AUO has completed three asset monetization deals this year, bringing in about NT$34.7 billion, according to a report cited in the source material. The transactions include the sale of a Gen 5 fab in Huaya Technology Park to Quanta for NT$19.7 billion, a Gen 5 color filter plant in Kaohsiung’s Luzhu Science Park to ASE for NT$6.3 billion, and the sale of its Singapore AFPD facility to Western Digital for about S$350 million, or roughly NT$8.753 billion. The latest market talk says Taiwan Semiconductor Manufacturing Co. has visited AUO’s L5C and L7A plants in the Central Taiwan Science Park. The discussions are said to cover not only the two facilities but also possible technical cooperation. The report says the two CTSP sites already have cleanroom, power, ultrapure water, and exhaust treatment infrastructure in place, with L7A next to TSMC’s existing CTSP operations and L5C adjacent to L7A. The report also links AUO’s process strengths in large-area substrate thin-film manufacturing, TFT Array production, precision alignment, and uniformity control to possible FOPLP development. Even if talks move ahead smoothly, outside estimates cited in the report say a formal agreement and closing would likely not come before the second half of next year.

After GPT-6 Astra went live, OpenAI’s developer community began sharing a striking run of projects built with the model across 3D, game development, and interactive design. ABMedia, citing a roundup from OpenAI Developers, said Astra is being used for more than code completion: developers are applying it to assemble editable digital environments inside tools such as Blender, Three.js, Unity, and Unreal Engine. The examples listed in the report span a 3D human anatomy site broken into 2,234 modeled parts, a Manhattan environment built street by street in Unreal Engine over a week, and a walkable "Van Gogh Town" created from six paintings. Other projects include a steam train reconstruction in Blender that produced 3,295 editable objects, a Backrooms scene with VHS effects and audio, browser-based games, and a LEGO design workflow that exports .ldr files using official parts. Across those cases, the article frames Astra as a coding agent that can work with geometry, space, interaction logic, and visual output inside structured 3D workflows rather than merely generating flat images or finished video.

Jacob Coxon, a 27-year-old AI researcher who worked on large-model pretraining at both OpenAI and Anthropic, said he resigned because he believes leading labs are racing toward self-improving superintelligence while effectively betting human lives on the outcome. His post on X drew more than 100 million views, according to the source article, and reignited debate over existential AI risk inside top model companies. The PANews report says Coxon was listed as a contributor to GPT-4o and a core contributor to GPT-4.5 before moving to Anthropic this year. After his post, Anthropic’s alignment lead publicly backed him and said the probability of such a failure scenario happening within 10 years was greater than 10%. Another Anthropic safety researcher also supported him in a personal capacity, while former scalable oversight lead Joe Benton said Coxon’s description of the industry was accurate. The article places Coxon’s warning alongside Tim Urban’s 2015 essay “The AI Revolution,” revisiting ideas such as exponential progress, intelligence gaps, recursive self-improvement, and the possibility that advanced AI could push humanity toward either extinction or what Nick Bostrom called “species immortality.”

A male fruit fly connectome published in Cell on Sept. 3 was turned into a Minecraft-based simulation just two days later, according to a MarsBit report citing posts by Georgia Tech graduate student Evan Smith. The dataset, called MaleCNS v1.0, was built over a decade by HHMI Janelia, the University of Cambridge, and Google Research, and includes 166,700 neurons and 125 million synaptic connections. Smith said the project is still in development, was built with help from GPT-6 Astra, and that the code and mod will be released soon. The report says the simulation uses neural activity rather than a standard game script to drive a virtual fly inside a glass-walled Minecraft room. The article argues that three factors made the rapid build possible: a fully open connectome, AI-assisted coding that reduced engineering overhead, and Minecraft’s existing modding and physics environment. It also stresses that the project demonstrates behavior-level simulation, not consciousness, while suggesting the lowering cost of whole-brain modeling could let individuals test ideas that once required larger lab setups, dedicated teams, and specialized engines.

NestWorks, a desktop CNC brand incubated within Elephant Robotics, has completed an angel and angel-plus financing round worth several hundred million yuan, according to the source article. Backers include Sequoia China, Lanchi Ventures, CDH Capital, Vertex Ventures and Unity Ventures, while Gaohu Capital acted as the exclusive financial adviser for subsequent fundraising. The company was formally established in November 2025, and its first product, the C500, raised more than $13 million on Kickstarter. The report says NestWorks built the product around its own in-house prototyping needs. The company wanted a machine that could sit in an office, cut aluminum alloy, stainless steel, brass and engineering plastics, and still be simple enough for engineers without formal CNC training to use. The C500 initially ships with an 800W spindle, up to 18,000 RPM, MQL lubrication, automatic tool changing and an intelligent tool management system. NestWorks argues the bigger opportunity is not only hardware, but also software and manufacturing knowledge. Its system connects model import, CAM generation, toolpath planning, simulation and machine execution in one workflow, with software and AI algorithms handling decisions that traditionally depended on operator experience. The company says desktop CNC will likely spread first through engineers, designers, research institutions and small-batch businesses before moving outward into broader user groups.

Goldman Sachs used its Aug. 25 review of second-quarter earnings in the U.S. internet sector to make a simple point: investors are still willing to fund AI spending, but they are asking much harder questions about timing, payback and margin pressure. The bank said the quarter confirmed three themes across the sector — rising AI investment, resilient digital consumption and a tougher balancing act between growth spending and profit discipline. Internet stocks gained 9% during the period, trailing the S&P 500, which rose 11%. The report also previews Goldman Sachs’ Communacopia technology conference in San Francisco, scheduled for Sept. 8 to Sept. 11. Forty public and private companies are set to appear, with the OpenAI CFO opening the event and executives from Google Cloud, SpaceX and Uber also on the agenda. Across digital advertising, e-commerce, cloud computing, travel, ride-hailing, streaming, gaming, dating and online betting, Goldman said the central debate has shifted from whether companies should invest in AI to how clearly management teams can explain the return profile. In cloud, Google Cloud and Amazon Web Services reported massive backlog growth and ongoing capacity constraints. In consumer-facing internet segments, spending held up, though higher-income and lower-income users are behaving differently.

Tencent Hunyuan has released and open-sourced Hy4 preview, a next-generation LLM with 770B total parameters, 49B active parameters and more than 1M context. According to ChainCatcher, the model sits in the top tier of open-source models for real productivity work, including coding, office tasks and scientific research. Hy4 preview expands model size, context length and data scale, and was developed with high-quality data from Tencent experts in software engineering, gaming, finance and security, plus deep integration with WorkBuddy. It can build complex frontend projects, generate playable game prototypes, handle financial audits across multiple documents, and speed up molecular dynamics simulations. The model is available in Tencent products such as WorkBuddy, CodeBuddy, Yuanbao and ima, with API access through Tencent Cloud Tokenhub and OpenRouter. WorkBuddy and CodeBuddy are running a two-week free trial. Tencent Hunyuan has released a major version roughly every two months since rebuilding its infrastructure, using a preview-first cadence before formal releases.

Tencent has released the preview version of its Hunyuan Hy4 large language model, built on a mixture-of-experts (MoE) architecture with 770B total parameters, 49B activated at a time, and a 1-million-token context window. The new generation is more than twice the size of Hy3 — which had 295B total parameters, 21B active parameters and 256K context — and its long-context capacity has grown roughly four-fold. Rather than lean only on general benchmarks, Tencent is steering Hy4 toward concrete tasks in coding, office work, gaming and scientific research. The model can build a Three.js 3D website from scratch, craft a playable game demo in Unity, or work through 72 financial documents at once to catch duplicate reimbursements, over-limit expenses and budget anomalies. Hy4 was trained and iterated along with Tencent products including WorkBuddy and CodeBuddy, and it has begun contributing to its own development — helping optimize training methods, data strategy, evaluation systems and underlying operators before running experiments and adjusting its approach based on the results. The preview is already integrated into WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub and OpenRouter, with Tencent having flagged earlier in its earnings report that Hy4 would surpass Hy3 and that real product feedback would be a key source for training.