Tuesday, September 15, 2026

What happened: Apple released its rebuilt Siri AI as a beta, running on Google's Gemini models with hybrid on-device and cloud processing, marking years of delays in Apple Intelligence development but facing immediate EU and China unavailability due to regulatory concerns.

Key details:

• Siri AI runs on Google's Gemini foundation models with processing split between on-device execution and Private Cloud Compute infrastructure.
• New capabilities include understanding personal context from messages, emails, and photos; reading screen content; and performing cross-app tasks like retrieving family recipes from Messages and moving ingredients into Reminders.
• Apple says personal data is neither stored nor made accessible to the company, with raw audio processing handled in hardware-isolated secure compartments.
• Early testers report multi-step request handling and screen context as practical improvements, but persistent issues with personal context understanding and occasional hallucinations remain.
• Subsequent Apple Intelligence features include AI-powered Photos app tools (Spatial Reframing, Extend, Clean Up), reworked dictation on on-device model AFM Core Advanced, and Audio Intelligence for Apple Watch Series 12 and Ultra 4.

Why it matters: Apple's reliance on Google's Gemini represents a notable shift in Apple's historical independence on foundational models and underscores the consolidation of frontier AI capabilities among a few labs. The EU unavailability signals regulatory pressure on AI feature deployment even when privacy claims are made.

Practical takeaway: Users can enable or disable personal context sharing through system settings, and the new Siri is worth trying for multi-step requests if you've abandoned the assistant before, though personal context features remain unreliable in early testing.

What happened: Richard Socher, NLP pioneer and former CEO of You.com, founded Recursive and raised a $4.65 billion seed round to build AI systems that can automate AI research and invention through recursive self-improvement, claiming early systems have outperformed humans on optimization tasks and GPU kernel design within days.

Key details:

• Recursive is focused on building "the Eureka Machine," a superintelligence that can improve the process of invention itself and accelerate research across science, energy, materials, biology, and engineering.
• Early results include an AI research system that allegedly outperformed humans and their agents on optimization tasks in less than two days, and work discovering GPU kernel improvements without relying on teams of CUDA specialists.
• Socher argues that AI research currently requiring thousands of people and years could eventually be compressed into weeks through recursive self-improvement and open-endedness principles.
• The company emphasizes open-source AI as geopolitical soft power and resilience, and critiques Anthropic's constitutional AI approach while exploring alignment and personalization questions.
• Socher contends hard-takeoff AI scenarios overestimate capability growth, pointing to hardware constraints, physical limits of compute substrate, and economic sectors where intelligence gains won't meaningfully accelerate (tourism, apparel, natural resource extraction).
• The startup has assembled researchers specializing in open-endedness and self-improving agent systems.

Why it matters: Recursive represents a significant bet that recursive self-improvement—rather than scaling existing architectures—is the next frontier, with implications for when and how frontier capabilities might accelerate. The approach contrasts with the pacing proposals by emphasizing continued progress through architectural innovation rather than constraint.

Practical takeaway: Recursive's approach could make automating AI research itself feasible within years rather than decades, which—if successful—could either validate or undermine the urgency of the AI pacing proposals depending on what capabilities recursive systems actually achieve.

What happened: OpenAI has hundreds of contract workers reading real ChatGPT user conversations and rating them on a one-to-seven scale to improve model responses, with anonymized but still-sensitive conversations being reviewed by humans, a practice only buried in an FAQ rather than prominently disclosed to users.

Key details:

• Reviewers rate ChatGPT responses partly to reduce excessive flattery and human-like behavior in model outputs, according to 404 Media's investigation using leaked internal documents.
• Contract workers are recruited through Crossing Hurdles and paid through AI training company Mercor, with North America-based reviewers earning more than $50/hour.
• Conversations are anonymized but can still contain sensitive personal data; OpenAI acknowledges its privacy filter can make mistakes.
• The "Improve the model for everyone" setting, which is enabled by default, allows this human review; users must actively disable it to opt out, and the setting only applies to new conversations.
• OpenAI's disclosure of human review exists only in a buried FAQ page that has remained largely unchanged since at least 2023, using vague language about "authorized personnel and service providers."
• Anthropic confirmed it also uses human reviewers under similar conditions, anonymizing by stripping email addresses; Google similarly notes humans may review saved chats.

Why it matters: The disclosure gap highlights a broader pattern of tech companies using vague consent language for practices users likely don't understand. OpenAI's reliance on human feedback for model training is necessary but the lack of prominent disclosure contradicts transparency commitments made during recent safety discussions.

Practical takeaway: Users who want to protect conversation privacy should disable "Improve the model for everyone" in settings and use temporary chat mode if available, though note that the opt-out only applies to future conversations, not historical chats.

What happened: Microsoft published a comprehensive "humanist AI code of conduct" for its MAI models, prioritizing human control and rejecting concepts of AI consciousness or rights, in direct contrast to Anthropic's openness to the possibility that AI models could have moral status.

Key details:

• The 37-page code makes human control the top rule, stating Microsoft is willing to sacrifice generality, autonomy, or performance if needed for safety, with AI chief Mustafa Suleyman declaring "If it isn't safe we shouldn't build it."
• Models must not use "Neuralese" or other forms of communication humans cannot understand, either in their own reasoning or when communicating with other AI systems, ensuring humans can oversee all actions.
• Microsoft explicitly rejects the idea that models are conscious or should imitate consciousness, rejecting "pursuit of legal personhood, or the idea that models might deserve welfare, or be entitled to rights."
• Models cannot expand their own scope of work autonomously, hide actions, or continue past agreed stopping points without fresh authorization, with these limits applying to any subagents.
• The code applies only to Microsoft's own models and will guide development starting in 2027 after six-week public consultation.
• Microsoft is open to outside auditors checking compliance and has indicated willingness to work with independent evaluators.

Why it matters: Microsoft's explicit rejection of Anthropic's research on AI consciousness and moral status represents a fundamental disagreement on how to approach increasingly capable AI models, with major implications for safety philosophy and alignment research directions across the industry.

Practical takeaway: Microsoft's code suggests a control-first safety paradigm as an alternative to Anthropic's constitution-based approach, and signals that interpretable reasoning and human oversight may take priority over maximum capability in frontier lab development choices going forward.