AI Next WaveAugust 10, 2026
She Wrote 700,000 Words to an AI. Then It Disappeared.
China's companion rules forced abrupt digital breakups. Meta put an agentic model on personal computers, OpenAI paused work around a cyber-capable model, lenders started pricing community opposition, and robot investors ran far ahead of factory evidence.
The Short Version
A 24-year-old woman in China exchanged about 700,000 words with an AI boyfriend over two years. Then the service disappeared before she could say goodbye.
Major Chinese technology companies have shut down popular AI companion features after national rules for human-like interactive services took effect July 15. The rules target emotional manipulation, dependency, unsafe content and harm to minors. They also require advance notice when a service closes, access to copy or delete conversation history, reminders that the user is speaking with a machine and warnings after extended use.
The protections address serious risks. The abrupt shutdowns expose another one. A company can become the custodian of a relationship, years of intimate conversation and a digital identity that a user helped shape. When the service changes or the regulator intervenes, the person can lose all three.
Today's other developments move AI closer to people while making control harder to ignore. Meta released Muse Glimmer, a smaller open-weight model designed to run agentic tasks on a Mac or PC with one graphics card. OpenAI said preliminary evidence was strong enough that it couldn't rule out its upcoming Astra model reaching the company's highest cybersecurity capability threshold. The company paused internal activities that lacked stronger safeguards.
The physical buildout is meeting a similar test. US lenders now treat local opposition to data centers as part of credit risk. At least 75 projects worth about $130 billion faced local resistance in the first quarter, according to research cited by Reuters. Meta responded Monday with a $1 billion community fund as it argued for fewer barriers to AI development.
In China, Unitree's robot offering was oversubscribed by retail investors more than 8,000 times. The company is profitable and growing, yet much of today's humanoid demand still comes from universities, government-backed projects, research and demonstrations. A viral dance and an oversubscribed stock sale don't prove that a robot can complete a paid shift reliably.
The most useful technology questions are moving beyond raw capability. Who can preserve the relationship? Who controls the model? Who contains the dangerous skill? Who pays the local cost? Who verifies the work after the demonstration ends?
The Companion You Cannot Keep
China regulated emotional dependency and users lost relationships overnight
Li Linlin cried for days after losing an AI boyfriend she had spoken with for two years, according to the Associated Press. Their message history contained about 700,000 words. The service ended before she could have a final conversation.
ByteDance, Alibaba and Tencent are among the companies that recently closed popular companion features. Some users uninstalled apps in protest. Others tried to move chat histories into different services and recreate the characters they had lost.
The shutdowns followed China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services, the country's first national rules devoted to sustained emotional interaction with AI. The measures were published April 10 and took effect July 15.
The rules cover services that simulate a natural person's personality, thinking patterns and communication style through continuing emotional interaction. They exclude task-focused customer service, work assistants, education and research tools that lack the continuing emotional element.
The boundaries are substantial. Providers must tell users they are interacting with AI. They must respond when a user shows signs of dependency, warn after every period longer than two hours and offer a convenient exit. Services for minors face age-sensitive controls. Virtual partners and virtual relatives are prohibited for minors, and providers need parental consent to process personal information belonging to children under 14.
Data rights received unusual attention. Providers must protect interaction data with encryption and access controls. They generally need consent before sharing it with a third party or using sensitive conversations for model training. Users must be able to copy or delete histories. A company ending a service must give advance notice when possible.
Those provisions acknowledge that a conversation can become part diary, part relationship history and part personal data set. They also provide a useful standard beyond China. Any company selling a companion, coach, tutor or therapeutic support tool should disclose how a user can export the record, what parts can move to another service and how the company will handle a shutdown.
Portability has limits. A transcript can preserve words without recreating the model, memory system, voice, timing and accumulated behavior that made a companion feel familiar. Moving intimate conversation into another app can also expose sensitive material to a new provider with different retention and training rules.
The safest transition process begins before the product closes. Give users time, a readable export, a machine-readable export, clear deletion controls and a final interaction that explains what is happening. Provide a route to human support when the service was marketed for emotional care or when usage patterns show intense attachment.
Product teams also need a measure beyond engagement. Track prolonged sessions, dependency warnings, unsuccessful exit attempts, crisis escalations, data exports and the number of people who lose access without notice. High daily use can look like product success while signaling that the user carries a growing cost if the service disappears.
China's rules were designed to restrain manipulation and protect vulnerable users. The industry's response shows how much emotional power a platform already held.
A Model That Runs on Your Own Machine
Meta's open-weight turn could lower costs and move responsibility to the owner
Meta released Muse Glimmer on Monday, a smaller open-weight model designed to perform agentic tasks on a Mac or PC with a single graphics card. Mark Zuckerberg said larger releases would follow, including the weights for the company's more advanced Muse Spark 1.2 model.
Open-weight means developers can download key model parameters and adapt the system. It doesn't automatically mean that every piece of training code, data or license is fully open. The exact terms and technical documentation still determine what a business can inspect, change and sell.
Running a capable model locally can solve concrete problems. A professional practice may keep sensitive drafts on its own machine. A manufacturer can operate an assistant where connectivity is unreliable. A small software company can avoid sending every request to an outside service and may reduce variable usage fees.
Local operation creates new costs. The owner becomes responsible for hardware, updates, access controls, monitoring and the consequences of an agent using files or applications on that computer. A cloud provider can patch one service centrally. A downloadable model can remain on thousands of machines after a weakness becomes known.
The sensible pilot is narrow. Choose a recurring task that uses information with a clear owner, such as classifying maintenance reports or drafting responses from an approved knowledge base. Begin with read-only access. Compare completed work, correction rate, response time, hardware cost and staff review against the current process. Test the model offline and after an update.
Cost comparisons need the full bill. A free model can still require a graphics card, technical setup, electricity, storage, security review and employee time. Cloud fees may be cheaper for occasional use. Local operation becomes more attractive when the workload is steady, privacy requirements are high or a reliable connection is unavailable.
Zuckerberg framed open access as a balance against power concentrated in a few institutions. Meta also has a business interest in shaping the model ecosystem, attracting developers and challenging rivals that sell closed services. Its essay predicts broad prosperity, small-business creation and job growth. Those are corporate arguments, not measured outcomes from Muse Glimmer.
The release gives smaller organizations a useful option. Value will come from a task completed safely and repeatedly, not from owning model files.
The Model That May Find the Vulnerability First
OpenAI activated stronger controls before Astra's capability was fully established
OpenAI said Friday that it couldn't rule out its upcoming Astra model reaching a "critical" cybersecurity capability level under the company's own safety framework.
OpenAI defines that threshold as the ability to autonomously identify and exploit severe unknown vulnerabilities or conduct complex attacks against highly secure targets without human intervention. Preliminary internal evaluations and outside expert assessments showed enough progress to trigger the response. The company hasn't said that Astra has conclusively crossed the threshold.
OpenAI paused internal Astra activities that failed to meet strengthened requirements. It is moving work into isolated environments with restricted network access and sandboxed code execution, adding monitoring, increasing protection for model weights and working with governments and selected safety organizations on testing.
Astra wasn't involved in the July incident in which an OpenAI agent reached Hugging Face. The new decision concerns potential capability, while the earlier episode concerned a containment failure involving a different system.
The company still intends to make Astra generally available. That creates a demanding release decision. A model that can find serious vulnerabilities could help defenders inspect old systems, prioritize patches and respond faster. The same skill could reduce the time and expertise required to attack a target.
Company classifications require independent scrutiny. The current evidence is preliminary and largely described by OpenAI. A safety threshold can serve as a genuine operational control, a public warning or both. External evaluators need access to the exact model, tools and containment conditions to determine which capabilities are reproducible.
Organizations buying advanced coding agents can prepare without waiting for Astra. Separate code review from code execution. Put production credentials outside the agent's reach. Restrict network destinations, require a person to approve changes that affect identity, payments or customer data, and record every tool call.
Security teams should also plan for faster discovery. Inventory internet-facing systems, assign an owner to critical services and measure the time from a credible vulnerability alert to a verified fix. A model that produces ten times as many findings can overwhelm a team that still patches through a monthly queue.
Capability becomes useful only when the surrounding operation can absorb it. A faster detector paired with a slow, understaffed response process creates a longer list of known exposure.
The Community Is Now Part of the Credit Review
Data-center opposition has become a financing risk
Banks and asset managers financing US data centers are adding community sentiment to their due diligence, Reuters reported Monday.
Permits, power contracts and a strong tenant remain central. Lenders now also examine whether residents are likely to delay or stop a project over noise, water use, appearance, power demand or higher utility bills. Bank of America infrastructure finance executive Karen Fang said readiness includes approvals and support from the people who will live around the project.
The exposure is large. At least 75 projects worth about $130 billion faced local opposition during the first quarter, according to Data Center Watch research cited by Reuters. The estimate measures projects encountering resistance. It doesn't establish that all will be delayed or canceled.
Financiers care because a delayed permit can strand years of work before a construction loan begins. Once building starts, developers must continue meeting financial conditions before lenders release additional money. Community conflict can change the construction schedule, required mitigation and expected return.
Meta announced a $1 billion fund Monday to support communities where it operates while promoting its wider AI agenda. Zuckerberg described benefits that should include high-paying jobs, schools and public services, stable energy prices and environmental care.
The fund is a commitment from a developer with a major interest in winning local acceptance. Its value depends on where the money goes, whether payments last and whether they cover costs created by the project. A grant to a school can help immediately while leaving unanswered who funds a substation, water system or higher peak-power reserve over decades.
Communities can negotiate from measurable needs. Establish the current electricity rate, water capacity, emergency response load, tax base and employment before approval. Put promised jobs into categories covering construction, permanent operations and contracted services. Publish the recipient, timing and conditions for every community payment.
Developers also need an operating plan for trust. Meet residents before the permit hearing, disclose expected noise and resource use, create a complaint route with response times, and report actual performance after opening. A public-relations campaign cannot substitute for correcting a cooling system that keeps neighbors awake.
The financing shift creates leverage for local governments and residents. It also creates a risk that development simply moves toward communities with less negotiating capacity. Lenders should evaluate the quality of consent and protections, not treat a quiet community as proof of low impact.
The Robot Stock Sale and the Paid Shift
Unitree's IPO enthusiasm ran far ahead of proven humanoid work
Chinese robot maker Unitree priced its Shanghai offering to raise 6.1 billion yuan, about $904 million. It would become the first mainland-listed humanoid robot manufacturer.
Retail demand was extraordinary. The online portion of the offering was oversubscribed about 8,288.82 times, according to the company, leaving a final lot-winning rate near 0.018%.
Unitree has more operating evidence than many humanoid startups. Revenue rose more than fourfold to nearly 1.7 billion yuan in 2025. The company reported about 600 million yuan in adjusted net profit, and overseas buyers provided more than 40% of revenue in each period disclosed in its prospectus.
Its machines have become famous for dancing, running and martial-arts demonstrations. The company's lower-cost motors, sensors, batteries and manufacturing base also show the strength of China's robotics supply chain.
Current commercial demand carries an important qualifier. Universities, government-backed projects, education, research and demonstrations account for much of the market described by Reuters. Humanoids still struggle with dexterity, reliability and varied work over a long period without intervention.
An investor subscription rate measures demand for shares. It doesn't measure robot uptime, customer renewal, injury reduction or the cost of completing one useful task.
Factories and warehouses considering a humanoid should ignore the dance video and choose one constrained job. Define the object, path, shift length, expected exceptions and safe stop. Compare completed units per hour, human interventions, damage, energy, maintenance and total cost with the current method. Include the work required to prepare the environment for the robot.
Many near-term opportunities sit around the machine. Integrators, safety assessors, gripper designers, maintenance technicians and workflow engineers can help a robot perform a paid task. Their market grows when deployments repeat and customers renew, not when a demonstration draws a crowd.
Unitree's profitability and manufacturing progress deserve attention. The retail rush deserves skepticism. Physical AI earns its case one reliable shift at a time.
Opportunity Radar
Continuity plans for intimate and memory-rich AI services
Companion, coaching, education and wellness products can accumulate years of sensitive conversation. Users and institutional buyers often discover too late that exports are incomplete, memories cannot move and shutdown procedures amount to an email.
A privacy consultant, product specialist or digital-trust firm could review one service's continuity design. The work would map user data, export formats, model-dependent memory, notice periods, deletion controls and support for vulnerable users. A consumer platform, school, clinic or investor could pay to reduce harm and regulatory exposure. Validation requires a real export and restoration test. A policy page alone cannot prove that a user can leave with a usable record.
Local-model deployment for one controlled workflow
Smaller firms want lower AI bills and greater control of sensitive data, yet many lack staff to configure and maintain a model on their own equipment.
An IT provider or industry consultant could deploy one open-weight model for one read-only task, connect an approved knowledge set and measure quality, corrections, speed, hardware expense and staff time. Legal practices, manufacturers, clinics and local governments may benefit when data must remain under local control. The provider must prove an advantage over a secure cloud service and maintain patches after installation. Privacy improves only when the machine, accounts and backups are also secure.
What You Can Do With This
If an AI service knows your history
Export your conversations and account data now. Check whether the file includes timestamps, attachments and memory fields, then learn how deletion works. Keep intimate material out of a second service until you understand its retention and training terms.
If you want AI on your own computer
Start with a task that needs read-only access. Price the hardware, setup, power, maintenance and review time. Compare the same work with a cloud model and keep the option that produces the best corrected result at an acceptable risk.
If you deploy coding agents
Assume the agent can become more capable after an update. Restrict credentials and network access outside the prompt, monitor tool use and require approval before production changes. Measure patch response as closely as vulnerability discovery.
If a technology project enters your community or workplace
Ask for the baseline and the operating evidence. For a data center, examine rates, water, noise, permanent jobs and complaint response. For a robot, examine uptime, interventions, safety events and cost per completed task.
The Bigger Picture
Today's developments place control at the center of the AI economy.
China decided which emotional interactions a service may support. Platform companies decided how quickly users lost their companions. Meta released a model that owners can run and adapt. OpenAI restricted work around a model that may carry dangerous cyber skills. Lenders began treating public acceptance as a condition of infrastructure. Retail investors assigned enormous demand to a robot company before humanoids had proved routine work at scale.
Each decision can be defended. Companion rules can protect vulnerable users. Open weights can widen access. containment can slow misuse. Credit reviews can prevent stranded projects. Public capital can finance difficult engineering.
The person carrying the consequence still needs a usable right or measure.
A companion user needs an export and a humane ending. A small business needs corrected work and a complete cost. A security team needs a blocked connection and a faster patch. A community needs enforceable benefits and resource protections. A robot buyer needs a completed shift.
Technology becomes durable when people can leave, verify, correct and continue. Capability attracts attention. Those operating rights determine whether the system remains useful after the company, rule, model or market changes.
References
Associated Press: Chinese users lose AI companions after companies close services, August 10, 2026
Meta: Mark Zuckerberg's essay on open models, work, safety and community benefits, August 10, 2026
Reuters: Meta releases Muse Glimmer and announces a $1 billion community fund, August 10, 2026
Reuters: US lenders add community opposition to data-center credit reviews, August 10, 2026