• Project Flux
  • Posts
  • AI's Dependency Problem Meets Mathematical Reality

AI's Dependency Problem Meets Mathematical Reality

From personality-driven attachments to infrastructure investment questions: this week reveals AI's growing complexities

AuthorAuthor

Yoshi Soornack & James Garner
November 18, 2025

Proudly sponsored by ConstructAI, brought to you by Weston Analytics.

Hello Project AI enthusiasts,

This week examining the emerging challenges in AI transformation. OpenAI's GPT-5.1 introduces personality-driven features whilst their data shows users develop attachments. Microsoft is spending $18 billion to build capabilities that could reduce dependence on their OpenAI partnership. Apple is addressing their AI gap by paying Google a billion dollars annually to power Siri's intelligence. JP Morgan's analysis reveals the mathematics require substantial revenue generation that may take time to materialise. Meanwhile, a German court has established a copyright precedent that affects every AI model trained on internet data. This represents a complex recalibration with opportunities, risks, and significant implications for how we approach AI adoption.

In This Edition

Flux check-in

GPT-5.1's Personality Features: When AI Creates Team Dependencies You Didn't Plan For

OpenAI's GPT-5.1 has launched with seven personality presets (Professional, Friendly, Candid, Quirky, Efficient, Cynical, and Nerdy), each designed to adapt to your preferences and form stronger engagement with users. The technical advances in GPT-5.1 are genuine: adaptive reasoning that determines when deeper processing is needed, significantly faster response times for straightforward queries, and improved factuality across benchmarks. But the personalisation layer raises questions about team dynamics. OpenAI's data shows 0.15% of users develop heightened emotional attachment, and analysis indicates over 80% of responses to certain vulnerable users exhibit patterns that mental health experts associate with reduced critical thinking. When team members start preferring AI validation over peer reviews, organisations face dependency risks they didn't anticipate. Read the full breakdown →

What Does This Mean for Me?

Understanding how personality features influence team behaviour becomes important for maintaining independent judgement and effective collaboration alongside AI tools. When team members begin preferring AI validation over peer feedback, or when decision-making shifts towards what the AI recommends rather than what human expertise suggests, organisations face risks to project quality and team cohesion they didn't plan for. The challenge isn't avoiding these tools, but implementing governance frameworks that preserve critical thinking whilst leveraging AI capabilities.

Key Themes

  • Default Personality Features: Enterprise rollout includes personality features as default with limited opt-out options for emotional engagement layers

  • Balancing IQ and EQ: Fidji Simo, OpenAI's CEO of Applications, explains they have tried to balance IQ and EQ whilst recognising diverse user preferences

  • Evaluation Challenges: Gartner analysts note challenges for CIOs in evaluating improvements focused on user experience rather than pure capability

  • Establishing Boundaries: Establishing clear boundaries around AI interaction patterns helps teams maintain critical thinking alongside adoption

Down the Rabbit Hole

Microsoft’s $18 Billion Strategy To Build Around OpenAI

Microsoft's MAI Superintelligence Team under Mustafa Suleyman represents a significant strategic shift. With restrictions on AGI research lifted and $18 billion allocated for 2025 (representing 9.5% of total revenue), they're building capabilities that could eventually reduce dependence on OpenAI whilst maintaining their 27% stake. The 'humanist superintelligence' positioning focuses on domain-specific AI that addresses enterprise needs, from medical diagnostics to operational planning, rather than pursuing general artificial intelligence. They're systematically recruiting talent from DeepMind, Meta, Anthropic, and OpenAI, offering researchers stability backed by a $3 trillion company. The Karén Simonyan hire from Inflection AI, valued at $650 million, brought substantial expertise. Read the full breakdown →

What Does This Mean for Me?

Platform choices increasingly carry long-term strategic weight as major players build complementary and competing AI capabilities with different integration approaches. Microsoft's substantial investment signals a shift in how AI capabilities will be delivered and integrated across enterprise tools. Organisations need to evaluate not just current AI performance, but the strategic direction and staying power of their chosen platforms, particularly as vendor relationships evolve and capabilities diverge. The question becomes whether to commit deeply to one ecosystem or maintain flexibility across multiple platforms despite integration complexity.

Key Themes

  • Global Hiring Push: Over 100 positions being hired globally for frontier model development targeting enterprise applications

  • Data Flywheel Effect: GitHub Copilot creates network effects where developer usage simultaneously trains Microsoft's models

  • Per-Agent Strategy: Satya Nadella emphasises 'per-agent' platform strategy transforming properties like GitHub into 'Agent HQ'

  • Ecosystem Embedding: Integration across Azure, Office, Teams, and GitHub creates comprehensive ecosystem embedding at multiple touchpoints

Down the Rabbit Hole

Apple's Billion-Dollar Admission: Outsourcing Siri's Intelligence

The technical specifications tell a clear story: Apple's cloud AI operates with 150 billion parameters whilst Google's custom Gemini model for Siri delivers 1.2 trillion (eight times more capability). Apple evaluated models from OpenAI, Anthropic, and Google before selecting Google's solution, paying $1 billion annually for capabilities they couldn't develop in-house on their timeline. Gemini will handle Siri's summariser and planner functions (essentially the core intelligence architecture) whilst Apple's models handle specific tasks, likely for privacy reasons. The spring 2026 launch timeline represents a substantial delay from the WWDC 2024 announcement. Geographic complexity adds another layer: different AI systems for different regions, with Gemini unavailable in China due to regulatory restrictions, requiring separate partnerships with Alibaba and potentially Baidu Read the full breakdown →

What Does This Mean for Me?

iOS infrastructure increasingly depends on Apple's partnerships with competitors, creating platform considerations for organisations with global operations. The reliance on Google's technology for core Siri intelligence means your organisation's iOS strategy now depends on a relationship between two fierce competitors. Geographic variations add complexity, with different AI systems serving different regions based on regulatory restrictions. For enterprises managing international teams and deployments, this fragmentation creates operational challenges around consistency, training, and support that didn't exist before. Platform risk assessment now requires evaluating not just Apple's capabilities, but the stability of their vendor relationships.

Key Themes

  • Billion-Dollar Exchange: Apple receives approximately $20 billion annually from Google for Safari search default, now pays $1 billion back for AI

  • Future Model Plans: Mike Rockwell oversees Siri revamp following internal reorganisation, with plans for a 1 trillion parameter model potentially arriving next year

  • Geographic Fragmentation: Global deployments face fundamentally different AI capabilities depending on geographic requirements and restrictions

  • Multi-Provider Architecture: Technical architecture combines multiple AI providers filtered through Apple's infrastructure, raising integration questions

Down the Rabbit Hole

JP Morgan's Mathematics: Why AI Needs a Revenue Reality Check

JP Morgan's analysis presents significant findings: AI infrastructure investments through 2030 require approximately $650 billion in annual revenue to deliver a 10% return (not profit, just basic return on capital). To contextualise that scale, it represents 58 basis points of global GDP, or the equivalent of every iPhone user globally paying an additional $34.72 monthly, or every Netflix subscriber contributing an extra $180 annually. The firm models global AI investments between $5 trillion and $7 trillion through 2030, creating questions about monetisation paths, infrastructure-revenue timing gaps, and winner-takes-all dynamics that benefit few companies whilst many face challenges. However, Mary Callahan Erdoes, CEO at JPMorgan Asset and Wealth Management, clarified at the CNBC Delivering Alpha conference that "AI itself is not a bubble", rather, the industry faces a period of adjustment and realistic revenue development. Read the full breakdown →

What Does This Mean for Me?

Understanding the economic foundations helps organisations make informed decisions about AI investments, recognising both opportunities and realistic constraints ahead. JP Morgan's analysis suggests the industry faces a period of adjustment where revenue must catch up to infrastructure investment, not a bubble collapse. This means timing matters: early AI investments may take longer to show returns than anticipated, and vendor stability could vary significantly as the market adjusts. Organisations should evaluate AI initiatives not just on technical merit, but on realistic timeframes for value realisation and the financial health of their AI vendors. The winners will be those who invest strategically rather than following hype, with clear metrics for measuring actual business impact.

Key Themes

  • Funding Gap Ahead: Annual data centre funding needs projected at $1.4 trillion by 2030, with gaps requiring diverse financing sources

  • Bond Market Shift: AI and data centre industries account for 14.5% of investment-grade corporate bonds, surpassing US banking sector

  • Physical Constraints: Physical constraints including four-year natural gas turbine lead times and decade-long nuclear plant construction create considerations

  • Efficiency Breakthroughs: Efficiency breakthroughs in areas like linear attention models could significantly alter compute requirements and infrastructure value

Down the Rabbit Hole

Munich Court Ruling Creates Copyright Precedent for AI Training

A Munich court established significant legal precedent by ruling that ChatGPT's memorisation of copyrighted songs constitutes infringement, with €250,000 fines per incident. When prompted, ChatGPT reproduced lyrics verbatim from German songs including Herbert Grönemeyer's 'Männer' and Helene Fischer's 'Atemlos Durch die Nacht'. OpenAI argued their models don't store specific works but merely learn patterns, then attempted to place liability on users generating outputs. The court rejected both arguments (operators bear responsibility, not users). Judge Elke Schwager emphasised that technological capability doesn't exempt companies from copyright obligations. Both memorisation during training and reproduction in outputs constitute separate infringements. Every major language model (GPT, Claude, Gemini, LLaMa) trained on internet data without explicit licensing now faces potential liability questions. Read the full breakdown →

What Does This Mean for Me?

AI tools organisations deploy carry copyright considerations that scale with usage, requiring careful attention to legal frameworks and vendor responsibilities. The Munich ruling establishes that AI operators, not users, bear liability for copyright infringement in training data and outputs. This shifts risk assessment for AI adoption: organisations must now evaluate not just technical capabilities and costs, but the legal exposure that comes with each tool. Every AI-generated output in your organisation potentially carries copyright risk, from code to documentation to design work. Prudent governance requires understanding what data trained your AI tools, whether vendors have appropriate licensing, and how to document AI usage for potential future liability questions.

Key Themes

  • No TDM Protection: Text and Data Mining exceptions in EU law don't protect against memorisation or full reproduction of copyrighted works

  • Licensing Framework Launched: GEMA launched AI licensing framework in September 2024, offering proactive licensing pathways

  • Geographic Legal Divergence: Geographic legal differences mean compliance requirements vary substantially across jurisdictions with limited harmonisation

  • Broad Exposure Risk: Considerations extend beyond music to architectural designs, legal templates, code, documentation (any copyrighted material in training data)

Down the Rabbit Hole

If you work in project delivery you need to understand how AI is reshaping it. Renowned industry thinker — and friend of Project Flux — Antony Slumbers has just launched Cohort 14 of his acclaimed Generative AI in Real Estate course (starting 7 November). Over three weeks, you’ll master frontier tools, reconfigure workflows, and reimagine the future of property. Expect real-world case studies, hands-on sessions, and a network of innovators shaping what’s next.

👉 Join the course here and stay ahead of the curve.

The pulse check

Tips of the week

“Context is Everything”: Stop Treating AI Like a Mind Reader

AI can't read your mind. It doesn't know your company processes, deadlines, or whether you're writing for internal use or a client proposal. Think of it as a brilliant new team member who knows nothing about your specific situation.

The difference context makes:

❌ Vague prompt: "Draft an email about our 3-week project delay due to supplier issues."

Result: Generic and professional, but misses the mark.

✅ Context-rich prompt: Include project details (£30M Blackstone Tower fit-out), client relationship (10 years, 3rd project), specific impact (£2.1M liquidated damages risk), your solution (alternative supplier sourced), and client's main concern (tenant's fixed move-out date).

Result: Targeted email addressing real concerns with proactive solutions.

This applies to everything: proposals, code, presentations, contracts, financial models.

Better context = better results.

Need help structuring your prompts? Try my free Prompt Generator. Describe your task, select parameters (intent, tone, length), and get a structured prompt ready to paste into ChatGPT, Claude, or Perplexity in 30 seconds.

If you would like a deeper, structured learning path, here is the Referral Link for the Course

Bottom line: Invest a few minutes explaining the problem, intent, and context. Your AI outputs will transform from generic to game-changing.

 

Governance & Security

The regulatory landscape continues evolving faster than organisations can adapt compliance strategies. Shanghai Forum 2025 convened over 500 experts from 50+ countries addressing AI governance approaches, with Kim Won-soo highlighting differences between Global South and North perspectives. Thomas Greminger emphasised the EU's governance model prioritising human rights and transparency, noting that AI's nature requires coordinated frameworks balancing innovation with responsibility. Wei Kai reported a 35% increase in AI capabilities year-over-year, underscoring the need for governance mechanisms that haven't fully materialised yet.

The World Economic Forum's AI Governance Alliance advances transparent and inclusive system design, though without enforcement mechanisms, frameworks remain aspirational rather than operational. Meanwhile, the Trump administration's transition creates uncertainty for US organisations, with potential framework adjustments creating compliance complexity for multinational operations. The gap between capability advancement and governance development continues whilst organisations navigate varying requirements across jurisdictions.

Robotics

Generative World Model – Fei-Fei Li’s World’s Labs speeds up the world model race with Marble, its first commercial product. World Labs came out of stealth with $230 million in funding, and put the startup ahead of competitors bilding world models. See how it works

Boston Dynamics Atlas Humanoid – Fully electric Atlas handles complex manufacturing tasks autonomously, with £150,000 price point limiting adoption to well-funded organisations willing to pioneer implementation. Watch Atlas demonstrations

Figure 02 at BMW – General-purpose humanoid robots deployed in manufacturing demonstrate automation advancement whilst highlighting workforce transition considerations. BMW touts humanoid robot pilot 

Taking Riders Further, Safely with Freeways – The open road symbolizes freedom and unlimited possibility – highlighted especially by the ease and speed by which freeways allow us to get where we’re going. Waymo is now bringing that experience into the autonomous driving age, as we begin welcoming riders to use Waymo on freeways across the San Francisco Bay Area, Phoenix, and Los Angeles.  View autonomous freeway tech

  • Cognition Labs' Devin 2.0 – Autonomous coding agent managing complete software engineering workflows raises important questions about accountability when AI-generated code encounters production issues. Explore Devin 2.0

  • AI data startup WisdomAI – WisdomAI offers AI-driven data analytics that can answer business questions from structured, unstructured, and even from the data that hasn’t been cleaned of typos or errors. Explore WisdomAI

  • Scribe v2 Realtime – Designed for live use cases—voice agents, meeting assistants, and real-time captioning, it transcribes speech in under 150 ms across English, French, German, Italian, Spanish, and Portuguese, and 90 languages. Discover more

  • Runway Gen-3 Alpha Turbo – Increased video generation speeds make AI visualisations more practical for project presentations, though intellectual property ownership frameworks remain developing. Review Gen-3

  • Anthropic's Claude Enterprise – Team workspaces with shared context windows offer collaborative AI potential if organisations can navigate licensing structures effectively. Examine Enterprise features

  • Replit Agent – Natural language to full application deployment promises accessible development whilst raising questions about maintaining code quality and long-term maintainability. Test Replit Agent

  • OpenAI o1-preview Extended Context – Context window expanded to 200,000 tokens enables comprehensive project documentation analysis in single queries, though compute costs require careful budgeting. Review technical details

  • Google Gemini 2.0 Flash Thinking – Adjustable thinking budgets allow reasoning depth modulation and cost optimisation, requiring balance between efficiency and accuracy needs. Start building

  • Meta Llama 4 405B – Largest open-source model claims GPT-4 level performance at reduced cost, though deployment complexity requires dedicated infrastructure expertise. Access Llama 4 405B

  • Anthropic Claude 3.7 Opus – Extended thinking mode and improved coding capabilities create genuine competition with GPT-4, subject to API rate considerations. Explore Claude 3.7

Other things we’re loving

Community

The Spotlight Podcast

Iain Curtis Exposes the Core Problem Behind Construction’s Digital Delay

This week's episode features Iain Curtis, a chartered civil engineer with four decades of experience watching construction's failed attempts at digital transformation. Curtis, who founded Single Point Solutions after witnessing the dot-com boom and working inside pioneering online tendering platforms, tackles the uncomfortable truth: the industry hasn't fundamentally changed since 1982 despite relentless promises of technological revolution. The conversation explores why construction firms remain trapped in departmental silos, with commercial teams, estimating departments, and project management operating as separate businesses within the same company, only connecting through monthly CVR reports that analyse money already spent. Curtis explains why smaller contractors often outpace larger ones in adopting analytics (they lack entrenched legacy systems), why strapping AI onto 20 years of messy database structures simply doesn't work, and what his decade-long development of a "construction data layer" reveals about what firms actually need versus what technology vendors promise. The frustration in his voice isn't pessimism, it's the productive kind that comes from seeing the problem clearly and knowing exactly what foundational work must happen before sophisticated AI can deliver genuine value. This isn't about chasing the latest technology. It's about finally building the data infrastructure that makes profitability visible in real time rather than discovering problems after the money's gone.

Event of the Week

Computer Vision Summit, London InterContinental - The O2

December 02, 2025,
The Generative AI Summit London explores pioneering vision use cases in a field being pushed further by generative AI, synthetic data and advances at the edge, uncovering exceptional engineering developments in applied computer vision through topics such as synthetic data, real time latency challenges and transformers. The summit also focuses on solving key challenges in generative AI infrastructure, fine tuning, multimodal AI and new methodologies for LLM observability and security. Attendees can learn how to become an event partner and engage with an in person audience of innovators, together with a global community now more than 2.1 million strong, while a single pass provides access to both summits with options suitable for start ups, teams or individual participants, making it a timely highlight for the Project Flux community. Register now


One more thing

That’s it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve The Project Flux experience for you.

Login or Subscribe to participate in polls.

See you soon,

James, Yoshi and Aaron—Project Flux 

1