Muse can handle complex tasks and support professionals who want to automate repetitive work. Digital marketers and content creators could use it to generate ideas, draft copy, manage social posts and free up time for strategy. The New York Times reports that users have been impressed by Muse’s capabilities, but the convenience comes with concerns about sharing emails, financial information and other sensitive data.

Meta says Muse prioritizes user control over data, yet questions remain about privacy practices, data handling and trust. Marketers and organizations using tools like Muse may need clearer data policies, stronger AI literacy and transparent guidelines to reassure users while preserving efficiency, originality and brand identity.

AI safety is the likeliest area for limited cooperation. Treasury secretary Scott Bessent has proposed an AI safety mechanism after talks with Chinese vice-premier He Lifeng, while China’s People’s Daily urged both sides to make AI a new frontier for cooperation. The White House state dinner is expected to include Jeff Bezos, Sundar Pichai, Sam Altman, Tim Cook, Elon Musk, Mark Zuckerberg and Jensen Huang.

Deep mistrust still limits the agenda. Beijing views US safety proposals and export controls as potential efforts to suppress Chinese development, while Washington remains concerned about China’s AI progress and semiconductor access. Trade tensions also persist as a temporary truce expires in November, with China seeking tariff relief and the US reluctant to give up leverage over rare earths and magnets.

Taiwan remains a flashpoint after Trump paused a $14bn package of arms sales following the May summit, a move criticized by Democrats. Senator Tammy Duckworth warned that earlier Chinese promises to buy 200 Boeing aircraft and $17bn in agricultural products had not materialized, and expressed concern that Trump could make further concessions to Xi.

Meta’s Muse assistant became a market catalyst after Wells Fargo raised Meta’s target share price from $640 to $796 and kept an “overweight” rating. Meta shares rose by nearly 12% on local time September 21, lifting Mark Zuckerberg’s estimated net worth by about 25 billion U.S. dollars in one day to 253.6 billion U.S. dollars. Amazon then blocked Muse from ordering on its platform, saying unauthorized AI agents violate its terms and may access user account and order data, while PayPal said it would support purchases and checkout through Muse among PayPal merchants worldwide.

The roundup also highlighted fresh AI model releases and infrastructure moves. OpenAI launched GPT-6 Sol and GPT-6 Luna with API prices 50% lower than the promotional pricing of GPT-5.6, Anthropic released Claude Opus 5.5 with typical-task operating costs about 40% lower than Opus 5, and Xiaomi open-sourced MiMo-V2.6. Alibaba CEO Eddie Wu said Alibaba Cloud aims for globally operated data centers to exceed 20GW by 2032.

At the Whitehead Institute for Biomedical Research, biochemist Anna Pertl used Co-Scientist to search for ways to target MYC, a cancer-linked protein that has long resisted drug-development efforts. After researchers corrected misunderstandings in the system’s framing, Co-Scientist reviewed more than 700 scientific papers and generated 108 possible strategies, ultimately proposing a counterintuitive approach: gluing molecular condensates together rather than dissolving them.

Similar platforms from Google, Anthropic, OpenAI, FutureHouse, Phylo, Huawei Technologies and Sakana AI are being tested across biology and other fields. Researchers have used these systems to identify drug combinations, explore tissue regeneration and reproduce theories about bacterial DNA transfer from public data.

The growing role of AI shifts more weight onto human judgment. Scientists still need to ask productive questions, catch errors, assess whether an idea is meaningful and verify claims experimentally. The same tools that promise faster discovery could also make training harder if students have fewer chances to build expertise through hands-on work.

Rome-based Exein raised $270 million at a $1.7 billion valuation to expand cybersecurity embedded directly into device firmware for robots, drones and edge devices. The company says Asia-Pacific now accounts for half of its revenue, with overall revenue growing 400% year-over-year, and it is training a foundation model for physical AI security slated for Q1 2027.

Madrid, London and New York-based Fever raised a $250 million primary equity round for its culture and live entertainment discovery and ticketing platform. Eindhoven semiconductor startup EUCLYD raised Over €200 million to develop AI chip systems intended to reduce data center energy needs, with former ASML CEO Peter Wennink joining as chairman of the board.

London-based Mantic raised a $25 million Seed round for AI systems that forecast real-world events for hedge funds and decision-making customers. Expanse, founded in 2025 and operating from London and San Francisco, raised $5.3 million for software that predicts AI workload compute requirements before jobs run.

A YouGov survey found that 85% of organizations currently use AI in some capacity, but only 18% use it extensively within operations. Larger firms are ahead, with 25% using AI extensively, compared with 11% of small businesses. A quarter of small businesses have yet to adopt AI at all, compared with 3% of large enterprises.

The UK government has already announced plans to train 10 million people in AI by 2030, including basic workplace uses such as drafting text, creating content, and completing administrative tasks. Researchers are calling for a Skills Tax Credit offering 20p for every £1 spent on accredited training, alongside a 30p ‘AI super-rate’ until 2030 and a dedicated AI Skills Intelligence Unit.

The initiative will roll out in about 100 countries, including Uganda, Angola, Sudan, Haiti and Mongolia. The companies said the specialized version is intended for regions where limited access to medical literature, specialist expertise and continuing medical education can affect patient care. Anthropic will provide back-end support, while OpenEvidence will tailor the system for regional healthcare infrastructure and local medical contexts.

OpenEvidence founder Daniel Nadler said US clinicians consulted the platform 42 million times in August alone. In 2026, several hundred million Americans will have been treated by a doctor who used OpenEvidence to help guide care, he said. Anthropic and OpenEvidence did not disclose financial terms for the collaboration.

The donation flow lets users ask about Trussell’s work and choose a monthly gift of £5, £10 or £15, with each option explaining how the money will be used before bank details are entered for authorisation. The trial positions Direct Debit and recurring pay-by-bank as potential foundations for agentic payments, where software agents support the payment journey while customers retain final approval.

GoCardless said consumer control will be central to wider adoption, citing research that found 64% of UK consumers are open to AI managing recurring payments if they keep control over spending limits or final approval. Trussell said the conversational format could make recurring giving feel more natural and reduce friction for donors.

Developers and researchers criticized the broader restrictions, citing refusals on prompts involving Alzheimer’s and immune response, bioinformatics, defensive security work, and some kernel development tied to frontier LLM capabilities. One commenter said ChatGPT 6 Pro answered a college student’s biology prompt without issue, while others argued that false positives make long-running coding and security workflows unreliable.

The discussion also raised concerns that strict safeguards could push users toward OpenAI, local models, open-source releases, or Chinese providers. Some commenters defended biology limits by pointing to gene-ordering services, biohacker culture, and the possibility that frontier AI could lower barriers to dangerous pathogen research.

Pricing through Xiaomi starts at $0.140 per million input tokens and $0.280 per million output tokens. Reused prompt prefixes are priced at $0.0028 per million cached input tokens, while the provider table also lists $0.140 input, $0.0028 cached input, and $0.280 output rates with a 1.0M input context.

The model is released under the MIT license, permits commercial use, and is described as open-weight, allowing users to download and run it on their own hardware. Official resources include Xiaomi provider documentation, an official playground, a technical report and model weights on Hugging Face, and an official launch post.