Between Calls #3
Anthropic's fear flywheel, radiologists vs. the prediction, and building an AI assistant that actually works
I did the reading. Here’s what matters.
What caught my eye
Anthropic’s fear flywheel
When I was listening to the All-In Podcast talk about Anthropic’s fear tactics the other day, I realized I’d actually done some research on this that I should have visualized for the dive. Because when you look at it, there does seem to be a lot more going on here, and the pattern doesn’t start with Anthropic. It starts back when Dario, Daniela, and Jack were working on GPT-2 at OpenAI. They said that was too dangerous to release too. Nine months later, OpenAI quietly admitted there was “no strong evidence of misuse.” Jack Clark himself later conceded they “didn’t do enough experiments to justify each misuse” claim they’d made.
Same three people. Same playbook. The latest one is Mythos.
When I think about the kind of playbook they’ve been running, they do this quite a bit. And the most interesting thing I found when I used Anthropic to do the deep research was that the timing is pretty coincidental, if you can even call it that. Every single major safety research drop lands within about 30 days of a funding event or product launch:
Sleeper Agents paper, $18B Series D
Bioweapons Senate testimony, Amazon $4B plus Google $2B
Alignment Faking paper, $3.5B Series E at $61.5B
Opus 4 “blackmail” scenario, revenue ramp to $14B ARR
Mythos “too powerful to release,” $800B valuation offers
That’s five examples. To me, five in a row is not a coincidence, that’s a business model.
At the end of the day, I do think this is one of those cases where the business model they’ve built genuinely aligns with their view of the world. They probably believe every warning they put out. They’re also probably running a sophisticated commercial operation around those warnings. Both can be true at the same time, and I think that’s the part most people are too tribal to actually say out loud.
Sources: All-In Podcast E267, Anthropic Mythos announcement (Apr 7, 2026)
Radiologists vs. the prediction
I know a lot of people in imaging jobs right now, and people who went to school for this around the same time everyone was saying these jobs wouldn’t exist. Back in 2016, Geoffrey Hinton stood up at a Toronto conference and said we should stop training radiologists because deep learning would do the job better within five years. This was before we had AI as we have it today.
One of my favorite things Jensen Huang said on the Lex Fridman Podcast is that tasks get automated, but purposes don’t, and that the human owns the purpose. When he made his radiology comment, it was pretty optimistic. But when I dug into the research, it looks more like the demographics changed, not the job.
The image above tells that story. Radiologist headcount grew 17% from 2014 to 2023. Real pay (inflation-adjusted) is up 23%, with nominal pay climbing 63% from $351K to $571K over the same decade. Imaging volume has climbed steadily, driven by an aging population and expanded use of CT and MRI. Hinton himself walked the prediction back in 2025, telling the New York Times he was “wrong on timing.” The job didn’t disappear. The work doubled.
The need for images outpaced AI, until now. Looking at the research, AI is just getting to a point where it can actually disrupt this field. I think the next five years are going to matter more than the last five. Every chance I get to do research like this, I try to find the optimistic side of job displacement, the version where automation frees people up to do work that matters more. Stories like radiology make that harder to land. The optimistic read isn’t that AI created demand for radiologists. It’s that when purpose outruns automation, the humans who own the purpose win. The harder question is what happens when automation finally catches up.
Sources: Lex Fridman Podcast #494 (Mar 23, 2026), Neiman Health Policy Institute / JACR
Real demand, real competition: why the AI bubble talk misses the point
There’s a lot of conversations around an AI bubble that people (myself included) have daily, weekly, monthly. A couple of things around that really stick out to me.
Anthropic has gone from $1 billion to $30 billion in about 15 months. In April, they actually passed OpenAI, so that tells me not only is there real demand, but there’s real competition.
I also think enterprise AI is such a fast-growing category, the fastest-growing in history. It’s gone from $1.7 billion to $37 billion since 2023. It’s insane.
I also think there’s so much corporate and enterprise usage of AI that governance is starting to become its own vertical. Companies like Kai just raised $125 million and JetStream Security raised $34 million, both focused on governing and securing AI agents inside the enterprise. Which means so many of these companies feel like they need help with governance because it’s such heavy usage.
Sources: SaaStr, Menlo Ventures
What Microsoft is up to
Microsoft flipped the switch on the largest AI cluster in history
It was really cool to see that Microsoft just flipped on the most powerful AI cluster ever built, ahead of schedule, in Mount Pleasant, Wisconsin. This went live on April 16th and uses hundreds of thousands of Nvidia GB200s. It anchors about $7 billion in investment from Wisconsin itself.
I wrote a couple of weeks ago about the largest round of investments in history for startups. I think it’s interesting that the same month we get that announcement, we get the largest AI data center in history flipped on.
Also, the buildout of this data center is really pulling renewables and new energy tech forward, because 27% of global data center electricity already comes from renewables. Hyperscalers are signing huge deals for geothermal, small modular nuclear, and other things like battery storage.
While a lot of AI investment gets talked about in the realm of software, this is really cool, showing that a third of U.S. GDP growth in the first nine months of 2025 came from real-world, physical updates.
Sources: Microsoft On the Issues, CRE Impact
Kids & tech
I’m starting to think 2026 is going to be known as the year kids’ tech stopped being a free-for-all, where companies can do whatever they want to kids. There are four stories over the last two or three weeks that have stood out to me around COPPA.
One example is that federal regulators actually drew a pretty hard line with data that takes effect April 22nd. It actually requires separate parental consent to train AI on kids’ data. Things like biometrics or personal info can be penalized for more than $50,000 a day. When I read that, it’s very surprising to me that that wasn’t already a violation.
I think the biggest data point I saw is that there’s over 1,500 AI bills across 45 states. Out of those 1,500, over 130 of them are specific to AI and education and how it can be used. The narrow, specific ones are passing really fast. Kids rely so heavily on a digital environment. Once they get into middle school and high school, I think this just has to be fast-tracked.
I mean, even seeing that Meta and YouTube were held liable for engagement-maximizing design choices on kids under 18 also speaks to this. I think this is just the beginning, and once this regulation starts getting approved, it’s going to be a free-for-all.
Sources: Federal Register (COPPA Rule), MultiState.ai State AI Bill Tracker
What I’m testing
I hired an AI assistant named Bill Carter
The reason I gave Bill Carter a persona versus just making him an AI agent was because I think of this as somebody I’ve hired to help in my daily life, and I only give him access to things as I start to trust him. Giving him a persona allows me to interact with him in a more natural way. If other people hear from Bill Carter via email or Facebook, they naturally kind of think he’s a person and interact naturally, so it just helps smooth out the entire process.
One of the things I really wanted to do when I started using AI agents was create a persona with its own email. That persona was Bill Carter, and he has his own Gmail account. He has his own login to my computer with his own access controls that I either allow him to have or don’t allow him to have for my daily vitals, where I get an email on my health based on my Garmin data and what supplements I’m taking. Those emails come from Bill Carter. They don’t come from my own email.
He’s also in charge of making sure my inbox stays at zero emails, so he runs several times a day, drafting email replies. Honestly, the most helpful thing he does right now is actually schedule calendar invites. One of the most recent things I’ll tell you he did for me was schedule a calendar invite for my vehicle renewal. I forgot to renew my vehicle, so he scheduled a calendar invite. It reminded me automatically one day last week, and I submitted the renewal. He puts all the info I need in the description. It’s actually very helpful.
The other thing is that, because replies come from him, I have actually found that people will go back and forth with Bill Carter, and I don’t even have a say in it. Sometimes he just replies to people, or they talk to him for the Olive & Oak coffee brand we have. He handles the social media engagement and has really long conversations with people, and so he’s been working out really well.
I think the next step with Bill Carter is just empowering him with more things now that I kind of trust what he’s doing. For example, right now for my personal inbox, he drafts email replies. He doesn’t send them. I think I’m going to start letting him send them to save me time, unless it’s somebody I have not interacted with before. I’ve been thinking about those controls.
He even has a profile pic and a few other things that I’ve given him access to, so he’s definitely throughout all of my workloads that I build out. He’s always at the core of it, organizing it. It’s not just one specific task.
The short list
Q1 2026 venture funding hit $300B, 87% went to AI. Top 5 deals (OpenAI, Anthropic, xAI, Waymo, Databricks) accounted for 73% of total value. Concentration is extreme.
Over 40% of seed and Series A rounds are now mega-rounds ($100M+). The largest seed ever: Advanced Machine Intelligence at $1.03B. The middle ground of traditional $5-50M seed rounds is disappearing.
Wayve raised $1.5B for London robotaxis in 2026. UK-based, zero-shot deployment in 500+ cities. A three-way race (Waymo, Tesla, Wayve) is forming.
Mistral published a European AI sovereignty playbook alongside $830M in GPU financing and a 1.2B euro data center deal in Sweden.
DeepSeek V4 shipped with a 1M token context window, open-weight, multimodal. Chinese labs continue closing the gap with frontier Western models.
FTC enforcement on “AI washing” is accelerating. Section 5 of the FTC Act is the primary lever. A dozen+ cases in 2025 targeting companies overstating AI capabilities.
27 states are pushing back on AI data center expansion, requiring developers to bear infrastructure costs. Maine has the first moratorium, through November 2027.
Only proofread with AI, never written.
- Alec