Cognition Unchained, One Year Later: What I Got Right (and Wrong) About How AI Would Change Our Workday
In July 2025, I gave a talk at Chicago Tech Week called Cognition Unchained: How AI Will Change Our Workday. The talk argued that AI might help knowledge workers manage the information that currently overwhelms them, and provided a few frameworks I was using to evaluate both tools and startups.
A year is a long time in the age of AI.
A year later, I’m struck by how much of the talk holds up (and by the ways the market managed to surprise me). In 2025, most of this felt like speculation. Now the contours of this new AI landscape are visible, even if we don’t know what the ultimate destination will look like. Before the reflections, here’s what I said on that stage.
The original talk, in brief
The Hyperactive Hive Mind
I opened with a scene I knew firsthand, and one most knowledge workers recognize. It’s late afternoon. Slack and email are stacked with unread messages. A memo is due today, and a larger LP update is due tomorrow. Meanwhile, you’re trying to reconstruct last week’s team meeting because your AI note-taker failed, and you’re in charge of this week’s agenda.
Productivity writer Cal Newport calls this constant pinging, checking, and updating the hyperactive hive mind. My point was that this way of working is structurally inefficient and, over time, unsustainable. For a long time, the systems we’ve operated in have kept pulling us toward shallow attention and away from deep work.
The Search for a Second Brain
Managing yourself, let alone an organization, has always been hard, and thinkers have disagreed for millennia about how to do it. In Plato's dialogue Phaedrus, Socrates argues that writing would make people forgetful, that offloading memory onto a scroll would weaken the mind instead of sharpening it. Centuries later, philosophers Andy Clark and David Chalmers made the opposite case in their 1998 paper “The Extended Mind”: memory doesn't have to live in your head to count as yours. Their example is a man named Otto who has Alzheimer's and relies on a notebook to navigate the world; they argue his notebook plays the same role his own memory would, so it counts as part of his mind.
As our societies have gotten more complex and our organizations have expanded, we've built systems that put that idea into practice, whether or not we ever called it philosophy: Gantt charts, Cornell notes, mind maps, then Evernote and OneNote, then Trello and Asana, then Notion and Airtable. Part of why we keep building these is a real cognitive limit: psychologist George Miller's famous 1956 finding put short-term memory capacity at about seven items at a time. Retention was never our strength.
That's the premise behind a second brain: offload the storage so your actual cognition stays free for judgment and creativity. AI is the next leap in this story. For the first time, a system doesn't just store what you give it, it can automatically build and maintain it.
Frameworks for AI Tools
Making the case that AI could build and maintain a second brain raises an obvious question: which tools should you trust with that job? I'd been testing dozens and listening to pitches for months, and had begun recognizing some patterns. Here are the frameworks I used to make sense of the volume I was seeing.
Framework 1: All-in-one vs. modular tools.
Every AI tool now sits somewhere on this spectrum. On one end, integrated suites like Google Workspace bundle AI across everything you already use. They're powerful, but slow to adopt and painful to leave once you're in. On the other end, there are modular point solutions like Granola for meeting notes, Perplexity for research, Gamma for slides. Each (ideally) excels at its one job, but using several together pushes the work of stitching them together onto you.
Framework 2: Individual vs. team tools, and the “collective brain.”
Some tools are built to make one person sharper: your calendar, your inbox, a drafting assistant. Others are built to solve a coordination problem for an entire team. The tools that do both are rare, and when they work, you get something bigger than a second brain: a collective brain, shared across a team.
Framework 3: The idea of “calibration.”
AI is shifting from generating content, drafts and summaries, toward shaping decisions. I was starting to hear people ask: does this number mean the quarter went well, or does it mean we're in trouble? The best tools earn your trust for that by doing more than removing work from your plate. They cite sources, explain their reasoning, and know when to act alone, when to loop in a person, and when to just watch. Matching that mode to the task in front of you is a skill both we and AI needed to develop.
Community as an edge
When I gave this talk, OpenAI had just launched Agent Mode and Y Combinator ran through 100+ personal-AI-assistant startups in that year alone. Making sense of all that required comparing notes with other people hitting the same walls, which is part of why I started a monthly community for exactly that.
With that as the baseline, here's how it's held up.
What I got right
All-in-One vs. Modular Stacks
Perplexity and Gamma, two of the modular tools I called out, have both cooled off in terms of popularity, a reminder that being the hottest startup in a given moment doesn't buy you staying power. I did not, however, expect the model providers to continue moving into the all-in-one space.
While not yet a true all-in-one product, they have taken on two roles: a layer that sits on top of a bunch of other apps, connectors reaching into your calendar, your docs, your whole stack, and a platform people build within, like a copilot. (Remember when custom GPTs were all the rage?)
Human <> AI Calibration
Most people calibrate on a philosophical spectrum: AI for everything, AI for nothing, or somewhere in between. But the real dividing line comes down to how ambiguous the task is. There are tasks you could've automated pre-AI, where the output is logic-based, so handing it off to a machine is easy.
Other tasks require judgment, and the more judgment a task demands, the harder oversight is to scale, because judgment calls on business data open to interpretation. A Claude research report on AI-generated dashboards found that without precise documentation of what a metric actually measures, even a human reviewing the output can be fooled into thinking a number is right simply because it looks plausible. You can't verify what you can't independently check, which is probably why judgment-heavy work will be the last terrain AI fully takes over.
The Race for Context
A few years ago, the people chasing a second brain were mostly individuals, a niche community comparing notes and stacking tools to build one for themselves. Now the apps have joined the hunt too. Every new agent, connector, and workspace tool wants to own that context for you, your files, your calendar, the record of how you live and work. That's the interesting turn for consumers: building a second brain used to be something you did yourself. Increasingly, it's something companies are competing to do for you.
What I underestimated
The Value of Portability
As models have gained the ability to do work and save it locally, on your own device, frameworks for personalized context files, ones you can carry between models, are starting to emerge. That choice, whether you own your context or hand it to whichever app is holding it, maps onto a bigger centralized-versus-decentralized debate playing out across AI right now. The centralized bet, the one apps are making in the race for your context, is to let one company's ecosystem hold your files and history for you. The decentralized bet is to own that context yourself, in a harness, the layer that manages your prompts, memory, and context, and treat whichever model is running it as replaceable. Bet decentralized, and the harness ends up mattering more than the model.
Individual vs. Team Leverage
This is the one I'd flag as the most underappreciated nuance.
I talked about individual-contributor tools and team tools as two points on a spectrum, with the rare tool that serves both looking like a “collective brain.” What I didn't appreciate is how lopsided the progress would be, and the portability I previously described is part of the reason why. The same personal context that makes one person dramatically more effective is built around their own files, habits, and shorthand, which is exactly what keeps a room full of those people from working as a team. The frontier right now is turning a team of individually 10x'd people into a 100x team.
Where we’re at
If I were giving this talk today, I'd keep the closing line: “the next leap in productivity won't happen alone, it'll happen together.” I'd just mean something broader by “together” now. In 2025, it mostly meant your network, peers comparing notes on the same walls, in parallel. In 2026, it also means your teammates: figuring out how you can work together to build proprietary intelligence inside your own organization.
This piece revisits a talk I gave at Chicago:Blend in July 2025 (watch it here).
If you're working through similar questions about AI and the future of work, I run monthly sessions on this and you can sign up here.