Axess Connect podcast, with Kris Haynes and Stefan Stickley · September 1, 2025 · Part 2 of 2, ~30 minutes. Listen to the episode.
Part two zeroes in on what AI means for day-to-day work: productivity definitions, job compression, career-progression disruption, build vs. buy, and practical advice. Contains early precursors to the second brain — the “advisory council” GPT project and “Bizarro Brian,” a contrarian custom GPT — and closes with a 12–18 month prediction for computer-using agents inside VDI environments.
Productivity is a loaded word
Productivity means different things to different people: producing more work, the same amount at better quality, or the same amount in less time. It isn’t a decision a company makes once — it can be different for every task, every project, every person. And the conversation about implementing AI has to be bigger than IT: executive leadership, HR, legal, and compliance all belong at the table. From an HR standpoint, if a job used to be half easy tasks and half hard ones, and AI takes all the easy ones, the job left behind is nothing but hard tasks, all day. You can’t just strip out all the relaxing parts and jam every day full of difficulty.
Career progression disruption
AI doing the easy jobs sounds fine in the abstract, except the easy jobs are usually someone’s entry-level position. As AI gets more capable, the tasks it can absorb keep climbing the ladder the same way a career does — task worker, junior manager, senior manager, director, with less task management and more broad business thinking at each step up. What you tell the AI evolves the same way: today it’s “go to LinkedIn, open a browser, click here, click here.” Soon it’s “check this on LinkedIn.” Eventually it’s just “how is my social post doing?” and the AI already knows that means LinkedIn.
Individual contributor amplification
As an individual contributor at Citrix: “I think I do the work of what a team of three or four people would have done five years ago” — a mix of high-brain and low-brain activity, just a lot more of it. That extra capacity leads to more conversations and more directions that wouldn’t have happened otherwise, which can end up creating new hiring rather than eliminating it.
The advisory council and Bizarro Brian
An early, pre-second-brain system: pulling the blogs Brian follows into a ChatGPT project, building a library of the people and ideas he finds interesting, then asking questions against that library — not “was this already answered here,” but “how might these people think about this new thing?” A companion project, Bizarro Brian, is deliberately contrarian — named for Bizarro Superman — built to gut-check a draft blog post: what’s the blind spot, what’s missing, what’s the contrarian read? Brian doesn’t have ChatGPT write the posts themselves, but uses both projects to brainstorm and stress-test his own thinking before publishing.
Build vs. buy
If a build takes three months, it might just ship as an inbox feature in the next model release. Bloomberg spent $10 million building a custom GPT trained on their own data, and the next version of ChatGPT did everything that system did, for free. The advice isn’t “never build” — if building genuinely excites you, do it — it’s don’t build for builder’s sake alone.
Networks of small models
Not every model needs to chase general intelligence. A small model trained only on medical data can outscore ChatGPT on medical exams while running on a Raspberry Pi. Networks of many small, specialized models talking to each other, rather than one enormous general model, is where a lot of the real leverage is heading.
Interface evolution
Text in, text out is the DOS era of this technology. Voice, vision, screen, gestures — whatever interface a person actually wants will exist. Don’t think of these models as fundamentally text-in, text-out systems; that’s an artifact of where the interface happens to be today, not a permanent constraint.
AI as the great equalizer
What AI offers is close to the world’s best coach or chief of staff — the kind of personal support only billionaires with full personal staffs used to have. Now everyone gets access to that same caliber of help, freeing people to do their best work in whatever way is actually fulfilling to them.
The 12–18 month prediction
Computer-using agent benchmarks stood at roughly 45% as of this conversation, up from 17% at the start of the year and just 6% the previous summer, against a human baseline of 72–74%. Within 12 to 18 months, that curve likely closes — not because AI takes every job, but because it becomes capable of doing everything it needs to on a desktop. The concrete prediction: agent identities inside Entra ID, AI agents logging into corporate VDI workspaces and performing tasks without supervision, using agent-to-agent protocols to coordinate — one worker’s agent talking directly to a coworker’s agent, agents picking up tasks straight out of an inbox. Basic by later standards, but squarely inside the next 12 to 18 months.
Key formulations
“I think I do the work of what a team of three or four people would have done five years ago.”
“If you spend three months building something, it might just be an inbox feature in v.next.”
“It’s like the world’s best coach, assistant, helper that billionaires with personal staffs can have already. Now we can all get that.”
“Doesn’t mean AI takes your job — means AI can do everything it needs to within a desktop.”