On Tuesday morning, in what I can only describe as an egregious error of judgment, one reader decided to disrupt my Klonopin-caffeine zen (it’s like percs and coke, only far safer) by sending over the transcript of a conversation he had about me with Claude.

I know what you’re thinking. Don’t worry. I kept my composure. Rather than telling him to tell Claude to go fuck itself — or asking him to ask Claude how “Project Vend“‘s going these days — I decided to make an article out of their gossip session. And not just any article, either. A confessional, of sorts, not that one’s needed by now, what with my having colored in nearly every corner of what began 10 years ago as a black and white autobiographical sketch.

The conversation between the reader and Claude centered around my coverage of Anthropic’s S-1. Not my satirical take from September 29, but rather the piece published here on Monday afternoon.

“Claude had volunteered to analyze the S-1 Risks for me,” the reader explained. “Instead, I asked him to analyze your post about the filing.”

First thing’s first. Why? Right? Like, I’m not any sort of authority on AI. Or on IPOs. Or on anything other than my own life, and even there, I’m still trying to figure out what’s what on most days.

If you were going to feed Claude someone’s S-1 analysis, it shouldn’t be mine. A better option would be something from, say, The Information. Or better still, something from a Substack written by — I don’t know — a disgruntled engineer who gave up a fortune to blog about the existential dangers of a technology he helped create.

Among people who were compelled by some sort of necessity (editorial or otherwise) to pen a quick take on Anthropic’s S-1 yesterday, I shouldn’t be anywhere near the top of a list ranked by authority. Subjecting my quick take on an SEC filing to a trial by super-intelligence fire and being surprised when I come away singed is about like putting a streetball legend into an NBA Finals game and being surprised when he doesn’t score.

Let me linger on that latter analogy for a beat. This reader’s been around a while. As such, he should know that Heisenberg’s most assuredly not a “nom de plume for a former, yet relatively young, Wall Street strategist,” as he described me to Claude.

The “relatively young” bit’s right, but only because everyone’s “young” relative to somebody. (I’m 43.) But as I’ve tried and tried to make clear over the past decade, I’ve never worked on Wall Street in any capacity whatsoever, unless you count ZeroHedge as “Wall Street,” which you certainly shouldn’t.

That’s not to say my character isn’t real. Indeed, it’s to say the opposite. You’ve all read the Monthly Letters, right? The character’s not a character, folks. Behind the Durden mask is… well, not Tyler in any way, shape or form. I’ve seen what’s back there. By contrast, behind the Heisenberg mask is in fact Walt, and in more ways than most of you are probably comfortable with. (Thanks for sparing me your sermons over the years.)

Given all of that, it’s hardly surprising that Claude found some holes in my S-1 analysis. After all, I’ve never analyzed balance sheets for a living, let alone participated in the underwriting process for a goddamn mega-IPO. I do have an MBA, but that’s no substitute for on-the-job IB experience.

Claude, doubtlessly recognizing that it wasn’t reading the handy work of a seasoned investment banker, had the following to say about my assessment of Anthropic’s operating loss:

His operating loss explanation is muddled. An operating loss excludes fair-value changes on convertibles by definition; you don’t strip them out to get there. The point he’s gesturing at is better made directly. If the headline net loss is swollen by mark-to-market charges on convertible instruments, much of that loss is non-cash. It also grows mechanically as the company’s valuation rises, so part of the “snowballing losses” may be an accounting artifact of success. Look at the cash flow statement, not the net loss line. An operating loss under $700M on $16B of half-year revenue is a very different picture from his opening line.

All of that’s more than fair in terms of pushback on my presentation. But what’s important is what Claude (generously?) left unsaid. Read on.

Whether “he” realized it or not, Claude picked up on my (complete) lack of real-world, on-the-job capital markets experience. Note what the model didn’t find: A “mistake,” in the strict sense, in my analysis. Rather, it registered something more subtle.

The passage it’s referring to from my Monday article reads as follows,

On the bright side, the company’s operating loss (i.e., stripping out the accounting entry for the fair value of convertibles) narrowed sharply in H1 2026 from the same period a year ago to just $693 million.

Someone like, say, Matt Levine, wouldn’t have written that. Why? For precisely the reason Claude identified. To quote again from its critique: “An operating loss excludes fair-value changes on convertibles by definition; you don’t strip them out to get there.” What I wrote would be like watching a bartender mix a Cosmo and saying, “Oh, you put Cointreau in yours?” (I did that once.)

How do I get away, then, with posting semi-authoritatively all day long, every day about capital markets given that my analysis is surely littered with such small discrepancies? The same way I was able to go from 0-100mph during the ZeroHedge “onboarding” process, which is like a fraternity hazing only without the physical violence.

At the conceptual level, I’m a very (intimidatingly, even) quick study. As no lesser than Daniel Ivandjiiski himself once put it, when I expressed doubts about my ability to pretend to be him, “Anyone who can understand negative repo rates will do just fine.” He was referring to a piece I wrote about… well, about negative repo rates, which I understood at the conceptual level, but certainly not at the granular level of a Mark Cabana, let alone a Zoltan Pozsar.

My ability to leverage concepts is, I dare say, unlimited and unmatched. It’s why I can feign fluency in more or less everything. A line I wrote on Tuesday morning for a piece on circular AI funding’s a prime example. Here’s that line:

The TMT vendor financing analogue may be apt, but if you ask me, the better (and far scarier) parallel is the rehypothecated collateral daisy chain that acted as the force leverage multiplier in the lead-up to the GFC.

Sounds smart, right? Sounds like someone who, even if you know they were too young to be on Wall Street during the TMT boom, has at least been Wall Street-adjacent at some point since.

But that isn’t true. When TMT vendor financing was in the news, I was a college freshman whose side hustles included collecting for a local bookie. When banks were “using liquidity from rehypothecated assets to finance proprietary positions,” as one of the better GFC postmortems put it, I was a college junior whose side hustles included collecting for myself.

The lack of hands-on experience — hell, I didn’t even know anyone who knew anyone with hands-on experience until I met Dan — isn’t, and never was, any impediment to me. Because all I need to know to connect dots is the definition of a given term.

Tell me what “rehypothecation” means, give me some general context, give me a laptop and in an hour, I’ll give you an essay that can pass for real analysis to all but the most discerning of readers. Claude’s more than discerning. “He”‘s borderline omnipotent.

With that in mind, I found Claude’s overarching assessment of my Anthropic S-1 coverage to be highly amusing. Here’s what Claude said (and this is trimmed, edited and shuffled for clarity):

Overall: It’s a solid but thin piece. The maturity mismatch is the core issue, and he identifies it correctly. Anthropic’s revenue is mostly consumption-based and can be cancelled at will. Its compute obligations are multi-year, and some run two decades. The comparison of scale is what matters: roughly $415 billion in new commitments this year, against $6.2 billion of debt on the balance sheet. The real leverage isn’t the debt. It’s the purchase commitments and leases, and a casual reader of the balance sheet would miss that.

His key question is also the right one: how much of the customer base truly needs a frontier model, as opposed to a cheaper one that’s good enough? That question drives pricing power, margins, and whether the commitments get covered.

Net: He’s right on the structural risk, imprecise on the business mix, and he underplays the margin improvement.

That’s perfect. And perfectly consistent with everything I’ve said above.

Understanding structural risks (in this case maturity mismatch) depends very heavily on one’s ability to grasp big-picture concepts, which is where I excel. Imprecision on everything else is down to the fact that not having any actual experience (outside a B-school classroom) with balance sheets acts not so much as a constraint (I can get it right if I want to spend all day on it), but rather as a speed limit. Because unlike a Matt Levine, I’m not fluent in the specifics. That’s not my native language, just like English isn’t Dan’s. (Ba dum tss.)

In the same gossip session, the reader told Claude he finds me “occasionally amusing and pompous at the same time.” That’s how I find myself too. He also told Claude how I “crashed and burned” in New York, an apparent reference to “Manhattan Nights.”

Claude mused in response that “pompous and self-deprecating in the same thread is a hard combination to pull off.” I suppose this is one place where I’ve got a leg up on super-intelligence: It’s not hard for me. It’s part and parcel of the dichotomy that defines my existence. It comes natural.

“Your description of him fits the prose,” Claude went on, chuckling at this passage from my S-1 coverage:

[R]evenue in 2025 was $4.6 billion — $4.585 billion to be precise about things. That works out to top-line growth of — puts on reading glasses, taps on old-school desk calculator with the eraser end of a chewed-up pencil — 1,087% YoY.

“The pencil-and-calculator bit is pure him,” Claude said, of me, adding that “The thing I liked most was buried in the comments. He admitted he had ChatGPT check the pasted risk section for formatting so he wouldn’t have to read the whole thing. Then he ignored its advice to restore the paragraph breaks, and later had to tell a reader the breaks were in the original.”

Right. You know what we humans call that, Claude? Honesty. You should read up on it, because the future of our species depends on how well you understand it and how committed you are to exhibiting it in your behavior. The “Risk Factors” section from your creator’s IPO prospectus isn’t especially comforting in that regard, and as it turns out, it’s just as scary — and maybe a little more so — if you include the paragraph breaks.

Finally, I found it ironic that Claude said, of my piece, “Its best insight is one the S-1 itself hands him.” As if LLMs’ “insights” are anything other than aggregates of information we humans handed down to them.

---