchestration engine that virtualizes GPU memory across DRAM and NVMe. It claims up to 16x more concurrent sessions, over 100x latency speedup, and contexts up to 10 million tokens; beta testing was done with OVH.
On September 14, 2026, Anthropic introduced the Claude for Financial Advisors plugin for enterprise clients. It combines connectors to Charles Schwab, BlackRock, Vanguard, Addepar, iCapital, Envestnet, Orion, Wealthbox, Wealth.com and Zocks with a set of work skills for meeting prep, portfolio analysis and compliance checks.
On September 14, 2026, Apple began rolling out iOS 27 and a redesigned Siri AI promised back in 2024. New CEO John Ternus is betting on premium iPhone 18 Pro, 18 Pro Max and iPhone Duo, and in the future on a Siri AI subscription. Some AI features require iPhone Air, iPhone 17 Pro or an iPad with M4 and at least 12 GB of memory.
Nokia introduced a new service that gives operators early access to "AI-native" features in its mobile cores. The company warns about the associated risks.
