Sunday Night Musings… AI fundamentals are accelerating and the innovation to usage flywheel keeps speeding up. Hard to internalize that none of this existed three years ago, and that the capabilities I’m excited to use daily weren’t possible last August. Things are compounding crazily. From a market perspective, we just had a legendary de-grossing in the market’s main theme and the indexes held. Success broadened into other sectors. Meh 7 became Mag again on the back of wildly impressive growth and some “new” ROI disclosure. We’ve even had a few “everything rallies”. With multiples undemanding, that bodes well for the opportunity set imho. Meanwhile, information asymmetry (“edge”) in large, well covered stocks has never been higher. The paradox exists because a higher % of people trading them know very little, and against fast secular change that makes for ripe opportunity. Watching people learn their xyz-th version of Jevons, or plant a flag on zero sum, is genuinely encouraging if you’re on the other side. And it looks like we can climb the macro wall of worry. The rates narrative can change as fast as it arrived (repeat after me: negative NFP) as wages lead and are consistently normalizing. We’re also starting to lap tariff impacts and, with a little luck, war effects can dissipate too. Ultimately, if the supposed hiking cycle turns out to be a never-was, it will do a lot for risk sentiment into year end. That said, the next two CPIs will be important. Where am I worried? Data-center political football and equity supply. The anti-AI narrative is short-term and bi-partisan convenient politically, but in the end I don’t think it engenders enough passion (faux hate?) to be substantive. I know it polls well, but as we’ve learned in 2016 / 20 / 24, polls have lost signal. It’s also simply too important to GDP at this point, and obvious solves exist (paying power bills etc). Push comes to shove some localities will extract their pound of flesh, same as it ever was. On supply, recent IPO unlocks plus a likely Anthropic deal is potentially a lot of stock to digest. But if that supply arrives alongside accelerating ARRs and compute deals, and that information gets democratized (particularly among the less informed but large capital macro community) the negative can flip into a positive. Interesting times to say the least. As always, godspeed :)

On this one, I don’t agree. This feels a bit like price => narrative, and the claims aren’t supported just stated. 1) Define “memory losing steam.” The stocks stopped going vertical? Good. That’s partly technical and partly the industry deliberately trading spot upside for contracted volume, floors and longer-duration FCF, i.e. the seeds of higher P/E And if LTAs are meaningless, you’re asking us to believe an oligopoly that knows the issue better than anyone given burned in 2022, knowingly gave up upside for no downside protection. Just a win-win for the customer. That strains credulity. 2) What exactly is “not coming up roses”? Cloud growth remains strong, lab ARR remains strong, and Anthropic exited the month in the low $80s after adding nearly $20bn. I suspect the big two collectively hit $350BN ARR at year end based on july momentum rolled fwd / compute coming online. What is the contrary datapoint for the labs? 3) Financing is the oddest claim. September hike pricing just moved materially lower (we printed neg CPI and neg NFP in succession for those not paying mind…), wage growth is trending consistently lower, and a hyperscaler just printed $25bn five times covered. The “Big unwind” kinda already happened? It could of course continue as a tape call but the best potential driver wasn’t mentioned: equity supply.

“They’re not going up. But they should go up. If they don’t go up, they won’t go up. They’re probably done. But fair value is 50% higher. Whoa!!! An anon account used to be bullish and is now bearish. I wish he told me 2 weeks ago before the stocks he now likes already outperformed. But US stocks need memory to work, and you should still buy no cash flow neos even tho they go down bc no one believes in memory, ” having a little trouble following? Me too!

I like to think about upcoming 'risk clusters', i.e. a confluence of important macro / micro events in a v short window. They can work for and against you. For ex, an upside EPS surprise can get an extra boost if it coincides w/ a favorable new macro tailwind (akin to a 'double jump', and it's often not priced appropriately in the vol mkt). Similarly, the fear of a 'bad parlay' across events often gives mkts anxiety in the approach. Like having two mid-terms in the same week. June had the latter issue (SPCX, cpi, Warsh + Seasonal momo fear). This dynamic is why today's NFP is more important than it looks. A negative (!) number on the back of a negative CPI, aside from being evidence we live in a simulation 😜, helps soften an upcoming risk cluster: jackson hole => anthropic ipo? => "fed hikes in sept!". And within the print, wages are the part that actually matters for that last leg. The hawkish case rests largely on inflation broadening into second-order effects, and wages are THE canonical second-order driver. AHE at a cycle-low 3.2% and still decelerating says that channel is going the right way. And that's with last year's cuts and the early '26 tax breaks already in the bloodstream! A Sept hike would now mean hiking into negative payrolls, 3.2% wage growth, a soft core PCE month, $77 oil and 2.2% breakevens. Good luck with that!

Ladies and gentlemen, I present to you the agent harness! And it IS important but not in the way you think… I’m frankly surprised people are surprised (even seemingly the index creators themselves?). Anyway, much of this is a mechanical mix shift from how models are actually used as usage and token consumption grow (sub agent fan outs, frontier synthesis). Strange not to understand this and be investing near the space. Btw, a simple Claude or GPT query can answer this.

I’ve been thinking about what the “AGI moment” might look like. My best guess is that it’s when context management disappears. Right now that layer is manual labor. For example, I’ve built an elaborate system around Claude and Codex that includes shared memory / source of truth, knowledge graphs, tagging systems, saving rules, learning loops, etc. (if by chance you don’t speak nerd, ask gpti). It’s materially better than using the apps, and more like working with a teammate (which is why I’ve recommended people spend the time to tinker with the coding version / co-work). But I only get that experience because I’m doing active and intentional context management. The end state though I think is that the models “just know.” They’ll know what matters, what changed, what’s current, what needs to get tossed. This will feel like intuition to the user. Under the hood it will take state coherence (shared memory / state, accessed and passed effortlessly between agents working toward a goal), facilitated by LOTS of upgrades to memory / gpus / networking. But nobody will ever experience it as context mgmt, it will just feel like intelligence. Interestingly, power users are living on a slope, starting with Claude Code’s fall kaboom. Better models, harnesses, context systems, etc. Most people though still evaluate AI through a *mildly* stateful chatbot. I suspect they may end up skipping this ‘messy middle’ entirely and jump straight to persistent and intuitive agents, i.e. a slope for early adopters, a step function for the rest. And of course it’s the step function that’s worth playing for.