To build a defensible, enduring business, it is essential to attract, convert, and retain users.
Ideally, customer reliance on your product should deepen over time, making the cost of switching to a competitor prohibitively high.
Consumer-facing AI companies are now grappling with how to achieve this.
At which layer of the stack should they play? Which product mechanic drives retention? How do they becomes the default choice amongst the sea of same-ish alternatives?
While early attention focused on benchmarks and model capabilities, another battle is being faught. The companies that win the general AI assistant race won’t necessarily have the smartest models. They’ll have the smartest context capture and utilization strategies.
Current Personalization is Broken
Before we dive into the stack wars, let’s look at how current personalization attempts are failing at basic relevance filtering. I’ll focus purely on general-purpose personal AI assistants and not vertical products such as Jenni AI for research, Flora for design, or Harvey AI for legal.
You may recall entering lengthy prompts about your identity, current projects, and recent readings, hoping for more relevant and personalized responses. Countless of articles and reddit threads have been created in the pursuit of writing better prompts. We can think of this as thread-level personalization.
Application providers noticed this issue and realized they’d need to build features to help users achieve the goal of deeply contextualized responses without asking the user to spend 20 minutes padding every single prompt. Some examples:
- Claude’s Projects ask the user to build context around part of their life by adding text and files. This approach gives us project-level personalizaton which is user-generated. Users then create threads nested under a project.
- ChatGPT’s Memories are more ambitious and attempt to achieve account-level personalization. Unlike projects, they are system-generated but allow for user modification.
- We’ve also seen the launch of first party data connectors and MCPs to pull context from ‘the outside world’ and run CRUD operations in external tools. This attempts to solve both personalization and making the general assistant the primary interface for work.
In the most part, current AI assistants fail to distinguish between relevant and irrelevant context. Most seem to simply inject memory, project info, or personalization notes into the prompt, sparing users from retyping but not improving true contextual understanding.
Where the Battle is Being Fought
Here’s how different layers of the stack are approaching context capture thus far:
| Layer | Players | Context Scope | Capture Strategy | Benefits | Challenges |
|---|---|---|---|---|---|
| OS (Hardware) | 1. Apple Intelligence 2. Windows Copilot | Complete | The default | • Installed by default • Complete device context • Zero permission friction • Perfect app integration • Cross-app handoffs | • 1984-esque privacy concerns •Regulatory scrutiny • Platform lock-in • No escape hatch |
| Desktop App | Rewind | Deep Local | Explicit privacy focus | • Deep local context • Stated as ‘local data’ • Cross-application awareness | • Setup friction • Not ‘installed by default’ • Performance impact • Single-device limitation • Technical user requirement |
| Browser | Dia | Web-focused | Browsing behavior | • 80% of knowledge work • Cross-platform • Less invasive than OS | • Misses native apps • Web-only context • Browser switching breaks context |
| Search Engine | Query-based | Search behavior | • Existing user intent data • Zero behavior change • Massive query dataset • Universal adoption | • Episodic vs continuou • Search paradigm disruption • One-shot interactions | |
| Web App | Many, such as: 1. Claude 2. Perplexity (now trying to enter Browser category) 3. ChatGPT | Session-limited | 1. Data connectors 2. Memory 3. Projects 4. Spaces | • Low friction to try • Rapid iteration • Cross-platform • Rich conversation context | • Narrow context window • Manual input dependency or connectors • Session-based memory |
The OS layer seems like the obvious winner.
Apple Intelligence and Windows Copilot have unparalleled access to user behavior. Precedent tells us the further down the stack you move, the more value you can capture, Google’s acquisition of Android and subsequent move into hardware exemplifies this.
However, context capture isn’t just about technical capability, we should also consider user acceptance.
Having the most context means nothing if users reject the experience entirely. The OS layer faces a fundamental tension: the same comprehensive access that makes them powerful also makes them feel invasive. There’s an inherent discomfort in an operating system observing everything you do, even if it is ‘stored locally’.
Might Dia Work?
This brings us to a potentially contrarian thesis: the winners might not be the companies with the most advanced context capture, but those who adhere to the MAYA principle: Most Advanced, Yet Acceptable (see Rory Sutherland on Lenny’s podcast).
Consider Dia.
Dia is the second browser product by The Browser Company. The first product, Arc achieved cult status for its novel design and left sidebar layout.
Abandoning Arc, a much loved product seemed odd, but it’s clear users will not pay for a browser, no matter the UX. In contrast, consumers are willing to pay for AI assistants, ChatGPT has over 20 million paid subscribers.
On paper, it should lose to OS-level competitors. It can’t see your Slack desktop app, your Figma designs, or your terminal sessions. But it captures roughly 80% of knowledge work context while feeling like an acceptable level of intrusiveness, at least to me.
The browser layer might be the perfect context capture strategy because it threads the needle between utility and privacy. It’s advanced enough to be genuinely useful (watching your research, writing, and communication) but contained enough that users don’t feel permanently surveilled.
The final benefit is that browsers traverse devices. If successful, Dia will build for mobile and be available across devices.
Web app strategies face the opposite problem.
ChatGPT, Claude, and Perplexity (who are also building a browser) are attempting to build memory and project features to increase switching costs, but they’re fighting an uphill battle against user behavior.
They require constant user input and curation to build useful context. Even with Desktop and iOS apps for cross-device use, they are not yet solving the fundamental problem of only knowing explicitly what happens in app.
A Third Device?
The current assumption, and Rewind’s original thesis is that most work happens on a laptop. As agentic AI gets better, more work will happen away from our screen.
Open AI’s Codex has given us a peek into a world where we can draft PR’s whilst on the go. Recently I used Claude to prototype an interaction via and artifact whilst on the train and shared it with the team.
On mobile, the OS layer advantage becomes even stronger.
Your phone knows your location, your communication patterns, your app usage, your photos, your payments. It’s the ultimate context-rich environment.
Mobile and smartphone combo has been the default device stack for the past ~15 years but it might be ripe for disruption. With OpenAI’s acquisition of Jonny Ive’s io, they’re clearly trying to define a new category of hardware.
It’s unclear if it will be complementary to the smartphone-plus-laptop combo or attempt to replace part of either.
A Long Ways to Go
We’re still early.
One thing I’ve learnt after building in AI for the last 3 years is that things change fast.
Most users haven’t decided how much context they’re comfortable sharing for how much AI utility.
My gut tells me privacy concerns will vanish so long as the utility gain is sufficient. We’ve seen the pattern play out before with resistance to online payments, biometric authentication, and cloud storage.
The winning strategies might be the ones that gradually acclimate users to increasing levels of context sharing. Building the best models is important, but who can build sustainable context capture methods that users will accept and embrace over time.
This parallel battle for context may ultimately determine who wins the general AI assistant war. The companies that crack the code on trusted, comprehensive context capture will build the deepest moats, regardless of whose foundation model they’re running underneath.
Of course, this will all seem quaint when we eventually accept brain-computer interfaces, but until then, the stack wars rage on.