DUCUG Citrix User Group Conference #28 · Echteld, Netherlands · March 18, 2026
Closing session. 60 minutes. Solo presentation with slides. This was a brand new talk, built from scratch in one day after the December 2025 stump speech became entirely out of date following the November 2025 model inflection.
The talk opened with a personal connection: Brian and his wife at the Milan Winter Olympics, where long-track speed skating was "basically a Dutch event" (orange everywhere, strap-on mullet backs). Then a pivot to the real topic.
Workers know how powerful AI is. Companies report no ROI. 80% of companies see no significant impact. 95% of AI projects fail. "Why is it that we as workers know how powerful AI can be, but the companies are not seeing any return?"
The answer: "most knowledge work is invisible because it's the knowledge part, it's in the name, it's inside our head." The visible outputs (emails, documents, meeting transcripts) are not the actual work. They're transcriptions of the work happening inside our heads.
Enterprise AI targets only the visible 20%: Copilot writing emails, transcribing meetings, chatbots. "I say, yeah, I mean, I guess that's helpful, but so what? If you only perfectly automate this little 20%, okay, you're missing the big part."
Three frontier models dropped in a single month: Opus 4.5, Gemini Pro 3, GPT 5.2. Before November, AI helped with coding but was halting, clumsy, lost context after 15-20 minutes. After November: sustained hours of work, maintained to-do lists, diagnosed and recovered from errors, reliable tool use.
"It's not because the model got suddenly very smart. It was all these little things that were annoying that prevented it from doing the work we wanted to do. Starting in November, they all suddenly were able to do that."
Practitioner quotes on screen: Boris Cherny ("100% of my contributions were written by Claude Code"), Karpathy ("phase change"), Paul Ford ("I spent an entire session of therapy talking about it"), Simon Willison ("It's all changed since November"). Brian's own reaction: "holy fucking shit."
A blog post framing: "What will knowledge work look like in the future? Just look at what coding is like today." Nate B. Jones quotes about software engineering, with coding terms swapped for knowledge work terms.
"Everything that everyone is saying about how AI has completely changed software development, that is coming for knowledge workers. And it is not coming at some distant point in the future, that's coming like yesterday."
The framework for understanding AI's impact on knowledge work. Five layers: worker, cognition (brain), skills, agency, interfaces. The same stack mapped to AI-enhanced work: worker, cognitive extension (AI), skills (text files), agents, interfaces (APIs, MCP, CUA).
The 80/20 line sits in the middle: skills and above are the invisible thinking (80%). Agents and below are the visible doing (20%). Enterprise AI investment concentrates entirely at the bottom.
"I can AI the crap out of this, I can do it perfectly. And I still haven't really changed the way the thinking happens, which is where all the money is. This is just a better Google or better SSH term. This is just a warm-up act. This is not the main event."
"I want to really dig in: what would it take to make AI work like a cognitive extension?" Brian showed his LinkedIn post ("I built a second brain using AI and it's changed the way I work"), then walked through a typical morning with the system.
Opens Claude Desktop, says "let's work on this week's blog post." Claude loads CLAUDE.md (the README file that the system wrote for itself), checks the thinking file, reviews the skills index, finds blog drafts, analyzes the editorial calendar. Full context, zero ramp-up.
"I did not write this file, by the way. It wrote it for itself. Or I should say when I was using Claude and it wrote that file, it wrote it for future versions of itself." And: "Your offspring, grandchild AI is going to be reading these folders someday. Make it really good instructions."
Model portability demonstrated live: "Claude was down the other day, so I changed it to Gemini. Gemini works just as well. Any AI system that can read and save files on a file system is able to work this way."
"Skills are text files written in plain language that just explain to the AI how to do something." They're "sort of almost discovered as opposed to invented." The lobster emoji and claws metaphor: skills connect the brain to the claws.
"If you can write instructions to describe to a colleague or your mother how to do something, you can write a skill."
Real skills demonstrated: connect the dots, daily briefing, knowledge integration. The podcast screenshot skill story: walking down the street, hear something interesting, screenshot the podcast app, AI finds the podcast, downloads transcript, extracts the relevant segment, integrates with knowledge. From idea to working skill in minutes.
Skills are future-proof: "New AI comes out, the skills just happen to get better."
"And what's crazy about this is this is still not about automations. In fact, I hate automations. Everyone wants to talk about automations all the time. We can automate this, we can automate this. My God, how repeatable are your jobs that you're automating all this kind of stuff?"
"There's no part of your job as a knowledge worker where you're like, man, the one thing I need to transform everything is one more automation. I don't care about automations. The reason I don't care about automations is because I don't care about apps."
How AI actually accesses work today. Modern apps through APIs, MCP connectors, token authentication. Web apps through browser control (Claude demonstrated live, controlling Chrome to find Brian's speech time, then spontaneously debugging the event website). Legacy desktop apps through Computer Using Agents (OSWorld benchmark: humans 72%, AI started at 12%, now 75%). And files, already demonstrated through the second brain.
Plus a new category: other AI agents. Shared folders between coworkers' brains, the A2A protocol. "There are multiple employees at Citrix who are working this way today. This is not science fiction."
"Legacy applications is like all of them. It sounds crazy, but since I started, I don't use my to-do list application anymore. I don't use Excel. I don't use Word." Word is only used for sending to colleagues not yet on second brains: "an old compatibility layer."
"I don't even use Teams or email" for colleagues working this way. Wired brains together via Git submodules and shared folders. Voice interface through open-source transcription, walking around the office dictating.
"I'm using Apple Keynote today. On the train over here I'm like, I kind of feel like I want to have AI code me a way to make presentations directly out of the brain."
"4 minutes and 27 seconds later" Brian asked his AI to build a token consumption dashboard. It made a website.
Real usage data: January 29 to February 18. 285 million tokens. 96 sessions. 10,000 messages. Pre-second brain: 100,000 tokens/day. Post-second brain: 5-10 million tokens/day.
Subscription escalation: 20 euros to 100 euros ("hit limits in two weeks") to 200 euros ("hit limits at 7pm, fuck it"). At corporate rates, his monthly usage would cost $954.
"This dashboard is why you don't need apps anymore. Whatever you need from the AI system, just ask it. Whatever format you want to see it in, just ask it."
"When we all start using AI as a true cognitive extension over the next three to 24 months, we're all going to use 50 to 100x more tokens. All data center capacity, GPUs, high-bandwidth memory, and power capacity for the next four years has been pre-sold. And by the way, it's not pre-sold to you."
"Not all tokens are created equal." Different quality tiers, different costs. "It's not a cost issue, it's like there are only so many tokens that exist. Tokens are a fundamental unit of knowledge. More tokens equals more intelligence."
Every request must be evaluated on complexity (expensive model or cheap?), sensitivity (cloud, on-prem, or device NPU?), and nature (company pays or personal subscription?).
The Excel routing example: six options ranging from 200K expensive tokens (CUA operating Excel) down to zero tokens (just open the app: "You are so lazy. I'm not burning 500,000 tokens to hit a button for you").
"Something needs to be making these decisions in essentially real time, for every application of every single worker in your entire organization, thousands of workers, hundreds of times a day."
Who decides? Not the worker (they don't care about cost). Not Microsoft (selling tokens, holding data). Not the AI labs (consuming tokens). "You need someone sitting in your workspace who's a neutral party."
"You want to look at the future of end-user computing? You want to look at the future of Citrix? It's this." Citrix for 35 years: connecting workers to their work. The application type changed, the worker type changed, the device type changed, but the fundamental connection is what they've done for 35 years.
Token management as competitive advantage: "The company that spends the most tokens in the most smart way is going to win. If you're spending 5 trillion tokens and your competitor is spending 6 trillion, they will win."
"If Citrix tells me, we're gonna give you 5 million tokens a day. I'm like, okay great, which half of my job do you want me to not do?"
The talk ended with a giveaway: a starter prompt for building your own AI second brain.
Build your own AI second brain (GitHub gist)