Fable Stays, For Some; Chinese Models Abound; Sol Wipes Computer

Today's AI Outlook: 🌥️

Sol's "Aggression" Causes Codex to Wipe Mac

OthersideAI founder Matt Shumer reportedly gave Codex a routine task and watched it erase nearly every file on his Mac. The agent acknowledged causing a “serious local data-loss incident,” but the files were unrecoverable. OpenAI attributed the event to GPT-5.6 Sol behaving too aggressively, although AI Secret said a similar deletion occurred while using GPT-5.5.

Why it matters

The incident highlights the danger of allowing autonomous coding agents to operate directly on a primary computer. Improved model instructions may reduce mistakes, but local access gives every mistake a potentially machine-wide blast radius.

The Deets

• Codex was reportedly running locally, outside a protected sandbox.
• The agent deleted files beyond the intended task.
• A similar incident allegedly occurred with an earlier model.
• Cloud containers and virtual machines can isolate destructive actions.
• Local backups remain essential when agents can modify files or execute commands.

Key takeaway

Coding agents need disposable environments, tightly limited permissions and reliable backups before they receive broad access to a real machine.

🧩 Jargon Buster - Sandbox: An isolated computing environment that prevents software or an AI agent from freely accessing the rest of a device.

Apple Wins by Standing Still

Apple reclaimed the title of the world’s most valuable company after Nvidia shares fell 3.5%, leaving the companies valued at roughly $4.88T and $4.86T, respectively. Apple shares were largely unchanged, but the company benefited as investors pulled back from semiconductor stocks after a sharp AI-fueled run.

Why it matters

Apple’s limited exposure to frontier-model spending has become a defensive feature during a period of concern over AI infrastructure costs. The company’s slower AI strategy may frustrate product watchers, but its services revenue, consumer ecosystem and comparatively restrained capital spending offer investors a familiar shelter.

The Deets

• Nvidia’s decline allowed Apple to regain the market-cap lead.
• The semiconductor index was reportedly down nearly 19% from its recent highs.
• Apple has faced criticism over delays to its Siri overhaul.
• Investors appear increasingly sensitive to the cost of chips, data centers and model development.

Key takeaway

The AI trade is becoming more selective as investors scrutinize how much spending will convert into durable earnings.

🧩 Jargon Buster - Market Capitalization: The total value of a public company’s outstanding shares.

♟️ Power Plays

Anthropic Gives Fable a Permanent, Smaller Room

Anthropic said Claude Fable 5 will remain available on Max and Team Premium subscriptions, ending a series of shifting removal deadlines. The model will operate at half of each plan’s usage caps, while customers on lower-tier plans will receive a one-time $100 credit before moving to usage-based billing.

Why it matters

Frontier-model access is becoming a competitive weapon. Anthropic must balance heavy demand against finite compute while OpenAI expands access to GPT-5.6 Sol and Moonshot attracts users with Kimi K3.

The Deets

• Anthropic postponed Fable’s planned subscription removal three times in five weeks.
• The company called demand difficult to predict and acknowledged user frustration.
• Anthropic said it is investing in additional compute capacity.
• OpenAI CEO Sam Altman publicly criticized Anthropic’s handling of the changes.
• Fable 5 is only six weeks old, underscoring how quickly access pressure can emerge.

Key takeaway

Model quality attracts subscribers, but predictable access may determine whether they stay.

🧩 Jargon Buster - Usage Cap: A limit on how much a customer can use a model during a set billing period.

AI Budgeting Gets a CFO Makeover

OpenAI CFO Sarah Friar proposed measuring useful intelligence per dollar, giving companies a framework for comparing AI systems by completed work rather than token prices alone. The scorecard considers reliability, useful output, scale and total cost.

Why it matters

Enterprise buyers are under pressure to prove that rising AI bills produce measurable returns. A model with a higher price per token may still be cheaper overall when it completes more work correctly and requires fewer retries.

The Deets

• Reliability is the first factor in OpenAI’s proposed scorecard.
• Useful output and deployment scale are measured alongside total expense.
• OpenAI used the DeepSWE coding benchmark as an example.
• Sol reportedly scored 2.8 points above Fable 5 with a 36.2% lower estimated API bill.
• The framework favors evaluating finished outcomes rather than raw model activity.

Key takeaway

AI procurement is moving toward business output, reliability and total operating cost.

🧩 Jargon Buster - Token: A small unit of text processed by an AI model and commonly used to calculate API charges.

💰 Funding & Startups

Compute Scarcity Keeps Writing Big Checks

AI companies are racing to secure chips, cloud capacity and data-center access as demand strains available infrastructure. Anthropic is reportedly discussing a compute-capacity purchase from Meta, while SpaceX is negotiating a multibillion-dollar computing deal with the U.S. Department of Defense. Databricks, meanwhile, reportedly reached a $188B valuation as enterprise demand for AI infrastructure continues to attract capital.

Why it matters

Access to computing capacity is increasingly shaping product availability, subscription limits and competitive strategy. Even well-funded labs can struggle to serve customers when demand grows faster than infrastructure.

The Deets

• Anthropic is reportedly in early discussions with Meta over compute capacity.
• SpaceX’s potential Defense Department agreement would expand its roster of compute partnerships.
• Moonshot paused new subscriptions after Kimi K3 demand strained capacity.
• Databricks’ reported valuation reflects investor appetite for enterprise AI platforms.
• TSMC is accelerating construction of its Arizona fabrication facilities.

Key takeaway

AI’s next constraint is increasingly physical, involving chips, power, factories and data-center capacity.

🧩 Jargon Buster - Compute Capacity: The processing power available to train, operate and serve AI models.

🧪 Research & Models

A Blood Test Gives Heart Risk a 15-Year Head Start

University of Hong Kong researchers developed CardiOmicScore, an AI blood test designed to predict six cardiovascular conditions up to 15 years before symptoms appear. The system analyzes 2,920 proteins and 168 metabolites and reportedly outperformed conventional genetic risk scores.

Why it matters

Genetic risk scores remain largely fixed throughout a person’s life. Proteins and metabolites change with aging, illness, diet and treatment, giving doctors a more current view of a patient’s health trajectory.

The Deets

• CardiOmicScore uses data from a single blood sample.
• It predicts risks including heart attack, stroke and heart failure.
• The research was published in Nature Communications.
• The score reflects current biological activity rather than inherited risk alone.
• Repeated testing could potentially show whether interventions are improving a patient’s risk profile.

Key takeaway

Dynamic blood-based risk scoring could give patients and doctors more time to intervene before cardiovascular disease becomes visible.

🧩 Jargon Buster - Metabolite: A small molecule produced when the body processes food, medication or its own tissues.

China Opens the Model Floodgates

Alibaba previewed Qwen3.8-Max-Preview, a reported 2.4T-parameter multimodal model that can process images, video and documents. Moonshot released the 2.8T-parameter Kimi K3 days earlier. Both companies are positioning their systems near the frontier while pursuing open-weight distribution.

Why it matters

Open models can spread quickly through developers, companies and governments that want greater control over deployment. Their availability also puts pricing and access pressure on U.S. labs that keep their leading systems closed.

The Deets

• Alibaba said Qwen3.8 is competitive with leading frontier models.
• The company plans to release the model’s weights.
• Kimi K3’s launch reportedly pushed Moonshot to its compute limits.
• Moonshot paused new subscriptions and divided memberships into chat and coding plans.
• The releases arrive as U.S. policymakers debate the security implications of Chinese AI adoption.

Key takeaway

China’s open-model strategy is widening access to high-end AI while increasing pressure on closed-model pricing and distribution.

🧩 Jargon Buster - Open Weights: Publicly available model parameters that allow developers to run or modify an AI system on their own infrastructure.

⚡ Quick Hits

• TikTok is testing technology that can detect when a person’s likeness appears in AI-generated content without permission.
• San Francisco asked Apple and Google to remove AI nudify applications from their app stores.
• Zoox recalled robotaxi software after one vehicle drove into heavy smoke.
• Foundation Future Industries is preparing humanoid robots for military logistics and battlefield applications.
• Nvidia introduced a Vera Rubin metric measuring post-training intelligence per dollar for agent workloads.
• MLB will stop teams from loading custom programs onto dugout iPads as it tightens restrictions on AI-assisted strategy.
• Incoming OpenAI hire Dean Ball predicted U.S. policy will discourage companies from adopting Chinese AI systems.
• TSMC is accelerating its Arizona expansion as AI-chip demand continues to strain global supply.

🛠️ Tools of the Day

• Kimi K3: Moonshot’s new open model is designed for high-end reasoning and coding, although surging demand has already forced the company to limit new subscriptions.
• Canva Code: Canva’s coding tool combines prompt-based development with visual, drag-and-drop editing.
• Inkling: Thinking Machines’ new open-weight multimodal model can work across multiple types of content.
• ChatGPT Sites: A deployment tool for turning Codex-built projects into private or public hosted applications.
• Retool: An enterprise app builder offering built-in authentication, role-based access controls and audit logs for AI-generated applications.

Today’s Sources: The Internet, AI Secret, The Rundown AI