A hundred days ago I built a machine that sounds exactly like me and asked it to read me the AI news every morning. Mostly to see if I could.
It could. It's done it every single day since, and somewhere around week three I stopped noticing that was strange and started noticing something else: I hadn't built a podcast. I'd built a logbook that happens to talk back in my own voice.
A hundred mornings in, I did something I hadn't done once in the whole run — I went back and read the thing start to finish, instead of living through it one day at a time. And that's the only way any of this became visible, because none of it was visible from inside a single morning. The machine writing Wednesday's script has no idea what it told me on Tuesday. I'm the only part of this arrangement with any memory at all.
Here's what a hundred days of that actually taught me.
Everyone kept asking who was winning. Wrong question.
In April, the story was still lab versus lab. Anthropic was quietly sitting on something called Mythos — a model so good at finding cybersecurity flaws that they wouldn't release it, "too dangerous for the public," while a Guardian piece the same week reported the Pentagon had reportedly used it to strike Iran. Make of that contradiction what you like. Anthropic went on to raise at $900 billion, then $965 billion six weeks later, officially the most valuable private AI company on Earth. OpenAI filed its own paperwork to go public. Two labs, racing.
Then the ground moved sideways. DeepSeek shipped a 1.6-trillion-parameter model running on Chinese-made chips — the whole stack, model and silicon, now functionally independent of anything Washington could restrict. OpenRouter, the platform developers use to shop between models rather than commit to one, saw usage jump five times in six months. And in late June, weeks after the US quietly ordered Mythos suspended for foreign users — no hearing, no published criteria, just an export directive — a Chinese lab called Zhipu released an open model that researchers said matched Mythos at finding bugs. Not eventually. Weeks later.
So the actual competition was never Anthropic versus OpenAI. It was: how long does an advantage last once you decide to sit on it. Governments spent the summer discovering they could switch a commercial model off entirely — for Anthropic in June, for OpenAI's new GPT-5.6 days after that, same restricted-release playbook both times, no law either time that said they were allowed to. By July, OpenAI was reportedly in talks to hand the US government a 5% equity stake ahead of its own IPO. Eighteen months ago the entire industry's position was "please don't regulate us." By July it had become "would you like a piece of the business instead?"
That's not a story about who wins between two labs. It's a story about who ends up owning them.
Then the same industry cut a lot of people loose, and quietly asked some of them back.
Cloudflare laid off 1,100 people in May and said the quiet part out loud — record revenue, and AI efficiency was the reason, no restructuring euphemism required. ClickUp, an actual SaaS company, fired staff and handed their jobs to its own agents. Remote grew revenue per employee fifty percent without adding headcount at all, which is either the future of software or the end of it depending on which side of the payroll you're standing on. I'd already gone on record months earlier saying the tidy "AI replaces jobs" story was too simple. Watching it up close all summer didn't soften that argument. It sharpened it.
Because six weeks after Cloudflare's layoffs, Ford quietly rehired engineers it had let go. Its automated design and production systems, it turned out, weren't as reliable as everyone had assumed — and the people who could have caught the errors were the exact people who'd already been shown the door. The institutional knowledge wasn't redundant. It was load-bearing. Nobody found that out until it was gone.
By the end of June, a report on companies actually using AI most heavily found headcount up over ten percent and entry-level hiring up twelve — the opposite of what the layoffs implied, at the same time the layoffs were still making headlines. Both things are true. Neither cancels the other out. The honest version isn't "AI takes jobs" or "AI creates jobs." It's that AI is cheap enough now to expose which parts of a job were ever load-bearing in the first place, and some companies are finding that out the expensive way, one Ford-shaped lesson at a time.
And underneath both of those stories, a quieter one kept resurfacing.
A psychotherapist who works with laid-off tech staff coined a phrase for it back in May: identity foreclosure. Losing income hits second. Losing the part of yourself that came from doing the job well hits first. Around the same time, a run of studies — PNAS, University of Arizona, Atlassian — found that roughly half of workers who use AI at work hide it, because the ones who admit it get rated as lazier and less competent, even when their output is identical. Effort has always stood in for something else: investment, competence, whether you can be trusted with the important stuff. Take away the visible effort and people stop trusting you, whether or not the work got done.
The research kept converging on the same unglamorous answer for what actually predicts trust in AI at work — not model quality. Whether people feel their concerns have somewhere to go. Glassdoor tracked a 240% spike in AI-related workplace anxiety through the first half of the year, with no corresponding drop in job satisfaction — which sounds like good news until you realise it's not resilience. It's dissociation. People are holding the anxiety in one hand and clocking in with the other, and calling that coping.
And then, right at the edges of both other stories: research on cognitive offloading found measurable changes in executive function among heavy AI users. Not "you'll forget how to do it." Something closer to — the capacity doesn't sit idle waiting for you to want it back. It softens. Which is, if you squint, the exact same mechanism as Ford's missing engineers, just running inside one person's head instead of across an org chart. Delegate the judgment for long enough, individually or institutionally, and you lose the ability to check that the delegation was ever a good idea.
None of this was foresight on my part. I want to be clear about that, because it would be a much better story if it were. I didn't call any of it back in April. The machine reading me the news every morning definitely didn't — it has no idea Tuesday and Wednesday are the same story wearing different clothes. I'm the only part of this whole arrangement that persists long enough to notice a pattern.
So a hundred days in, the honest answer to "what do you actually know now" isn't any one of the three things above. It's that the news was never really the headlines. It's the shape that only turns up once you've sat through all the ordinary days in between — which is either a genuinely useful way to watch an industry moving this fast, or an elaborate way of building myself a machine so I'd finally take my own advice. Possibly both.
Episode one hundred went out this week. If you want the version without the hundred-day lag, it's a daily habit now, not a one-off: harrysharman.com/briefly-ai, or wherever you get your podcasts.