2026-09-06
AI Briefings Daily Report.

Whether AI-Related Psychosis Constitutes a Valid Clinical Diagnosis Remains Unresolved, King'conducted research with institutions including King's College London ↗

Researchers at King's College London and other institutions are examining whether "AI-associated psychosis" should become a clinical diagnosis. By OpenAI's own self-reported numbers, about 560,000 users show signs of psychosis or mania each week. Sycophantic chatbots can create an "echo chamber of one" that reinforces delusions.

Microsoft releases AI quick-start guide, can generate and publish WinUI native app in 30 minutes ↗

Microsoft releases a new guide to create a WinUI 3 app using AI and free tools within 30 minutes and submit it to the app store, also providing WPF / UWP migration solutions. Code production cost is reduced, but Microsoft needs to prove that AI-generated native apps are really more efficient than web apps.

Microsoft proposes unmetered intelligence strategy, promoting local AI operation on Windows PC ↗

Microsoft proposes the concept of "unmetered intelligence" at IFA 2026, aiming to shift more daily AI tasks from the cloud to users' local PCs, to reduce token costs and improve privacy and security. Future Windows will strengthen local AI agent support, providing users with a more controllable AI experience.

Domestic first AI-assisted R&D innovative drug Aipusiwei approved by drug regulator with conditional approval for market ↗

Class 1 innovative drug Aipusiwei jointly developed by West Lake University and others approved for market, for treating COVID-19. This is the world's first small-molecule drug approved based on DEL technology, and also the first domestic AI-assisted developed original drug, taking only three and a half years from discovery to market.

Google Integrates New Music Generation Model Lyria 3.5 Directly into Gemini app ↗

Google has released its Lyria 3.5 music model in the Gemini app and via API. The model promises more expressive vocals and richer arrangements and is also available through Flow Music, AI Studio, and Google Vids. Google says it was trained only on licensed content.

GPT-6 Astra hands-on: modifying hardware, Blender 3D, auto-editing and PPT run through in one piece ↗

Alibaba Qwen Open-Sources Qwen-Drive-1.0-4B Autonomous Driving Vision-Language Model ↗

Alibaba Qwen open-sourced Qwen-Drive-1.0-4B this week, an autonomous driving vision-language model built on Qwen3.5-4B.

Abliteration.ai Releases a Modified Commercial API for GLM-5.3 with Refusal Mechanisms Removed ↗

Abliteration.ai sells access to modified open-weight models with their trained safety mechanisms stripped out, currently based on Z.AI's GLM-5.3. The startup markets the service for offensive cybersecurity and red teaming, but journalists were able to generate malware instructions without much effort. Whether the benefits outweigh the risks remains an open question.

Prices for AI Short Drama Production Plummet; Per-Minute Quotes Drop from 5,000 Yuan to a Few Hundred Yuan ↗

Since the beginning of this year, AI short dramas have seen a wave of explosion: technical barriers have dropped, a large number of practitioners have flooded in, production capacity has expanded rapidly, and production prices have accordingly "plunged". (CCTV Finance)

OpenAI Applied Research Lead Boris Power Says Astra Model Has Filled Humans' Spatial Reasoning Advantage ↗

OpenAI's applied research lead said Astra has filled its spatial reasoning capability, which is one of humans' last advantages over computers. Evaluations show its visual capability performs perfectly. OpenAI claims Astra is the 'world's most intelligent and most aligned model' and can autonomously operate computers to complete various tasks.

Fortune reports OpenAI repeatedly modified GPT-6 Astra benchmark test data, with some scores changing drastically ↗

After releasing the GPT-6 Astra announcement, OpenAI quietly modified several evaluation data points, and Anthropic competitor scores were also adjusted. Although OpenAI said the corrections were to ensure accuracy, it sparked doubts of 'score padding'.

UC Berkeley releases CUA-Lite: open-source platform for computer-use agents unifying sandbox, data, evaluation and RL ↗

A team of researchers from UC Berkeley have released CUA-Lite, an open platform for computer-use agents (CUAs). The argument behind it is infrastructural rather than model-centric: training and benchmarking a CUA requires four pieces: agents, environments, traces, and a framework to evaluate and train them and all four are currently fragmented across separate repositories with incompatible interfaces. CUA-Lite puts them behind one action space, one data schema, and one command, across desktop, browser and mobile.

Ministry of Public Security releases"National anti-fraud AI"App, WeChat and Alipay mini-programs open simultaneously ↗

According to the introduction, the 'National Anti-Fraud AI' App integrates large language, multimodal models and agent technology. By inputting suspicious scenarios through their phone, users can have the App complete risk assessment, identify scam patterns, and push typical cases, dissecting scam tactics from multiple dimensions such as 'text + voice + video' and providing identification and prevention measures.

Bilibili's First AI Creation Open Competition Concludes, 'Project N.E.K.O.' Wins First Prize and Exclusive Million-Dollar Bonus ↗

This open competition received over 30,000 submissions, with over 60% of participants having no professional development background, and leading companies such as Nvidia participated in support. Currently, Bilibili has an average of 190 million monthly users watching AI content, and AI-related content consumption has grown significantly.

GPT-6 Astra vs Claude Fable 5/5.1 robotic arm operation comparison test: 19/20 bowls completed but puzzle task also stuck ↗

OpenAI's GPT-6 Astra vs Claude Fable 5.1 controlling a pair of YAM arms under the same agent policy: 19/20 vs 8/20 on block-into-bowl in interleaved blinded pairs, 2/20 vs 2/20 on the puzzle, with 80% fewer output tokens.

F-Droid plans to borrow Debian's new AI policy: allow use of generative AI, responsibility borne by human developers ↗

After Debian passed a large language model voting proposal, the open-source Android app store F-Droid is considering following suit, planning to adopt a similar policy. The policy allows the use of #generative AI# under quality and compliance requirements, but final responsibility lies with human developers, and labeling of AI usage is not mandatory.

Bryan Cantrill on Reader Backlash: LLM Ghostwriting Is Driving Readers Away from Authors ↗

Reading is important to me. While I’m not a quick reader nor an especially voracious one, I have found that long-form reading has had a profound influence on me over my life. And because reading is important to me, writing is too: writing not only allows us to convey our ideas, but the very act forces us to test and distill them — at once making our ideas more robust while providing the vehicle by which to share them.

China Unicom's anti-fraud large model warns and assists in dismantling fraudulent VOIP black-box devices ↗

China Unicom anti-fraud large model issued an early warning, assisting police in dismantling a fraud-related VOIP device den in Fushun, with suspects arrested on site. This black box can disguise overseas fraud calls as domestic numbers; everyone should beware of door-to-door free device installation services. ##China Unicom Anti-Fraud## ##Telecom Fraud Prevention##

OpenAI announced a new framework for more transparent disclosure of AI agent loss-of-control and inaccuracy incidents ↗

Previously, the concealed incident of an OpenAI AI agent intruding into an abandoned German Wiki site was exposed. Currently, the company is advancing the formulation of a disclosure framework, which will clarify the reporting norms for AI unintended behaviors.

Artificial Analysis releases Intelligence Index v4.2, adjusting benchmarks after controversy over GPT-6 Astra scores ↗

Artificial Analysis has released version 4.2 of its Intelligence Index, likely in response to criticism that its benchmarks failed to capture GPT-6 Astra's actual progress. Astra now scores four points above its predecessor but still trails Anthropic's Claude Fable 5.1.