The contradiction is almost too perfect to be accidental.

Last week, Apple published its NeurIPS paper on STARFlow—a breakthrough in generative AI that achieves high-resolution image synthesis using normalizing flows, rivaling diffusion models while maintaining exact likelihood modeling and faster inference. This is serious, cutting-edge AI research that positions Apple as a genuine player in the generative AI space.

The same week, Apple released its Christmas campaign: "A Critter Carol," a meticulously crafted handmade puppet production. Filmed with real puppets, real puppeteers, real craftsmanship. Even the typography was created using woodblock print techniques. The behind-the-scenes content proudly showcases the human effort, the visible imperfections, the tactile reality of it all.

The implicit message? Human-made. Real.

The Familiar Pattern

This echoes what we explored in MC.84: AI and the Handmade Tale, the entertainment industry's predictable cycle of creating false binaries between "authentic handmade" and "artificial computer-generated" content.

But here's where it gets interesting: Apple isn't just marketing against AI. It's simultaneously advancing AI research while publicly celebrating craft in its consumer messaging.

As usual, hand made is always computer corrected

Here's what nobody's talking about: both "A Critter Carol" and the recent Apple TV glass icon sequences went through post-production. Color grading. Sound design. VFX teams. Compositing. Digital integration. The woodblock typography was digitized and composited. The glass icons were rendered and refined.

Nobody says "this video was made without electricity" or "this film was shot without computers." We don't obsess over the technology stack, we celebrate the creative vision that shaped it.

The same should apply to AI. The question isn't "was AI used?" It's "was the work shaped by intentional creative choices?"

Celebrating Craft, Not Tool Purity

What Apple is actually demonstrating is something far more optimistic than the "AI vs. no AI" debate: craft as a principle transcends any specific tool.

When you watch "A Critter Carol," you're witnessing:

• Puppeteers making split-second performance decisions

• Directors choosing imperfection over polish

• Production designers building tangible worlds

• Cinematographers composing light and shadow

• Sound designers layering texture and emotion

These are human creative choices. The fact that they went through digital post-production, color grading, and technical refinement doesn't diminish that. It enhances it.

Similarly, the Apple TV visual identity—those gorgeous glass apple icons—represents intentional aesthetic choices, sophisticated design thinking, and creative vision. The tools used to realize that vision are irrelevant to the quality of the idea.

The Future Isn't Either/Or

We're not heading toward a world where you have to choose between craft and technology. We're heading toward a world where craft is how you use technology.

The most compelling creative work will come from teams that:

• Use AI to accelerate iteration, not replace thinking

• Leverage automation to handle the tedious, freeing humans for the intentional

• Combine computational power with human judgment

• Treat every tool—whether it's a camera, a render engine, or a generative model—as a means to creative expression

Just as no one asks "was this film made without electricity?" we'll stop asking "was this made without AI?" The real question will always be: Does this work demonstrate intentional creative vision?

The Paradox Dissolves

Apple's publishing cutting-edge generative AI research while celebrating handcrafted puppetry isn't a contradiction. It's the clearest possible statement: we believe in craft as a principle, and we'll use whatever tools—including AI—to serve that principle.

The paradox only exists if you believe the tools matter more than the vision. Apple is betting they don't.

And they're probably right.

Until next Thursday 🎉
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

"If software is eating the world, Anthropic is eating software." This week, I want to talk about a significant development that I believe profoundly changes how we'll interact with technology and how work gets done: Anthropic's Claude Opus.

Having spent some time delving into the latest iteration of Opus, my primary takeaway is clear: Opus is deeply oriented towards code, and in an increasingly digital world, everything is becoming code.

Anthropic isn't just releasing another powerful model; they're strategically pushing the boundaries of code execution, a move that, while subtle, has massive implications. Two recent announcements perfectly illustrate this direction:

1. Advanced Tool Use: This powerful feature allows Claude to dynamically discover, learn, and execute tools. Imagine an AI that can intelligently navigate vast tool libraries, calling upon only the necessary functions for a given task. This drastically reduces the context window burden and significantly enhances accuracy for complex, multi-step workflows, moving beyond simple function calling to sophisticated, programmatic orchestration.

2. Claude For Excel: This integration brings Claude directly into your spreadsheets. It's not merely a chatbot; it's an intelligent agent capable of understanding entire workbooks—from intricate nested formulas to multi-tab dependencies. It can debug errors, test scenarios, and even build models, all through intelligent code execution directly within Excel.

These aren't merely incremental updates; they represent a concerted effort to empower AI to perform "real work" by leveraging code at its core.

We are witnessing a fundamental paradigm shift. Instead of humans learning the intricacies of various software applications, we're moving towards a model where a request is made, and an AI constructs the entire artifact by abstracting and executing code. This means less time spent on manual operations and more focus on strategic outcomes.

The Strategic Power of Opus Pricing

The most telling sign of Anthropic's strategic intent, beyond the technical capabilities, lies in their innovative pricing for Opus. Historically, high-tier models like previous Opus versions (e.g., Opus 4.1's Input: $15/MTok, Output: $75/MTok) were often cost-prohibitive for widespread adoption in daily workflows.

However, the new pricing for Opus 4.5 dramatically changes the landscape, bringing it down to $5/million tokens for input and $25/million tokens for output. This isn't just a discount; it's a strategic realignment that positions Opus as a highly competitive and “accessible” model.

This aggressive pricing strategy, bringing Opus's cost significantly closer to models previously considered more affordable, signals a clear message from Anthropic: they intend for Opus to become the go-to model for a vast range of tasks. It makes the cutting-edge capabilities of Opus—especially its proficiency in code and agentic workflows—available to a much broader audience, from individual developers to large enterprises.

This pricing shift, coupled with features like Advanced Tool Use and Claude for Excel, indicates that Anthropic sees code execution as the future of AI-driven productivity. They are making a bold move to commoditize high-performance, code-centric AI, establishing it as the new baseline for how we interact with software and produce tangible results.

We're moving beyond a simple "bigger is better" model race. The future is about smarter integration, efficient execution, and specialized agents that transform raw instructions into finished products through the elegant power of code.

Until next Thursday 🎉
Olivier

Bonus: I love when the model thinks outside the box. Instead of using code execution and Python libraries to create a PDF presentation, Opus simply uses a headless browser to save an HTML presentation as a PDF. Genius

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

The initial benchmarks came in for Gemini 3.0 on November 18th, and the reaction was... measured.

Not disappointing. Not revolutionary. Just... measured.

After months of teasing, after the Cloudflare outage that made it feel like Google was literally breaking the internet with the weight of their model announcement, after the promises of "multimodal understanding" and "agentic capabilities," the response from the community was remarkably subdued: it's good, but not transformative.

This matters more than you might think. Because what we're witnessing isn't the end of AI progress. It's the end of the hype cycle around model releases. And that's actually a sign of maturity.

Time to defy hype and gravity

The Hype Cycle is Dead. Long Live the Hype Cycle.

For the past 18 months, every major model release followed a predictable pattern:

1. Teasing phase: Cryptic tweets, leaked benchmarks, speculation

2. Launch day: Massive benchmark jumps, breathless coverage, "this changes everything"

3. Reality check (24-48 hours): Users discover the model hallucinates, ignores instructions, or performs worse than expected on real tasks

4. Rationalization: "Well, the benchmarks are still impressive, even if..."

We saw this with GPT-4, Claude 4 Opus, Grok 4.1, and now Gemini 3.0.

But something shifted this time. The reality check didn't come 48 hours after launch. It came during the launch.

Within hours of Gemini 3.0 becoming available, users reported:

• Better performance on benchmarks, but inconsistent on real-world tasks

• Aggressive code rewriting that ignored user preferences

• Creative writing that felt mechanical compared to Claude

• Math reasoning that was solid but not revolutionary

The benchmarks showed a 31% improvement on ARC-AGI-2. The lived experience showed... incremental improvement.

What the Plateau Reveals

This plateau reveals something crucial: we've hit a point where benchmark performance and user experience have decoupled.

Gemini 3.0 is genuinely better at many things. The benchmarks aren't lying. But the marginal improvement in what users can actually do with the model is smaller than the marginal improvement in the numbers.

This is what maturity looks like in technology.

We're no longer in the phase where each new model unlocks entirely new capabilities. We're in the phase where each new model refines existing capabilities. The jump from GPT-3 to GPT-4 felt transformative because it was. The jump from GPT-4 to GPT-5 feels incremental because it is.

And that's fine. That's actually healthy.

The Real Story: Consolidation and Specialization

While everyone was watching Gemini 3.0's benchmarks, the more interesting story was happening elsewhere:

1. Coding Models are Diverging

OpenAI released GPT-5.1-Codex-Max with "Extra High" reasoning and claims of 24+ hour autonomous operation. Google released Antigravity, their agentic IDE. Anthropic is quietly shipping Claude with better tool use.

These aren't competing on general capability anymore. They're competing on specific workflows. Coding. Search. Agentic tasks.

The era of "one model to rule them all" is ending. The era of "the right model for the right job" is beginning.

2. Open Source is Catching Up (Slowly)

Deep Cogito's Cogito v2.1 hit the leaderboards. It's not beating Gemini 3.0. But it's close enough that for many use cases, the cost difference matters more than the capability difference.

This is the pattern we saw with open source software. Eventually, the gap closes enough that factors like cost, control, and customization become the deciding factors.

3. The Infrastructure Layer is Where the Real Competition Is

While everyone watches model releases, the actual competitive advantage is shifting to:

• Inference optimization (vLLM, SGLang, Ollama)

• Context management and retrieval (RAG systems, MCP)

• Agent orchestration frameworks

• Fine-tuning and adaptation tools

The model itself is becoming commoditized. The stack around it is where differentiation happens.

What This Means for Builders

If you're building AI products, the plateau is actually good news:

1. You can stop chasing the latest model: The marginal gains from switching to the newest model are getting smaller. Focus on using the current generation well.

2. Specialization pays: Instead of "works with all models," build "works best with X model for Y task." That's a more defensible position.

3. The moat is in the data and workflow, not the model: The companies winning with AI aren't the ones with the best models. They're the ones with the best data pipelines and the best understanding of their users' workflows.

4. Stability matters more than novelty: Users care about reliability and consistency more than they care about the latest benchmark improvements.

Le french CunCun

Here's what nobody wants to say out loud: we might be approaching the limits of what scale alone can achieve.

Gemini 3.0 is rumored to be around 10 trillion parameters. That's massive. The benchmarks are impressive. But the lived experience is... fine. Good. Not revolutionary.

This doesn't mean AI progress is stopping. It means the nature of progress is changing. We're moving from "bigger models = better results" to "smarter training = better results" and "better integration = better results."

The next breakthrough probably isn't another 10x parameter increase. It's probably:

• Better reasoning at inference time (which OpenAI is exploring with their reasoning models)

• Better training data quality and curation

• Better alignment and instruction-following

• Better integration with external tools and knowledge

These are harder problems than just scaling. They require more creativity, more domain expertise, more careful engineering.

The Hype Cycle Isn't Dead. It's Just Maturing.

We're not at the end of AI progress. We're at the end of the easy progress.

The Gemini 3.0 plateau isn't a sign that AI has peaked. It's a sign that we've moved from the "exponential growth" phase to the "refinement" phase. It's the same transition that happened with:

• Mobile: From "smartphones are revolutionary" to "which smartphone has the best camera"

• Cloud computing: From "cloud is the future" to "which cloud provider has the best pricing and features"

• Deep learning: From "neural networks can learn anything" to "which architecture is best for this specific task"

Each transition felt like a plateau. Each transition was actually a sign of maturity.

What Comes Next

The real story of AI in the next 12 months won't be about model releases. It will be about:

1. Agentic systems that actually work: Not just in benchmarks, but in production

2. Specialized models for specific domains: Medicine, law, finance, etc.

3. Better integration with human workflows: Not replacement, but collaboration

4. The infrastructure layer solidifying: The winners will be the ones with the best inference, the best retrieval, the best orchestration

5. The cost curve flattening: Cheaper inference, better efficiency, more accessible AI

The plateau is here. And it's actually fine.

Until next Thursday 🎉
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

The music industry stands as a revealing barometer of technological disruption.

As a hypercompetitive sector alongside fashion and film, it demonstrates how innovation reshapes creative work.

The industry's unique history—from guerrilla marketing tactics like mixtapes and street concerts to the democratization of home studios—positions it as an ideal lens through which to examine the emerging relationship between artificial intelligence and human creativity.

Operator, are you ready?

The Current Moment: AI as Creative Collaborator

Recent technological advances have introduced compelling new tools for collaborative creative workflows.

Producer.ai exemplifies this shift. The platform functions as a personal AI music agent, transforming imagination into studio-quality music through an intuitive conversational interface. Users interact with the system as they would with a producer in a recording studio, requesting help with lyrics, sound design, and arrangement.

Welcome to Producer.ai

Welcome to Producer.ai by Seth Forsgren

www.producer.ai/blog/welcome-to-producer

This democratization of music production raises an essential question: What comes next? The natural evolution appears to be live, real-time AI music generation at concerts—a frontier that feels simultaneously inevitable and unexplored.

The Lineage of Sampling: From Analog Ingestion to AI Training

To understand where AI music creation is headed, we must first recognize where it has been. The practice of sampling—extracting sonic material from existing recordings to create new compositions—represents an early, analog iteration of what modern AI models now do algorithmically.

When a hip-hop producer samples a James Brown drum break or a soul vocal, they are performing an act of cultural ingestion and transformation. They are:

• Extracting meaningful patterns from existing work

• Recombining those patterns with new intention

• Creating something that honors the source while establishing its own identity

This is precisely what large language models (LLMs) and music AI models do, but at scale and with mathematical precision. Just as a sampler ingests audio to create new compositions, modern AI models ingest vast datasets of existing music to learn patterns, structures, and possibilities. The fundamental mechanism is identical; only the execution has evolved.

The key difference: Sampling operated within clear cultural and legal frameworks (however contested). Artists acknowledged their sources. The new AI paradigm operates in murkier territory—the training data is often opaque, consent is rarely explicit, and the line between inspiration and appropriation becomes increasingly blurred.

The Critical Gap: Master Files and Individual Tracks

Yet this forward-looking enthusiasm often obscures a crucial technical limitation that deserves serious attention.

In a recent interview, the Dewaele brothers of 2ManyDJs/Soulwax highlighted a distinction that reveals AI's current boundaries.

When working on projects with designer Dries Van Noten, they had access to master files containing individually separated tracks from seminal recordings. This access enabled them to deconstruct and reimagine classic compositions with unprecedented creative freedom—isolating vocals, drums, basslines, and instrumentation to create entirely new arrangements.

Here is the AW11 show, using David Bowie’s Golden Years.

The current reality: AI remains unable to reliably extract individual tracks from finished recordings with sufficient quality for professional use. While AI-powered stem separation tools exist, they cannot yet match the fidelity of original master files.

The future possibility: When this capability arrives—and it likely will—it will unlock an entirely new creative frontier. Artists will gain access to the raw components of existing works, enabling:

• Novel remixes and reinterpretations of classic recordings

• Creative collaboration across decades and genres

• Derivative works built on the foundations of canonical music

Simultaneously, this same technology will generate vast new training datasets for music AI models, creating a feedback loop that accelerates AI capability development.

This emerging tension reveals something profound about AI's impact across all creative industries. We now inhabit an expanding spectrum:

One end: The raw origins of music—individual musicians, live instrumentation, analog processes, human error and spontaneity

The middle: Hybrid workflows where human creativity and AI capability merge—artists using AI as a collaborative tool rather than a replacement. This includes:

• Sampling (the analog precursor)

• Remixing and recontextualization

• AI-assisted composition and arrangement

• Producer.ai's collaborative workflow

The other end: Fully AI-generated compositions requiring no human intervention, created instantaneously from text prompts

What This Tells Us

The distance between these poles—from human-made music to AI-only generation, with sampling as a historical bridge—illuminates several truths about the future of creative work:

1. Technology expands possibility without eliminating craft: Sampling didn't eliminate musicianship; it expanded what musicians could imagine and execute. Similarly, AI tools like Producer.ai augment rather than replace human creativity.

2. Precedent matters: Sampling established that recontextualizing existing material can be legitimate creative practice. This precedent will shape how we think about AI ingestion of training data—but also highlights the need for ethical frameworks that sampling sometimes lacked.

3. Access democratizes but raises questions: When tools become widely available, more voices enter the conversation. But democratization also raises urgent questions about attribution, consent, and compensation that the sampling era struggled to answer.

4. The human element remains irreplaceable: The most compelling creative work will likely continue to emerge from the intersection of human intention and technological capability—just as the greatest samples became greatest when paired with human artistry.

5. History rhymes: The anxiety about AI in music echoes earlier anxieties about synthesizers, drum machines, and sampling itself. Each technology was initially viewed as a threat to "real" musicianship. Each ultimately expanded the definition of what music could be.

Conclusion

The music industry's response to AI will likely shape how other creative sectors navigate similar disruptions. The challenge lies not in resisting technological change, but in thoughtfully integrating new capabilities while preserving the irreplaceable human elements that make art meaningful.

Sampling taught us that ingesting existing material to create new work can be legitimate. AI now asks us to extend that lesson—to recognize that algorithmic ingestion of training data follows the same creative logic, but at unprecedented scale. The question is not whether this will happen, but whether we will establish ethical frameworks to govern it.

The future of music—and creativity broadly—will be determined not by the capabilities of our tools, but by the wisdom with which we choose to use them.

Until next Thursday 🎉
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

Hello everyone,

We're witnessing something remarkable: the same economic forces that reshaped open source software over the past two decades are now repeating themselves in the vibe coding revolution. And if you're not paying attention to this parallel, you're about to get blindsided by a maintenance crisis.

Here's the uncomfortable truth: AI has commoditized the wrong part of software development. We've made creation trivially cheap while massively expanding the operational burden. And just like open source proved 20 years ago, the money was never in building the software—it was in deploying, maintaining, and operating it at scale.

The Open Source Playbook: Give Away Creation, Monetize Operations

Open source software represents an estimated $8.8 trillion in value if firms had to create it from scratch. Yet the companies behind this software discovered something counterintuitive: the real money isn't in the code—it's in everything that comes after.

Software maintenance typically consumes 50-80% of total cost of ownership, often costing 3-4 times the original development over a system's lifetime. Red Hat built a multibillion-dollar business on this insight, literally making billions "selling nothing"—free software with premium operations.

The most successful open source companies monetize through:

Cloud-hosted services: MongoDB Atlas generates 72% of MongoDB's total revenue—$1.5 billion annually—not by charging for database code, but by managing deployment, backups, and scaling.

Support and maintenance: Companies willingly pay 15-25% of original development costs annually just for guaranteed reliability and rapid incident response.

Professional services: While software is free, the knowledge to deploy it effectively is valuable. Integration and deployment services alone hit $8.8 billion in 2023.

The message is clear: code itself has become commoditized, while deployment, scaling, security, and ongoing operations retain significant value.

Vibe Coding: History Repeating at 10x Speed

Vibe coding—using AI to generate code through natural language prompts—is now doing to all software development what open source did to infrastructure: making creation dramatically cheaper and more accessible.

Leading startups in this space saw their combined valuation explode from $7-8 billion in mid-2024 to over $36 billion in 2025—a 350% increase. Lovable reached $100 million ARR just eight months after crossing $1 million. McKinsey estimates AI can improve developer productivity by 35-45%.

But here's the critical parallel: vibe coding commoditizes creation while massively expanding the deployment and maintenance challenge.

This democratization means millions of non-developers can now create software. A marketing manager can build a custom CRM. A teacher can create educational apps. The barrier between having an idea and building it is disappearing.

Yet this abundance creates new problems. As Wired reported, "vibe coding tends to yield a rough draft of code that may overlook specific contexts." Even more critically, "the essential safeguards and accountability mechanisms traditionally present in open source are either missing or severely fragmented."

The Trust Problem: Why Developer Track Records Will Become Currency

Here's where the parallel breaks down—and where things get interesting.

In the open source world, trust was built through transparent contribution histories. GitHub profiles became resumes. Commit histories became credentials.

But vibe coding introduces a fundamental opacity problem: you're not betting that software works—you're betting that it will be maintained.

When AI generates code, traditional trust signals disappear. You can't evaluate the developer's expertise because there is no developer in the traditional sense. You can't assess the architecture decisions because they were made by a language model optimizing for syntactic correctness, not long-term maintainability.

This creates what I call the Vibe Coding Maintenance Trap: code that looks production-ready but is actually a ticking time bomb of accumulated technical debt.

AI-generated code creates a "Tech Debt Loop." Code that looks syntactically correct gets merged without scrutiny. When issues emerge, teams use AI to fix them, introducing further patches that compound rather than resolve underlying problems. The accumulated debt has real consequences: increased maintenance costs, slower development, and fragile systems harder to scale.

This is where developer track records become the new currency.

In a world where anyone can generate code, the differentiator isn't creation ability—it's operational reliability. Organizations will increasingly ask:

1. Who deployed this, and what's their uptime record?

2. Who's maintaining this, and what's their incident response time?

3. Who's responsible when this breaks at 3 AM?

4. What's their track record for keeping AI-generated systems running reliably?

Just as Red Hat built trust through guaranteed reliability, the winners in the vibe coding economy will be those who can demonstrate operational excellence and maintenance competence.

The Coming Operational Crisis

Both open source and vibe coding follow the same economic trajectory—creation becomes cheap, while operational excellence becomes the differentiator.

Technical debt accumulates faster: When issues emerge in AI-generated code, teams often use AI to fix them, creating layers of patches that compound problems. Organizations adopting AI-generated code will face the same brutal reality open source companies discovered: enhancements after deployment typically cost 3-4 times the original development.

Operations expertise becomes scarce: While millions can now generate code, the skills to deploy securely, scale efficiently, and operate cost-effectively remain specialized. This is the gap that can't be automated away.

Monetization Strategies: The Deployment Economy

Both paradigms reveal consistent patterns focused on deployment and operations:

Platform-as-a-Service models: MongoDB Atlas doesn't charge for database code—it charges for managed hosting with automated backups and multi-cloud support. Deployment expertise and operational reliability are far more valuable than the code itself.

Consumption-based pricing: Usage-based billing captures revenue as systems scale. This works particularly well for vibe-coded applications, where initial creation is cheap but scaling requires substantial infrastructure investment.

Support and SLA guarantees: Enterprises pay premium rates for guaranteed uptime. The value isn't in the software—it's in the promise of reliability.

Integration and deployment services: As vibe coding democratizes creation, demand grows for professionals who can integrate AI-generated components into enterprise systems securely and reliably.

The Bottom Line: Trust Becomes the Moat

We're entering an era where software creation is becoming commoditized, but operational excellence remains expensive and difficult. The winners won't be those who can generate the most code—AI handles that. They'll be those who can deploy reliably, maintain efficiently, and operate cost-effectively at scale.

Just as Red Hat built a multibillion-dollar business on "selling nothing"—free software with premium operations—the most valuable companies in the vibe coding era will monetize operational excellence, not code generation.

And in this new economy, developer track records will become the primary trust signal. Organizations will pay premium rates for teams with proven maintenance capabilities, documented uptime records, and demonstrated incident response competence.

The parallel is clear: whether code comes from open source communities or AI language models, the fundamental economics remain the same. Software creation is becoming commoditized. The money lies in deployment, maintenance, and operations—the unglamorous work of keeping systems running reliably.

The irony? We've spent the past year celebrating how AI makes everyone a developer. But the real opportunity isn't in creation—it's in becoming the person organizations trust to keep AI-generated systems running when everyone else's vibe-coded prototypes collapse under production load.

Your track record for operational reliability is about to become your most valuable asset.

Until next Thursday 🎉
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

Hello everyone,

The browser wars are back. After 15 years of Chrome's unchallenged reign, 2025 has unleashed four serious contenders—each backed by billions, each betting that agentic AI represents the inflection point that finally breaks the incumbent's stranglehold. But here's the uncomfortable truth: most of them will fail, not because their technology is inferior, but because acquiring users is brutally, systematically harder than building better software.

The Problem With Browser Acquisition

Let's start with the brutal math.

Chrome commands 68% of the global browser market—3.45 billion users—and that share is still growing. Firefox, once a genuine competitor, has withered from 50% to 3% despite superior privacy features and technical excellence. Why? Because switching browsers is friction-heavy: users must actively seek alternatives, download them, install them, change default settings, and rebuild their digital muscle memory. It's a multi-step process that most people never complete.

The distribution advantages protecting Chrome are almost unfair. It pre-installs on 71% of global smartphones (Android). Google pays Apple $20 billion annually and Mozilla $100 million+ to remain default search. Developers optimize for Chrome because that's where users are, creating a self-reinforcing cycle. Network effects compound: 250,000+ Chrome extensions create switching costs, and deep integration with Gmail, Drive, Maps, and YouTube makes leaving feel like abandoning an ecosystem rather than just changing browsers.

Most critically: the browser market doesn't reward better technology. Firefox proved this. It had privacy, speed, and developer love. It lost anyway. The lesson is harsh—distribution beats differentiation in markets where switching costs are high and user awareness is low.

This is the mountain AI browsers must climb.

The Four Challengers (And Their Odds)

Dia (The Browser Company, acquired by Atlassian for $610M) is betting on simplicity. Built on Chromium, it embeds ChatGPT into the URL bar, adds AI-powered tab comparison, and includes a Skills system for reusable workflows. The Atlassian acquisition provides enterprise distribution through 300,000+ customers. But here's the catch: it's macOS-only with Apple Silicon M1+ chips. No Windows ETA. No mobile. The addressable market is roughly 5% of Chrome's. They're playing in a sandbox.

Comet (Perplexity, $1.22B raised) is the most aggressive. It launched free globally in October 2025 after initially charging $200/month. The product is genuinely impressive—true agentic browsing where the AI navigates autonomously, integrates with Gmail/Calendar, and completes multi-step tasks. Perplexity processes 780 million monthly queries and has 15+ million MAU. They're discussing Samsung Galaxy pre-installation, which could be game-changing. But Comet tracks "everything users do online" to sell hyper-personalized ads, and the company faces copyright lawsuits from Forbes, Wired, the New York Times, and others. The business model is controversial; the legal exposure is real.

Genspark (Eric Jing, ex-Baidu/Microsoft, $160M raised) is the efficiency play. 2+ million MAU, $36M ARR in 45 days post-launch, 50 employees. The differentiation: 169 on-device AI models running locally with zero internet required. Privacy-first, no subscriptions, no cloud dependencies. The problem: Trustpilot ratings average 2.5/5 with complaints about billing and security. LayerX Security found a 90%+ compromised webpage execution rate. It's fast and efficient, but reliability concerns are serious.

OpenAI Atlas (launched this week) is the heavyweight. ChatGPT-native design where the new tab page starts with ChatGPT, not a search box. Agent Mode automates tasks. Browser Memories provide persistent context. The distribution advantage is staggering: 800 million weekly ChatGPT users as a ready-made audience. This is the one Chrome should fear—not because it's technically superior, but because it's backed by OpenAI's resources and user base.

The Uncomfortable Truth About "AI Differentiation"

Here's what nobody wants to admit: most AI browsers are solving the wrong problem.

They're adding AI features to browsers. But the real issue isn't that Chrome lacks AI—it's that Chrome lacks agentic capabilities. Users don't want AI bolted onto a browser; they want AI to replace the browser for certain workflows. They want to say "book me a flight" and have it done, not "use AI to help me book a flight."

The differentiation angle in the knowledge base is correct: browser acquisition is hard because it requires massive effort to get people to switch. Most browsers are Chromium-based, so differentiation is marginal. And AI browser adoption has been minimal so far—despite billions in funding and impressive technology, none of these products have achieved meaningful market penetration outside early adopters.

Dia is enterprise-focused but platform-limited. Comet has traction but faces legal and privacy backlash. Genspark is efficient but unreliable. Atlas has distribution but is still unproven at scale.

The Real Winner (Spoiler: It's Probably Google)

Here's the scenario nobody discusses: Google wins by doing nothing special.

Chrome already has 68% market share. Google can integrate Gemini, add agentic capabilities, and maintain dominance without disrupting its $150 billion+ annual search and ad revenue. The company has the resources, the distribution, and the user base to evolve faster than any startup.

The more likely outcome isn't "AI browser disrupts Chrome"—it's "Chrome becomes an AI browser and nobody notices because it was already installed."

But there's a second scenario worth considering: market bifurcation. Chrome remains dominant for general users (70%+ share). AI browsers carve out 10-15% among professionals, developers, and enterprise segments where agentic capabilities justify switching friction. The market doesn't consolidate around one winner; it segments by use case.

This is the uncomfortable middle ground: AI browsers succeed without disrupting Chrome, creating a sustainable niche rather than a wholesale replacement.

What Actually Matters

If any of these browsers achieve real traction, it won't be because of superior AI. It will be because of:

1. OEM pre-installation (Samsung-Perplexity discussions are the real story here)

2. Enterprise adoption with demonstrated ROI on agentic workflows

3. Regulatory intervention forcing Chrome divestiture (the DOJ case is ongoing)

4. Ecosystem lock-in through integrations that make switching painful in the opposite direction

The technology is table stakes. Distribution is the game.

The Bottom Line

The browser wars are back, but they're not being fought on the battlefield of features or speed. They're being fought on distribution, regulation, and the willingness of users to overcome switching friction for genuinely new capabilities.

AI browsers have a real shot—not at dethroning Chrome, but at capturing meaningful segments where agentic capabilities justify the friction of switching. The market is probably expanding rather than redistributing. Chrome remains dominant, but the future of browsing is genuinely uncertain in ways it hasn't been since 2008.

The irony? The browser that wins won't be the one with the best AI. It'll be the one that makes switching easiest.

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

Hello,

The adult industry is about to enter the OpenAI ecosystem, and it is the biggest deal we can think of.

The Industry That Validates Technology

Here's something most people don't know: when Mistral launched, their biggest token consumers weren't enterprises or developers. They were sex AI applications. Why? Because Mistral's models were largely uncensored, making them perfect for adult content platforms.

The adult industry has always been technology's canary in the coal mine. It's hyper-competitive, forcing constant innovation and differentiation. When adult entertainment adopts a technology, you know it's ready for mainstream, and profitable.

OpenAI's Calculated Risk

By opening the door to adult apps, OpenAI is making several strategic bets:

1. Age Verification as Infrastructure: To allow adult content, OpenAI must solve a massive problem, reliably determining if users are 18+. This isn't just about compliance; it's about building guardrails that have enormous business implications across content moderation, advertising, and legal frameworks.

2. Confidence and Support Use Cases: OpenAI has clearly identified that many ChatGPT interactions are about emotional support and confidence. After the PR disasters of users making poor medical decisions or worse, they're acknowledging these intimate use cases exist, and trying to channel them properly.

3. The Transaction Problem: Adult entertainment lives on conversion. Users must pay, and fast. But OpenAI's app ecosystem currently offers minimal transactional capabilities. How will an industry built on immediate monetization function in a documentation-heavy, transaction-light environment?

Dressed to Prompt

The French Irony

For French tech observers, there's delicious irony here. We're essentially returning to Minitel Rose, the prehistoric technology that launched France's internet economy. The same business model that built Free and made Xavier Niel (and early investor in Mistral) a billionaire is now being reimagined for the AI age.

Sex sells. Always has. Always will.

What This Really Means

OpenAI's move reveals their economic desperation as much as their pragmatism. If adult apps can build viable business models on ChatGPT's platform, it validates the entire ecosystem. If they can't, if the constraints are too tight, the transaction friction too high, it exposes fundamental flaws in OpenAI's app strategy.

The adult industry will stress-test every assumption OpenAI has made about their platform. Age verification, content moderation, payment processing, user privacy, regulatory compliance, all at scale, all under intense competitive pressure.

If OpenAI's apps can satisfy this industry, they can satisfy anyone.

Sex might sell tokens. But can tokens sell sex?

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

Hello,

remember when the first App Stores launched? Those platforms promised unprecedented autonomy for developers and freedom of choice for users.

With ChatGPT "Apps", OpenAI is rewriting the rules, but in one direction only: absolute control. The long-awaited ecosystem looks less like an open marketplace and more like a greenhouse under glass.

ChatGPT "Apps": A Controlled Showcase

Imagine opening a shop in a mall. You're promised prime real estate, but you can't choose your storefront or interior design. Your "shop" is really just a showroom where someone else decides when customers can see you.

That's the reality of "Apps" in ChatGPT. These third-party modules primarily serve up information, though they can also trigger actions like booking rides or ordering food. But here's the catch: the AI model, not the user, decides which app gets activated at any given moment.

OpenAI's Economic Strategy Revealed

This "Apps" rollout exposes OpenAI's long-term play. Power sits entirely in their hands: they decide which applications get presented and, crucially, which ones get called in response to user queries. Ask for plane tickets? OpenAI chooses whether to activate Booking, Skyscanner, or neither.

There's no real "App Store" in the traditional sense. Even if developers register their tools, there's zero guarantee of activation. OpenAI positions itself as gatekeeper, dictating all interactions. (NB: there is a notion of App Directory thought)

This stranglehold on app selection lays the foundation for a business model that could rapidly pivot to advertising. Like web giants and existing app stores, OpenAI will likely monetize visibility. Developers could soon pay to "rank higher" in suggested results or secure preferential activation, transforming what presents itself as a service into a new, tightly controlled ad channel.

Severe Constraints on User Experience

Beyond opaque selection, developer constraints are numerous. Technical documentation offers little transparency about how apps get called. UX limitations: color codes, interfaces, are draconian, leaving minimal room for customization or differentiation.

ChatGPT "Apps" aren't complete applications but rather "app fragments", bits of functionality reduced to their simplest expression

You can have any color, as long as it is GPTblack.

Beyond the Walled Garden: Total Dependency?

We're moving past the "walled garden" concept where platforms set the rules. With ChatGPT, it's a private party where OpenAI is the exclusive host. They decide when you get to speak, who you speak to, and how you speak.

This centralization raises legitimate concerns. While some will see opportunity in reaching ChatGPT's massive user base, they must recognize the total dependency they're creating. It's a relationship where third parties serve OpenAI's interests, not the other way around.

What This Means for Developers and Users

The ChatGPT "Apps" promise is tempting, but the reality is rigid control. For developers, this means careful navigation in an ecosystem where innovation bows to OpenAI's decisions. For users, it means less choice and more algorithmic curation.

The stakes are high: how do you preserve value and autonomy when visibility and interaction are so tightly controlled? As AI becomes more central to how we work and create, these questions of platform power become increasingly urgent.

The app ecosystem we're building today will shape the AI landscape for years to come. Whether it becomes an open, competitive marketplace or a series of controlled fiefdoms remains to be seen.

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

Hello,

Following up on our discussions about AI as a collaborative partner, I want to shift gears this week and talk about a critical, yet often overlooked, aspect of AI's integration into our lives: the app.

As I reflect on the recent launch of Sora 2, it's clear that the strategic decision to prioritize an app-centric approach is laying the groundwork for the next generation of social interaction and AI adoption.

The Strategic Playbook: "There's an App for That"

The biggest takeaway from Sora's launch isn't just the model's capabilities, but its distribution strategy. By introducing an invite-only iOS app, initially restricted to the US and Canada, OpenAI is consciously building user "captivity" and fostering loyalty. In a world where AI models can quickly replicate each other's core functionalities, the real battle is for user habits and sustained engagement.

This move leverages the novelty factor, creating a kind of "Ghibli Effect" (or Giphy Effect, as some might say), a push for users to download and engage directly within their proprietary ecosystem.

Beyond the Tool: Building a Social Fabric

What's particularly fascinating is how OpenAI is borrowing from the social media playbook, notably TikTok's. The inclusion of watermarks on generated videos and the emphasis on portrait format signals a clear intention: this isn't just a video creation tool; it's a social network in the making. Instead of being "network first," Sora is "tool creation first."

This means attribution and brand building are happening from day one, positioning the app to potentially evolve into a new kind of social platform where creative expression is paramount.

Why Apps Are Key to AI Loyalty

This trend of major frontier models understanding the need for recurrence and habit creation will only accelerate. And there's no better mechanism for cementing user loyalty and daily habits than a dedicated app. Think about it:

1. Notifications: Keeping users engaged with timely alerts.

2. Home Screen Presence: A constant visual reminder, easily accessible.

3. Daily Habits: Integrating into routines, becoming a go-to for specific tasks.

We can anticipate features like discovery, sharing, and remixing becoming central to these AI-powered apps. We've already seen this strategy with platforms like Suno, where music distribution was baked into the app experience from the outset.

Final thoughts

The future of AI isn't just about impressive models; it's about how these models seamlessly integrate into our daily workflows and social interactions. By focusing on app-based ecosystems, AI providers are not just offering tools, but cultivating communities and habits. This signals a shift where the AI itself becomes the platform, driving a new era of "tool-creation first" social networks.

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

Hello,

This week, I want to share some fascinating insights from OpenAI's massive study on how people actually use ChatGPT. The findings challenge the dominant narrative about AI replacing workers and reveal something far more nuanced: AI as a collaborative partner.

The Real Numbers Behind AI Usage

OpenAI just released comprehensive data analyzing over 2.5 billion daily messages from 700 million users, nearly 10% of the world's adult population.

Here's what caught my attention: only 27% of ChatGPT usage is work-related, and that percentage has been declining over time.

But here's the kicker, when people do use AI for work, they're not asking it to replace them. They're asking it to help them think better.

AI as Your Research Assistant, Not Your Replacement

The study reveals three dominant use cases that account for nearly 80% of all interactions:

1. Practical Guidance (29%) - "How should I approach this problem?"

2. Seeking Information (24%) - "What do I need to know about X?"

3. Writing (24%) - "Help me improve this draft"

Notice the pattern? These aren't "do my job for me" requests. They're "help me do my job better" conversations.

The researchers introduced a brilliant framework: Asking vs. Doing vs. Expressing. About 49% of all messages are "Asking", seeking advice and information to make better decisions.

“Only” 40% are "Doing", requesting actual task completion.

Even more telling: Asking messages consistently receive higher satisfaction ratings and are growing faster than Doing messages.

The Decision Support Revolution

What we're witnessing isn't job automation, it's decision augmentation. The study found that AI usage maps heavily to two core work activities:

• Obtaining, documenting, and interpreting information

• Making decisions, giving advice, solving problems, and thinking creatively

This pattern holds across virtually every profession studied, from management to engineering to healthcare.

The Hidden Complexity Behind the Magic

The most intriguing finding? Writing dominates work usage, accounting for 42% of work-related messages. But here's the twist: about two-thirds of writing requests aren't asking AI to create from scratch, they're asking it to modify existing text.

This mirrors what we see at Alter. The real magic happens when humans and AI collaborate: humans provide context, judgment, and creativity; AI provides speed, scale, and refinement of the rough edges.

Final thoughts

Perhaps the most encouraging finding: AI excels at the collaborative aspects of work that humans value most. It's becoming the research assistant who never gets tired, the writing partner who's always available, and the brainstorming buddy who brings fresh perspectives.

The future isn't human vs. machine, it's human with machine. And the data shows that future is already here.

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

Hello,

This week, I want to unpack the buzz around Claude’s new ability to create and edit native Excel, PowerPoint, Word, and PDF files, all within a chat.

On the surface, it looks like magic. Just tell Claude what you need, upload your data, and get polished files back ready to use.

But here’s the catch: behind this “magic” lies a sophisticated engineering trick, Claude isn’t directly manipulating documents like a human would. Instead, it writes and executes code in a server-side sandboxed environment using Python and Node.js libraries to build or modify files.

The hidden complexity and the tension it creates

This architectural choice offers powerful capabilities that transform hours of work into minutes of conversation. However, it also raises important questions for AI developers, users, and the industry as a whole:

• Transparency vs. Abstraction:
  Should AI tools fully reveal these “behind the scenes” mechanisms to users? Hiding them offers simplicity and smooth experiences but risks being perceived as misleading or opaque.

• User control and understanding:
  Without insight into the intermediate code execution, users have limited ability to troubleshoot unexpected results or customize workflows at a granular level.

• Security and responsibility:
  The sandbox has limited internet access to download packages, which introduces risks of injection attacks or data leaks. The responsibility falls on developers to communicate these risks clearly and design safeguards , and on users to monitor usage carefully.

Deception or responsible abstraction?

Is the lack of detailed visibility a form of deception¹ ? I’d argue it’s a trade-off, creating manageable, scalable user experiences while wrestling with substantial engineering complexity and potential security threats.

As AI capabilities grow, so does the responsibility of AI developers to balance these tensions thoughtfully:
to be as transparent as possible without overwhelming users, to provide control without confusion, and to build trust without oversimplification.

Final thoughts

Claude’s impressive new feature marks a milestone toward AI truly doing “real work”, but it’s no magic wand. It’s an elegant hybrid of code-driven automation wrapped in conversational ease.

The heart of the challenge lies not just in innovation, but in communication, transparency, and ethics. How much should users see? How much should they control? How do we avoid the illusion of magic slipping toward the shadow of misunderstanding or unintentional deception?

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

OpenAI recently opened a position for a Content Strategist at ChatGPT.com, something that raised eyebrows, even amusement.

After all, isn’t AI supposed to replace roles like this? If AI’s promise is to automate content strategy, why is OpenAI hiring a human expert?

This contradiction reveals a deep tension in the AI conversation, in marketing, positioning, and expectations.

The Common Misconception: AI as a Full Replacement

It’s tempting to say AI will soon do all your work, making some roles obsolete. The idea of an AI autonomously running content strategy sounds futuristic and even scary. But those who use AI daily experience a different truth:

AI is a powerful accelerator and augmenter, not a replacement.

At Alter, we’ve integrated AI at the very core of our workflows. The real magic happens when humans and AI collaborate, humans provide context, judgment, and creativity; AI provides speed, scale, and automation of repetitive parts.

Why Humans Still Matter

The new Content Strategist role at OpenAI showcases what’s essential in the post-AI job landscape:

1. Human in the loop: You need experts to oversee AI output, make final calls, and ensure responsibility. AI can assist but can’t replace nuanced decision-making and ethical judgement.

2. Opportunity spotting: Humans are best at sensing opportunities and shaping strategic directions where AI cannot independently foresee impact or cultural nuance.

3. Voice and tone mastery: Crafting voice that resonates globally requires empathy and insight beyond pure data or automation.

4. Complex workflows: AI helps on sequences of the workflow, writing drafts, SEO, content experimentation, but expert guidance integrates these tasks seamlessly.

From Replacement to Augmentation

This hiring move by OpenAI invites us to rethink the narrative. Instead of rushing to call it proof that “AI doesn’t work,” it’s better to deconstruct what the human role truly is and how AI supplements it.

The future of work with AI is not AI or human. It’s AI plus human.

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

Remember when GPT-4 first dropped and suddenly everyone was having conversations with AI that felt... different? Not just better than GPT-3, but fundamentally more capable, more reliable, more there? We just had that moment again, except this time it's happening with images.

Google's Gemini 2.5 Flash Image, which was dominating benchmarks under the mysterious codename "nano-banana," isn't just another incremental improvement in AI image generation. It's the moment when AI image editing crossed from "impressive tech demo" to "holy shit, this actually works."

Why This Feels Different

I've been testing AI image tools since DALL-E first blew our minds with surreal dogs wearing spacesuits. But there's something fundamentally different about what Google just released. Four things make this feel like a watershed moment:

It's stupidly simple to use. No more crafting the perfect prompt like you're casting a magic spell. You can literally tell it "make this person wearing a red shirt instead of blue" and it just... does it. Correctly. The first time.

The adherence is borderline uncanny. Ask for a specific change and Gemini actually follows your instructions instead of giving you something vaguely related. Want to change someone's hair color while keeping everything else identical? It understands the assignment.

It's fast enough to feel conversational. We're talking 2-3 seconds per generation. Fast enough that you can iterate, experiment, and refine in real-time rather than waiting around wondering if this attempt will be the one.

It's practically free. At $0.039 per image, you could generate a thousand professional-quality edits for less than $40. Compare that to hiring a photographer or graphic designer.

One of the most impressive demo so far - Generate an isometric asset from a night picture

Images Hit Different Than Text

Here's the thing about text AI that we've gotten used to: there's always been a subtle uncanny valley. You can usually spot ChatGPT's writing, it's a bit too polished, slightly generic, missing that human spark. It's incredibly useful, but rarely passes for genuinely human. (NB: I love when LLMs are so meta)

Images don't have that luxury. When you look at a photo, your brain makes an instant judgment: real or fake? And with Gemini 2.5 Flash Image, that judgment is getting a lot harder to make.

Unlike text, which you process sequentially and analytically, images hit you viscerally and immediately. There's no gradual realization that something feels "AI-generated", it either looks real or it doesn't. And increasingly, it looks real.

The Manipulation Problem Just Got Real

This is where things get complicated. When you're generating images from scratch, "create a photo of a cat wearing sunglasses", it's obvious what's happening.

But when you're editing existing photos? The line between "enhanced" and "fabricated" becomes incredibly blurry.

I spent an afternoon editing real photos with Gemini 2.5 Flash Image. Changing someone's outfit, adjusting their expression, moving them to different backgrounds. The results were seamless enough that even I started losing track of which versions were real.

This isn't theoretical. We're already seeing perfect deepfakes created by people with zero technical expertise. When the tools are this good and this accessible, every photo becomes potentially suspect.

Watermarks vs. Reality

Google's solution? They're embedding invisible watermarks (called SynthID) in every generated image. It's technically impressive, the watermark survives compression, cropping, and minor edits. But here's the problem: it only works if everyone uses Google's tool.

What happens when the inevitable open-source clones emerge? When someone builds a version without watermarking? When bad actors specifically seek out unwatermarked alternatives?

Watermarking feels like bringing a policy solution to a technical arms race. It's a good start, but hardly sufficient for what's coming.

Platform Responsibility in the Age of Perfect Fakes

The bigger question isn't technical, it's social. Who's responsible for identifying and labeling AI-generated content?

Should platforms like Twitter, Instagram, and Facebook be required to detect and flag AI images? Should they ban them entirely? Should creators be legally required to disclose when content is AI-generated?

The technology is advancing faster than our social systems can adapt. While we're still debating disclosure requirements, millions of people are already creating perfect fakes that spread faster than fact-checkers can keep up.

Texture transfer is also a cool feature from Nano Banana (don’t create a honeypot with it!)

Living in the Post-Photography Era

Whether we're ready or not, we've entered the post-photography era. The question isn't whether AI will be able to create convincing fake images, it already can. The question is how we adapt our media literacy, our legal frameworks, and our social norms to a world where seeing is no longer believing.

Google's nano-banana moment is remarkable not just for what it enables, but for what it forces us to confront. We've crossed a threshold where the technology has outpaced our ability to manage its implications.

The genie isn't going back in the bottle. The question now is whether we can build the social and technical infrastructure to live responsibly with perfect fake images, or whether we'll let the chaos unfold while we figure it out.

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

GPT-5 launch perfectly captured the current moment in AI: sky-high expectations crashing into underwhelming reality. Sam Altman hyped it as something he'd "never want to go back from," posting Death Star images on social media. The anticipation felt like Christmas Eve for tech enthusiasts.

Then came the letdown. Users found GPT-5 insisting there were three "b's" in "blueberry" and mislabeling U.S. states with fictional names like "New Jefst" and "Miroinia." The model that promised PhD-level intelligence couldn't count letters or identify basic geography.

The backlash was swift. Gary Marcus called it "overdue, overhyped and underwhelming." Reddit users branded it "hot garbage." Even OpenAI had to quickly restore the previous GPT-4o model after user complaints about the new version's cold, robotic tone.

Behind the Curtain: Altman's Candid Moment

In a rare extended dinner interview, Altman admitted: "I think we totally screwed up some things on the rollout." Yet he also revealed the contradictions at AI's heart—while critics panned GPT-5, OpenAI's API traffic doubled and they're "out of GPUs."

Altman also confirmed we're in an AI bubble, comparing it to the dot-com era: "Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes. Is AI the most important thing to happen in a very long time? My opinion is also yes."

The Economics Don't Add Up

The numbers tell a sobering story. OpenAI raised $8.3 billion by August—suggesting a burn rate that dwarfs any startup in history. At a $300 billion valuation, who could possibly acquire them?

Meanwhile, the supposed AI success stories are crumbling:

• Cursor, the AI coding darling, had to degrade its service and introduce a $200/month tier after Anthropic raised prices

• Character.ai and Windsurf weren't really "acquired"—their founders got paid billions while employees got nothing

• No major AI company has gone public or found a real exit strategy

The pattern is clear: these companies can't make money at scale, can't be acquired at their inflated valuations, and can't go public without revealing their unsustainable economics.

But don’t worry, Anthropic new round is “too popular”.

The Infrastructure Mirage

Perhaps most concerning is how AI capex has become a pillar of U.S. economic growth. According to analyst Paul Kedrosky, AI infrastructure spending now represents 1.2% of GDP and accounts for more than half of recent economic growth.

But this "boom" is really just four companies—Microsoft, Google, Meta, and Amazon—spending hundreds of billions on data centers for products that don't generate meaningful revenue. It's a private stimulus program built on hope rather than returns.

When this CAPEX spending inevitably slows (you can't build data centers forever), what happens to the economy that's become dependent on it?

The Scaling Law's End

The technical reality is equally sobering. The 2020 "scaling laws" paper that launched the current AI frenzy suggested models would keep improving as they got bigger. But that's clearly breaking down.

OpenAI's internal "Orion" model disappointed, leading to GPT-5's focus on "post-training improvements", essentially souping up existing models rather than building fundamentally better ones. It's like the difference between building a sports car versus adding racing stripes to a sedan.

The Bittersweet Reality

We're witnessing AI's bittersweet symphony: genuine technological progress shadowed by unsustainable economics, revolutionary potential undermined by fundamental limitations, and transformative promises colliding with disappointing reality.

The technology isn't worthless, GPT-5 does show improvements in coding and reduced hallucinations. But it's not the paradigm shift that justifies the hundreds of billions invested or the economic disruption it's causing.

As one observer noted, we're building "short-lived, asset-intensive facilities riding declining-cost technology curves", not century-spanning infrastructure like railroads, but rapidly depreciating assets that require constant replacement.

The symphony plays on, but the crescendo may be behind us. The question isn't whether AI will matter, but whether the current bubble can find a sustainable rhythm before the music stops.

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

The GPT-5 launch was supposed to be OpenAI's victory lap.

Instead, it became a masterclass in how the relentless pursuit of hype can collide spectacularly with user reality.

When Sam Altman faced a Reddit firing squad demanding the return of "inferior" models, he wasn't just dealing with user complaints, he was confronting the fundamental contradiction at the heart of the AI industry.

The contradiction is this: AI companies need to feed an endless hype cycle to justify massive fundraising and capital expenditure, but users have reached a plateau where "better" isn't actually better for them.

The Hype Imperative

Consider the economics driving AI development. Training GPT-5 likely cost hundreds of millions of dollars. Michelle Pokrass from OpenAI's research team admitted they "would have loved to get longer context up to 1M in GPT-5, partly because of compute cost we couldn't yet."

These astronomical costs create an imperative: each new model must be positioned as revolutionary to justify the investment.

The hype machine must keep churning, the benchmarks must keep improving, and the narrative of exponential progress must be maintained, regardless of what users actually want or need.

But what happens when that narrative crashes into reality?

The Presentation Disaster

The answer came in OpenAI's GPT-5 presentation, where misleading bar charts became a symbol of the industry's desperation. When users called out the deceptive graphs, Altman's response was telling: "the numbers here were accurate but we screwed up the bar chart / presentation. we should never have shipped that slide."

This wasn't just a presentation error, it was a symptom of an industry so focused on maintaining the illusion of progress that it's willing to manipulate visual representations of data. The charts showed marginal improvements dressed up as dramatic leaps, revealing how thin the actual advances have become.

The User Revolt

Meanwhile, users were experiencing something entirely different. As one noted: "I've got my workflow, I've built my agents... why should I pay more for features I don't need?" Another captured the frustration perfectly: "GPT 5 is such a step backward for users like me... it feels like cost cutting measures or an attempt to make the product more appealing to the general public at the expense of users who value depth and nuance."

The pattern is clear: users have invested significant time and resources building workflows around existing models, and "improvements" that break these workflows aren't improvements at all—they're disruptions.

The Consistency Economy vs. The Hype Economy

This creates a fundamental tension between two different economic models:

The Hype Economy (what AI companies need):

• Continuous narrative of revolutionary progress

• Justification for massive capital raises

• Premium pricing for "cutting-edge" capabilities

• Media cycles built around breakthrough announcements

The Consistency Economy (what users actually want):

• Reliable, predictable behavior

• Preservation of existing workflows

• Cost stability over feature additions

• Long-term support for tools that work

Christina Kim from OpenAI's research team revealed this tension when she explained that "we've made a dedicated effort with gpt-5 to train our model to be more neutral by default." Users didn't want neutral, they wanted the personality and quirks they'd grown accustomed to. But "neutral" sounds better in investor presentations.

The Capex Trap

The economics are becoming unsustainable. AI companies are caught in what could be called the "Capex Trap":

1. Massive upfront investments in training and infrastructure

2. Marginal improvements that don't justify premium pricing

3. User resistance to changes that break existing workflows

4. Competitive pressure to keep investing despite diminishing returns

Meanwhile, users are discovering that older, "inferior" models often work better for their specific needs. The result? Companies spending billions to build products their customers actively don't want.

The Funding Reality Check

This dynamic is particularly dangerous given the current funding environment. AI companies have raised unprecedented amounts of capital based on promises of continuous exponential improvement. But what happens when that improvement curve flattens and users prefer the older versions?

The GPT-5 backlash suggests we're approaching that inflection point. When users actively demand access to "inferior" models, it signals that the hype cycle has disconnected from actual value creation.

The Coming Reckoning

The industry is facing a choice: continue feeding the hype machine with increasingly marginal improvements, or pivot to what users actually value: reliability, consistency, and cost-effectiveness.

Some signs suggest the latter is already happening. Altman's promise to bring back GPT-4o "for plus users" and his admission that "we should have something unlimited!" indicate that OpenAI is beginning to recognize the mismatch between their development priorities and user needs.

¹

What This Means

We may be witnessing the end of the AI hype cycle and the beginning of a more mature market where:

• User retention matters more than benchmark improvements

• Operational efficiency trumps cutting-edge capabilities

• Sustainable business models replace venture-funded growth-at-all-costs

• Incremental improvements are valued over revolutionary claims

The companies that survive this transition won't be those with the most impressive demos or the biggest funding rounds. They'll be those that can deliver consistent value at sustainable prices, even if that means admitting that sometimes, the older version really was better.

The age of AI hype is ending.

The age of AI utility, with all its boring, profitable realities, has begun. 

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

From Free Web to Paid Access

Tim Berners-Lee's original vision for the World Wide Web was radical: a decentralized network where information flowed freely, accessible to anyone with a connection. Cloudflare's new "pay per crawl" system fundamentally challenges this principle by turning webpages into potential tollbooths.

We've moved from "information wants to be free" to "information wants to be compensated."

Rethinking Web Neutrality

Traditional web neutrality meant treating all internet traffic equally. But Cloudflare's system introduces content neutrality: should human browsing cost the same as machine crawling?

Human browsing is fundamentally different from machine consumption. A person reads one article, maybe shares it, possibly makes a purchase. An AI crawler systematically harvests everything to build trillion-dollar models. The economic value extracted per visit isn't comparable.

Budget-Based Internet Architecture

Cloudflare's pay-per-crawl system mirrors token budgets in AI reasoning models. OpenAI's o3 has computational budgets allocated per task. Now websites can have access budgets.

This creates economic signals in digital consumption. When your AI agent has a $10 research budget, it forces prioritization decisions. Which sources are worth paying for? Which information commands premium pricing versus commodity rates?

The Subsidy Question

The critical question: Will AI companies actually pay these fees, or will Big Tech's advertising subsidies absorb the costs?

Google can afford to pay website owners because they monetize that data through advertising. The current AI boom is heavily subsidized by venture capital and tech giant cash reserves. When that funding model changes, we might see two distinct internets emerge:

1. Premium Web: High-quality, human-curated content behind paywalls

2. Commodity Web: Ad-supported, freely crawlable content

The Advertising Model Problem

This creates challenges for advertising-based business models. How do you show ads to an AI agent? When GPT-5 will research "best running shoes" for a user, does it see sponsored content? Does it disclose paid placements?

The advertising ecosystem assumes human psychology and behavior. AI agents don't impulse-buy or respond to brand awareness campaigns. They optimize for user objectives, not advertiser objectives.

The advertising subsidy model breaks down in an agentic world.

Market Maturation

What we're seeing is the internet economy maturing. The "free" web was never actually free, it was subsidized by advertising, data harvesting, and venture capital. We're moving toward direct value exchange.

Content creators can charge based on actual usage rather than hoping for ad revenue. AI companies access premium content by paying market rates. Users get better AI responses because models train on higher-quality, compensated data.

This forces both content creators and AI companies to optimize for value rather than volume. Better content gets rewarded. Better AI gets access to superior training data.

The web isn't dying, it's finally getting a sustainable economic model.

Modern Chaos explores the intersection of technology, business, and society in an age of rapid transformation.

P.S. If this analysis resonates with you, forward it to a colleague who needs to hear it. We're building a community of leaders who think critically about AI implementation.

Reply and share your story—I read every response and feature the best insights in future newsletters.

Until next Thursday 🎉

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

(This issue was inspired by Corridor Crew's latest episode)

We're witnessing a familiar pattern emerge in the entertainment industry, one that perfectly mirrors what happened with CGI over the past 15 years.

Hollywood spent years marketing "practical effects" as superior to digital ones, creating a false binary between "authentic handmade" and "artificial computer-generated" content.

Now, as AI enters creative workflows, we're seeing the exact same marketing narrative resurface. And it's just as misguided.

The Automation Overpromise and Coming Backlash

We're currently in the peak of AI overpromising, with companies pushing visions of fully automated creative processes that will supposedly replace human artists entirely.

This "full automation" narrative is setting the stage for an inevitable backlash. As these overpromises fail to deliver the promised results and quality, we'll see a predictable swing toward "No AI was used" badges and marketing campaigns.

The industry is essentially creating its own rejection cycle by overselling what current AI can realistically accomplish in creative contexts, rather than focusing on the more practical reality of AI as a powerful tool within human-guided workflows.

The False Binary Returns

The entertainment industry has always thrived on technological innovation, but it loves to pretend otherwise. For years, studios have promoted the idea that practical effects are inherently superior to digital ones, that "handmade" equals better quality, more honesty, more authenticity.

Some stunt person trying to keep up with the flow of AI news.

This narrative ignores a fundamental truth: the best productions have always used hybrid approaches.

Practical effects provide crucial reference points: physics benchmarks, lighting guides, tangible elements for actors to interact with.

This real-world anchoring becomes essential for VFX artists, even when the final product is heavily digitized. The magic happens in the alliance between practical and digital, not in the purity of either approach.

AI as the New Scapegoat

As AI tools enter creative pipelines, we're already seeing the emergence of "AI-free" marketing badges. This is the beginning of a predictable cycle:

1. New technology emerges in creative industries

2. Early adopters integrate it strategically

3. Marketing narratives emerge promoting "traditional" methods

4. False binaries are created (human vs. machine)

5. Reality settles into hybrid workflows

6. Technology becomes invisible infrastructure

Claiming something is "made without AI" is becoming as meaningless as saying you wrote a text without spell-check. The tools are already embedded in our workflows, often invisibly.

The Infrastructure Transformation

AI is rapidly becoming like GPUs, RAM, or cloud computing: invisible infrastructure that users don't think about.

When you use photo editing software, you don't question whether it uses GPU acceleration. When you send an email, you don't worry about the distributed servers processing it.

You care about functionality and results, not the underlying technology stack.

The same invisibility is happening with AI in creative work. What matters isn't whether AI was used, but whether the final output serves its purpose effectively.

Beyond the Purity Wars

While purists will emerge on both sides, those who refuse any AI assistance and those who go fully automated, the most effective approach will always be the hybrid model. The future belongs to human-augmented AI and AI-augmented humans.

Successful creative workflows integrate:

• Human creativity and judgment

• AI processing and efficiency gains

• Traditional techniques where appropriate

• Quality validation and refinement

The goal isn't purity, it's optimal results through intelligent tool selection.

The passion of Joan d’AI-rc

The Coming Marketing Divide

We're about to see a new wave of marketing positioning around AI usage. Some productions will tout their "AI-free" credentials, while others will embrace AI enhancement. But just like the practical vs. digital effects debate, this positioning will become irrelevant as the technology matures.

The market will reward results, not process purity.

The Invisible Revolution

The most successful AI integration will be the kind you don't notice. Just as the best VFX work is invisible to audiences, the best AI-assisted creative work will seamlessly blend human creativity with machine capability.

We're not heading toward a world where AI replaces human creativity, we're heading toward a world where the distinction becomes irrelevant because the integration is so sophisticated.

The age of hybrid creativity has begun. The question isn't whether you'll use AI in your creative process—it's how skillfully you'll integrate it.

Modern Chaos explores the intersection of technology, business, and society in an age of rapid transformation.

P.S. If this analysis resonates with you, forward it to a colleague who needs to hear it. We're building a community of leaders who think critically about AI implementation.

Reply and share your story—I read every response and feature the best insights in future newsletters.

Until next Thursday 🎉

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

Executive Summary

The Windsurf acquisition reveals a profound truth about AI's economic reality: we're witnessing the birth of a new asset class where exceptional human talent commands unprecedented premiums.

While the $2.4B headline grabs attention, the real story exposes how the AI revolution has inverted traditional value creation: transforming brilliant engineers into the ultimate scarce resource and forcing us to rethink everything we know about building defensible businesses in the age of artificial intelligence.

The Deal Anatomy

Let's decode what really happened here.

Microsoft's exclusivity clause basically killed the initial OpenAI acquisition before it could get off the ground.

Google pulled off a reverse acqui-hire: the top team joins DeepMind while they secure a non-exclusive license deal.

Cognition swooped in for the clean-up, grabbing $250M worth of remaining asset.

But the hidden structure is where it gets interesting: a complex package that includes lost equity compensation, performance bonuses, and licensing fees.

But here's what's been eating at me...

The Uncomfortable Truth

Windsurf's core IP isn't worth $2.4B. Not even close. What Google actually bought was

• Ten exceptional brains who can rebuild anything in seven weeks

• A proven ability to execute at lightning speed

• A deep understanding of the AI application layer.

What they didn't buy?

Defensible technology moats, unique model architecture, or irreplaceable IP.

This isn't an acquisition, it's the world's most expensive talent grab disguised as a technology purchase.

The Talent Arbitrage Game

We're watching something pretty wild unfold. Talent is being valued at 10-100x their "rational" worth, execution speed beats IP every single time, and the best teams can clone any AI application in weeks, not years.

The Anthropic → Anysphere → back to Anthropic pipeline? Just another example of this pattern.

We're not investing in companies anymore. We're investing in genetic lottery winners who happen to code really well with AI.

The Three Questions Keeping Me Up at Night

After diving deep into this deal, I keep coming back to three questions that should terrify every AI entrepreneur.

What's Actually Defensible? If the Windsurf team can rebuild their product in seven weeks, what stops anyone else? The brutal answer: almost nothing. The moat isn't technology, it's how fast you learn and execute.

Who Actually Captures the Value? Look at the structure: founders and early employees got rich. Investors? That's... less clear. The $2.4B bet assumes these people can create similar magic inside Google. That's a bet on humans, not code.

What's the Sustainable Model? If you're building an AI application today, how do you avoid becoming just another demo that exists to showcase talent for the next acqui-hire?

My Creative Destruction Playbook

Here's what I'm telling founders who corner me about AI valuations: Stop trying to build defensible IP.

Build learning velocity instead, teams that get 10% better every week.

Focus on creative synthesis, finding unique ways to combine existing models.

Create community gravity where users stick around for the journey, not just the destination. And above all, develop brand authenticity, a voice that can't be replicated.

If You're a Founder: Build for learning speed, not scale. Your moat is evolution velocity. Document your process publicly and turn learning into content gravity. Focus on creative synthesis.
The magic is in uniquely combining existing pieces. Design for humans because AI is the commodity; human insight is the differentiator.

If You're an Investor: Value learning velocity over current metrics. Bet on teams who can rebuild in under eight weeks. Look for creative synthesis, not technical breakthroughs. And accept that you're betting on genetic lottery winners.

The Meta Lesson

The Windsurf deal reveals the ultimate irony of our AI moment: the most valuable thing in the age of AI isn't the AI, it's the human ability to learn faster than the AI improves.

Every founder building AI applications needs to internalize this brutal truth. You're not in the technology business. You're in the meta-learning business. The question isn't "Can we build something defensible?" but "Can we learn faster than anyone can copy us?"

Final Thought

As I wrap this up, I'm struck by how the Windsurf deal captures everything fascinating and terrifying about right now. We're watching the entire concept of "company value" get rewritten in real-time.

The $2.4B isn't for what Windsurf built. It's for what they proved was possible when exceptional humans dance with powerful AI. And honestly? That might be the most valuable insight of all.

Modern Chaos explores the intersection of technology, business, and society in an age of rapid transformation.

P.S. If this analysis resonates with you, forward it to a colleague who needs to hear it. We're building a community of leaders who think critically about AI implementation.

I want to hear from you: What's your biggest AI reality check moment? When did you realize the gap between promise and practice?

Reply and share your story—I read every response and feature the best insights in future newsletters.

Until next Thursday 🎉

Cheers,
Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

In a sea of sleek, minimalist AI startups, Jupi is making waves by drawing inspiration from surrealist artists like René Magritte, Joan Miró, and Marcel Duchamp.

Their homepage and overall branding scream "surreal simplicity", a tagline that plays on words while promising an effortless user experience, an exquisite work from agency HowHow.

The Claude touch

Why Surrealism? A Refreshing Take on AI Branding

Jupi collaborated directly with a creative agency HowHow to craft a look and feel that's worlds apart from the typical AI aesthetic.

Instead of cold tech vibes, they've infused poetry, lightness, and beauty into their B2B identity.

In an industry often criticized for being artificial and uniform, this approach feels like a breath of fresh air. It's a reminder that even in the world of algorithms and data, there's room for artistic flair.

For me, this branding choice isn't just cosmetic, it's a meta-commentary on AI itself.

• Cadavre Exquis and LLMs: Surrealists played games like "exquisite corpse," where participants added words or images randomly to create unexpected art. This mirrors how large language models (LLMs) generate text probabilistically, building on patterns without full context. It's collaborative chaos turned into something meaningful.

• Hallucinations as Beauty: One common critique of AI is its tendency to "hallucinate", producing unreal or incorrect outputs. But surrealism celebrates the unreal: dream-like images that bend reality to create beauty. Jupi's branding flips the script, suggesting that these "errors" can be artistic and valuable, much like a Magritte painting that challenges our perception of what's real.

The discreet charm of the AI bourgeoisie

What Would Surrealists Do with AI Today?

The discussion poses a thought-provoking question: If surrealists were around now, how would they use AI tools? Imagine Marcel Duchamp prompting an AI to redefine "ready-made" art, or Magritte generating impossible scenes that question creation itself. Recent viral videos—like ASMR clips of glass fruits being sliced echo this: AI creates obsessive, non-existent beauty that evokes real emotions.

Jupi's launch prompts us to rethink digital branding. In a fast-paced ecosystem where everything looks the same, embracing difference through art could be the key to standing out. We hope more companies follow suit!

Fair enough, Perplexity's branding has a similar vibe but with a more “vintage sci-fi cover book in a wet dream” angle to it.

Final Thoughts

Jupi's surrealist-inspired branding isn't just clever marketing; it's a philosophical nod to AI's creative potential. As we hurtle toward more advanced models, let's remember to inject some humanity (and whimsy) into the mix.

Modern Chaos explores the intersection of technology, business, and society in an age of rapid transformation.

P.S. If this analysis resonates with you, forward it to a colleague who needs to hear it. We're building a community of leaders who think critically about AI implementation.

Until next Thursday 🎉

Cheers, Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉

We're witnessing a disturbing phenomenon that's infecting the AI and startup ecosystem: the absurd escalation of numbers. A perverse race where attention substitutes for real value.

The $100 Million Syndrome

Sam Altman just dropped another bombshell: Meta allegedly offering $100 million packages to OpenAI employees. A claim as spectacular as it is absurd.

Why this is problematic:

• Pure destabilization technique

• Zero internal economic logic

• Creates impossible-to-manage salary tensions

Picture this: Meta's top levels earn $15-25 million annually. How do you integrate an engineer at $50 million without exploding internal balance? It's like a dysfunctional football locker room.

The Funding Frenzy Flight Forward

This week again: $34 million raised by an Israeli company in seed phase with Index.

The amount impresses, but let's dig deeper: orchestration under the guise of ultra-ambitious promises. Again.

The recurring pattern:

1. Outrageous promises

2. Spectacular numbers

3. Substance... optional

The Obscenity of Contrast

What makes this numbers race particularly indecent? Microsoft simultaneously announces 9,000 layoffs.

On one side, the apology of easy money. On the other, drastic cuts in workforce. This brutal contrast reveals the toxicity of our moment.

The Sorare Precedent

Let's remember Sorare in 2021: poster child of funding absurdity, SoftBank funding round, employees bragging about multi-million packages.

Result today: catastrophic situation, employees left with zero.

Back to Fundamentals

This obsession with numbers betrays deep panic. We've forgotten:

• Model fundamentals

• Usage fundamentals

• Balanced company fundamentals

When you have nothing left to say about your product, you talk numbers. It's the frog trying to be bigger than the ox.

The Path to Sanity

The AI industry needs maturity. Let's prioritize:

• Substance over spectacle

• Real usage over promises

• Internal balance over flashy moves

Your Reality Check

Does this numbers race concern you? Have you observed this drift in your ecosystem?

Share your perspective: when did you realize the gap between promise and reality in AI?

The Bottom Line: We're clearly in an absurd bubble situation where communicating about numbers like this is, for me, panic behavior. We've forgotten the fundamentals.

Modern Chaos explores the intersection of technology, business, and society in an age of rapid transformation.

P.S. If this analysis resonates with you, forward it to a colleague who needs to hear it. We're building a community of leaders who think critically about AI implementation.

Until next Thursday 🎉

Cheers, Olivier

Like this newsletter? Forward it to a friend and have them sign up here.

Until next Thursday 🎉