Field Notes From the Edge

Silicon Valley, 2025

Over the past few weeks, I’ve been deep in Silicon Valley — invited into closed rooms across Nvidia, Netflix, Apple, OpenAI, VCs, start-ups, robotics labs, and beyond.

These are my field notes from the edge: signals, provocations, and practical realities that shape what’s now, next, and coming over the horizon.

If you’re leading a brand, building a product, or simply trying to stay ahead of where the world is going, here’s what you should be preparing for.

NOW - The New Foundations

LLMs Are Becoming Super-Apps — By Design

One note echoed everywhere I went: a significant proportion of 16–25-year-olds are already using OpenAI products as their primary search, problem-solving, or discovery tool.

But this shift isn’t just organic. It’s intentional.

Leaked internal memos from OpenAI describe a vision for ChatGPT as a “Super Assistant”—a single, unified layer mediating search, memory, task execution, and online transactions. This isn’t simply how users happen to use LLMs; it’s how model providers are actively redesigning the internet’s interface.

And they’re not alone. Meta, Google, Perplexity, and others are throwing their hats into the ring, each accelerating the race to build the dominant super-app. Whether it’s Google integrating Gemini deeper into Android, Meta weaving AI into WhatsApp and Instagram, or Perplexity pursuing agent-based shopping assistants, the same strategic goal is clear: own the conversational layer that mediates how we search, shop, communicate, and navigate the digital world.

Instead of opening five apps, users will increasingly engage one conversational layer that routes, curates, and executes on their behalf.

The critical question for every brand: When an AI assistant sits between you and your customer, whose logic, values, and incentives define the relationship?

Because once customers build habits around a super-app, loyalty migrates upstream. Not to you—but to the assistant mediating the entire experience.

LLMs as Identity Layers

Beneath this shift lies another profound transformation: identity.

“Log in with ChatGPT” isn’t science fiction—it’s a logical reality. The same applies to Meta’s moves toward AI-driven identity layers tied to its social graphs and Google’s experiments with personalised Gemini profiles.

Picture a persistent AI profile that carries your interests, shopping history, taste preferences, and personal context across every service and platform. LLMs will become identity brokers, orchestrating hyper-personalized experiences wherever you go. They’ll remember conversations, anticipate needs, and coordinate tasks across brands and channels.

First-party data goes poof.

Instead of belonging to individual brands, customer memory will live inside super-apps. Every insight, every behavioral cue, every loyalty signal risks being swallowed by the assistant layer sitting between you and your audience.

This is the architecture of an AI operating system for daily life. And it raises existential questions for brands:

• Who owns the persistent record of customer context?

• How do you build loyalty when memory and relationship context sit with an intermediary?

• How does your brand show up meaningfully inside someone else’s UI?

The super-app race makes this even more critical. The winner won’t just mediate your customers’ choices—it will own the layer that remembers who your customer is.

The Rise of LLM SEO

If LLMs become the front doors to the internet, discovery changes fundamentally.

We’re moving from keyword search to conversational prompts like:

“Plan me a sustainable dinner party for under £40.”

“Find me running shoes like the ones I bought last spring, but waterproof.”

LLMs, and soon, their super-app incarnations, will decide which brands and products surface in those answers. Welcome to LLM SEO.

To get indexed into these AI knowledge bases, brands will need:

• Machine-readable product data and rich metadata

• Structured content maps (think next-gen sitemaps)

• Semantic descriptions of products, offers, and contexts

Without this, brands risk disappearing from AI-led recommendations altogether.

And as super-app ambitions intensify, the stakes rise further. The same few AI ecosystems may soon control what customers see, buy, and believe. Check out BluefishAI and REFiBuy for more.

Agentic Commerce: Nascent, but Accelerating

Another strong signal from the valley: the future of commerce won’t be humans browsing pages—it will be agents negotiating with agents.

Right now, agentic commerce is embryonic. But it’s coming faster than you think. Perplexity already have an offering, as does Amazon. Google, OpenAI and Microsoft are hot on their heels. It’s months, not years away.

Imagine:

• Personal AI assistants negotiating prices, hunting for deals, or curating shopping lists on your behalf

• Brand-side agents responding dynamically with offers, inventory checks, and tailored bundles

To compete in this world, brands must make their data machine-visible and machine-readable. That means:

• Semantic Search: AI agents will look for meaning, not just keywords. Product catalogs must include metadata, context, and relationships so agents can “understand” your assortment.

• Machine-Readable Sitemaps: Think of them as next-gen sitemaps, designed specifically for agent consumption.

• Dynamic Offers: Systems capable of crafting personalised offers in real time in response to agent queries.

If your brand isn’t machine-readable, it won’t exist in agentic ecosystems. That’s the stark reality ahead.

And as the super-app race heats up, agentic commerce is poised to become one of its core battlegrounds.

Closed Models as Fast Labs — Not Forever Homes

One clear-eyed reality from the valley: even the world’s AI leaders aren’t betting their futures solely on closed platforms.

Instead, they’re taking a far more pragmatic view:

• Use external closed models (like OpenAI, Anthropic, or Google) for fast testing, cheap iterations, and rapid concept validation.

• Once validated, build proprietary models tailored to your data, your brand DNA, and your strategic moat.

It’s not either/or. It’s:

Closed models → fast experimentation

Proprietary models → scalable differentiation

Because while plug-and-play power is seductive, strategic sovereignty is priceless.

Don’t run headfirst into an enterprise solution and shackle your company’s goldmine—your data, your insights, your competitive advantage—to a big provider who will ultimately hold your future in their hands and outside your organisation.

Super-app ecosystems will tighten their grip. And the more you build on top of someone else’s closed model, the more they own:

• Your customer data

• Your behavioural insights

• Your product logic and differentiation

If you can’t detach, pivot, or take your data elsewhere, you’re not a customer—you’re captive.

The choice is simple but existential:

• Experiment fast with closed models.

• But own your core.

Because sovereignty over your data and your models will be the only real power you have left when the super-app layer controls the interface to your customers.

NEXT — The Emerging Blueprint

If the “NOW” is about new interfaces and fast experiments, the NEXT is about designing the systems, culture, and architecture that will either amplify—or strangle—future progress.

This is where Silicon Valley’s best minds were focused: how to scale innovation without losing control.

Mixture of Experts: Max Flex × Max Speed

One of the clearest signals from Nvidia, OpenAI, and even smaller startups was the rise of the Mixture of Experts (MoE) architecture.

Instead of a single, giant LLM handling every task, MoE uses small, specialised models — each tuned for a specific domain (e.g. financial language, retail SKU logic, creative copywriting).

• It’s cheaper: Not every query hits a massive model, reducing compute costs.

• It’s faster: Only relevant experts “activate,” improving speed.

• It’s more precise: Specialists outperform generalists in nuanced tasks.

The big lesson: The future of AI isn’t monolithic. It’s modular, dynamic, and orchestrated like a symphony.

For businesses, this means your tech stack needs to be composable enough to plug in new “experts” without major rewiring. That’s how you keep pace as the frontier keeps shifting.

Citizen Developers: Scale Without Chaos

Every leader I met agreed: AI cannot remain locked inside central tech teams.

Netflix, Nvidia, and even Apple are unleashing non-technical employees to build lightweight tools, automations, and prototypes. They’re doing it safely — not with chaos, but with smart governance:

• Internal APIs and secure sandboxes

• Templates for fast experimentation

• Approval workflows to mitigate risk

The real risk isn’t letting non-tech people build things. The real risk is forcing every idea to stand in line behind an IT backlog.

Citizen development does two powerful things:

1. It speeds up problem-solving at the edges of the business, where frontline teams often spot needs first.

2. It creates cultural ownership of innovation. People feel empowered, not sidelined.

The valley’s message was clear: Build guardrails, not gatekeepers. A bottleneck in this era isn’t just annoying, it’s existential.

Meaning Over Momentum: Make It Matter

Apple’s conversations were a masterclass in restraint.

Their discipline: don’t just be first — be first to make it matter.

• They’ll skip entire product waves if they don’t believe they can elevate the human experience.

• They delay launches until emotional resonance and real-world utility align.

• They avoid flashy demos if they know the tech isn’t yet transformative.

The Apple principle: “Don’t ship novelty. Ship significance.”

For brands, this translates into sharp questions:

• Does this new tool solve a uniquely human problem?

• Does it elevate the experience beyond merely digitising a task?

• Will users feel it’s intuitive — or like tech for tech’s sake?

The valley’s best minds are prioritising meaning over speed. That’s how you build true differentiation.

Organisational Intelligence as Nervous System

A standout insight from Silicon Valley is that data and insights aren’t enough. The advantage is how fast you can act on them.

Three powerful examples:

• Netflix: “Farm for dissent, then commit.” They actively surface disagreement during decision-making, but once a decision is made, the entire team aligns and moves quickly. It prevents groupthink while maintaining velocity.

• Nvidia: Their “Top 5 Signals” process requires every individual to list the five most interesting signals they’ve seen in their domain every two weeks. These signals flow into a shared knowledge pool where leaders scan for patterns, weak signals, and emerging trends. It’s crowdsourced radar for what’s coming next.

• Apple: Loop acts as a living, digital hive mind. Employees across levels share insights, tweaks, and observations from the front lines. It makes even the biggest enterprise feel connected and responsive.

The lesson: Don’t just store knowledge — wire your organisation to act on it.

For the valley’s best, learning is not an afterthought. It’s infrastructure.

FUTURE — What Lies Beyond the Horizon

If NOW is about new interfaces and NEXT is about composable systems and cultural rewiring, then FUTURE is about the deep shifts that will redefine how humans interact with technology itself.

In closed rooms across Silicon Valley, there was a clear signal:

We’re moving from gadgets in our hands… to intelligence all around us.

Robotic Brains Without Robotic Bodies

In several closed-door sessions across Silicon Valley, one trend stood out:

We’re not waiting for robot bodies. We’re racing to build robot brains.

A handful of stealth robotics start-ups are focused on developing highly sophisticated “robotic brain” models. These systems are trained on staggering amounts of multimodal data. The result being AI capable of reasoning across both digital and physical domains.

These robotic brains can:

• Predict how to unscrew a jar lid.

• Understand the mechanics of sweeping under a shelf.

• Optimise repetitive warehouse tasks.

• Learn patterns of joint movement to avoid collisions or errors.

Yet here’s the catch:

The body hasn’t caught up.

Despite these huge strides in cognition, dexterous, multi-use humanoid robots remain a distant reality. Hardware remains expensive and slow to develop. Mechanical precision for fine motor tasks is still far from human capability. And designing machines flexible enough for messy, unpredictable environments like stores or homes is immensely complex.

As one founder put it privately:

“Robots doing human jobs in chaotic real-world settings are further out than headlines suggest.”

The prevailing Silicon Valley thesis is clear:

• Invest heavily in the intelligence layer now.

• Wait for hardware to mature.

• Deploy robotic cognition first in virtual spaces and highly structured physical tasks.

That said, targeted physical breakthroughs are coming — particularly in environments where tasks are repetitive and controlled:

• Industrial arms handling precise, repetitive movements.

• Warehouse robots navigating structured layouts.

• Simple bots scanning shelves for inventory accuracy.

While mass humanoid adoption remains years away, the intelligence layer is progressing fast. Any business dependent on physical operations should watch robotic cognition closely for focused automation breakthroughs that could transform costs, accuracy, and speed in specific use cases.

The Sunsetting of the Phone Era

Meanwhile, another seismic shift is unfolding.

We are witnessing the slow sunset of the smartphone as the centre of digital life.

Apple, Nvidia, and OpenAI are all leaning into an ambient, multimodal future:

• Spatial Computing: Apple’s Vision Pro is just the start — digital information woven into the physical environment.

• Voice & Gesture: Interfaces are moving beyond glass screens to talk, glance, and movement.

• Proactive Agents: Systems that act without waiting for a typed query — anticipating needs, surfacing insights, blurring the line between device and human cognition.

This creates profound questions for brands and businesses:

• How do you remain relevant when there’s no screen to dominate?

• How do you embed yourself in a world where customers speak their desires rather than tap them out?

• What happens when AI intermediaries mediate every interaction between brand and human?

The valley’s insight: “We’re not designing apps anymore. We’re designing presence.”

The companies thinking furthest ahead are already prototyping for this world:

• Commerce will be conversational, not navigational.

• Brand interactions will be proactive, not reactive.

• The frontline of loyalty will be invisible.

In the ambient future, brands will compete not for screen real estate, but for mental bandwidth and emotional resonance.

Where This Leaves Us

Spending the past few weeks inside closed rooms in Silicon Valley made one thing clear:

The future is arriving unevenly—but faster than many think.

From LLMs morphing into superapps…
…to robotic brains learning human dexterity…
…to the quiet sunsetting of the smartphone era…

We’re shifting from an age of standalone tools to systems that anticipate, interpret, and act on our behalf.

And while this piece captures some of what I’ve seen, it’s only a snapshot. Much of what I heard and witnessed, I can’t share publicly. But know this:

This is not hype. It’s real.

I’m not drunk on Silicon Valley Kool-Aid. I’m seeing it with clear eyes. And my conclusion is simple:

Take it seriously.

Be curious about everything.

Bet boldly—but build wisely.

This is the scaffolding of the next decade. Let’s treat it with the urgency, caution, and imagination it deserves.