substrated

ex-hedge fund. techno-utopian. thoughts are my own, not investment advice

[substrated @hardinmallow]

i think the next leg of the AI trade could be emerging and frontier markets.

[substrated @hardinmallow]

seeing a lot of debate on the ant ipo and i genuinely can’t believe pod brainrot has made its way into the discussion. no one is going to give a shit about margins/open source until it shows up in numbers. 1. this is the average person’s first, and really only, chance to hedge the career risk posed by AI, pod bois won't admit this because they are one trick ponies but compute is not the best way to express the long term AI trade anymore, the 'labs' are, 2. it is in every serious player’s interest for anthropic to succeed (trump, nvda, google, etc.), so every down day will be bought aggressively or defended via headlines. 3. everyone thinks people are worried about margins and open source, so they will buy aggressively on the break. 4. anthropic’s pitch is not going to be “we are selling tokens.” it’s going to be “we have a near monopoly on scaled compute and we sell the smartest model in the world at the lowest price, and eventually can build every business ourselves”. I think ant probably goes to $5T before anyone blinks.

[substrated @hardinmallow]

twitter is great because legendary investors used to make a ton of money and retire and everyone would just assume they were geniuses but now they go straight to twitter and prove they are complete and utter morons with zero differentiation

[substrated @hardinmallow]

so i’ll be honest, i don’t know the indian economy super well, but all the people in your comments saying “no idea boss, india is cooked” should give you some pause. i agree the next couple years look rough. IT is cooked, FPIs have been dumping, and it’s still rly expensive vs other emerging markets. but i can’t help but think that once they get off the opiate that is outsourced white collar work, they should actually be one of the larger beneficiaries of AI. i’m working on an EM trade writeup i’ll publish soon so adapting some of that here… there are obviously a ton of hurdles for a country like india, but decades of research on why some countries get rich says one of the biggest drivers is workforce skill, and india’s bottleneck has always been bad schools, weak managerial know how and bad healthcare, which your AGI pilled friends will tell you are all solvable with AI. there’s a good paper (bloom iirc) showing indian factories got ~17% more productive just from adopting western management practices, and now their entire economy will have all of that at their fingertips basically for free. in theory that flows into higher wages, more consumption and a bigger formal economy. same logic at the country level. india is like 30x poorer per head than the US, and catch up growth usually comes from cheaply copying what richer countries already know, which AI just made way cheaper. i just wouldn’t want to be structurally short when that turns.

hopefully some of that made sense/helps

[substrated @hardinmallow]

i have said it before and i will say it again, if you think that a growth decel from the frontier labs will lead them to do anything short of putting legacy enterprise saas in the fking ground, you have lost your mind

[substrated @hardinmallow]

new writeup coming on my emerging markets thesis, investors are so asleep at the wheel it is not even funny

[substrated @hardinmallow]

usually skeptical of doc huberman but i thought his ep on ILTB was solid, s/o @patrick_oshag. he was saying the frontier labs become biotech companies (i think he probably overstates this but intg take ntl). made me think about BCIs/metaverse convos that i have been having with my most AGI pilled friends + i've just finished snow crash for the tenth time (highly rec if you have not read) and have been thinking more about this stuff. until recently i was focused on AI removing the content friction in ar/vr, now i'm wondering whether the bigger theme here is AI pulling the BCI timeline itself forward.

my timeline is full of techno optimists and AGI pilled semis analysts, and i barely see anyone spend real time on the BCI angle, even though they are bullish on neuralink. most of the discussion still seems to live in healthcare circles but let me know if there is anyone interesting talking about the externalities of this stuff.

i figured i would take a first stab (that i have seen) at getting the trade/investing convo started, mostly on the physical bottlenecks / supply chain side for now. so, first stab, and credit to a neuroscience friend who put me onto this one (although i fear he is not be a good investor): iridium. he metnioned paradromics' uses solid platinum-iridium microwires for their connexus product. Claude is telling me the makeup is 421 wires per module, up to four modules per patient, 40 micron wire, 1.5mm deep, with a 90/10 alloy Pt-Ir, (which i couldn't confirm) and it comes out to ~7mg of iridium per implant. the whole world only produces about 7 tonnes of it a year, nearly all of it as a byproduct of south african platinum mining, so hard to quickly increase supply. (matthey has iridium in deficit this year because of green hydrogen) 10m implants a year is ~70kg. 100m is ~700kg, which is 10% of global supply. and at under $2 of iridium per device the buyer is probably fairly price insensitive. unfortunately I think that is the extent of the Ir angle, neuralink says "high-volume production" this year, which is understood to be hundreds to low thousands of implants, so we're four or five zeros away from any of this mattering, with a few other caveats below.

There are some other interesting angles that I came across

packaging. every channel needs its own wire through the wall of a sealed implant, and that wall needs to be leak proof. the feedthrough industry is mostly pacemakers and stimulators, with only a handful of external pins. the new neural devices want something like 1,000 connections per sqcm, there is allegedly a patent out there that claims nobody sells a hermetic one that does that (was unable to find myself even with AI). the closest thing i can find is a platinum/ceramic composite out of heraeus called CerMet, licensed to a french company called SCT ceramics, which quotes up to 800 channels per sqcm. both are private. maybe the move is just to get long the people who supply the pacemaker companies, so maybe $ITGR?

surgery. this is the one that moved this year with the neuralink dura breakthrough. synchron seems to be going a different way and threads its device up through the jugular, with no neurosurgeon in the room. my napkin says a robot doing 8 a day gets through ~2,000 a year, so 10m implants is ~5,000 robots plus the rooms and people around them, so seems like a tough one to really gain exposure to. neuralink builds its own robot, and all of the cool ones are private, so the listed angle is $CLPT, which makes the MRI-guided navigation other people use to place things in brains.

electrodes. out of my element here and i have no contacts to check this one against, but precision neuroscience went and bought an entire foundry outside dallas in 2023 because they had to "rethink the medical device supply chain" (when a 2 year old startup buys a fab, the thing it wanted probably wasn't for sale). the hard part is lifespan. you can't speed up ten years of sitting in salt water being attacked by an immune system (someone should run the replacement cycle math on brain computers lol). everyone who actually makes these is private. the nearest ticker is medtronic ($MDT), which partnered with precision, but the stock won't move on BCI stuff for a LONG time.

chips. the physics here makes this intg but trade is not clear. the implant can't warm the tissue around it by more than about a degree, and sending raw data from 1,000 channels takes roughly 250mw against a budget of about 35mw, so compression has to happen inside your skull on custom ultra low power silicon.

the caveat that eats all of the above: if ultrasound works (merge labs angle) you skip the electrodes and the feedthroughs, and most of the surgery, and a completely different supply chain wins. the listed name there is butterfly ($BFLY), whose ultrasound chip is licensed into forest neurotech.

iridium angle is probably just fun napkin math, and packaging and surgical throughput are the ones i'd bet survive if implants win. for the people who want tickers: $ITGR, $CLPT, $MDT, $BFLY. BCI is ~0% of revenue for all four and the suppliers that matter most are private, so this is a watchlist and nothing more. not advice. will keep digging when i have time.

[substrated @hardinmallow]

hi brett, big fan. here's a less glib version of my take. i agree w/ your definition of judgment, weighing unique views/insights and comparing ur view to what the market thinks. my personal view is that the next big unlock is probably not finding more creative ways to use excel with ai, it is finding a more sophisticated way to automate the judgement.

to step back quickly, to me, ai represents everyone having the same facts with zero friction, when everyone has the same facts, the near term gets priced faster and more accurately and the edge moves further from the facts at hand, and closer to what those facts will mean. the problem is that the tree branches the further out u go, competitors respond, capital comes in, discounts go up, regulators show up, mgmt manages to the stock, small changes have time to compound and rip up your base case and, oh, don't forget about timing, an excel model and its variations are usually just one path w/ have baked scenarios stapled on and the pm's gut is doing the weighing.... my view is that simulation could be the answer to how you actually map the distribution in a rigorous, auditable manner.

before ppl say 'excel can already do monte carlo sims' that is not what i am saying. the paths should interact. right now excel doesn't have a way to represent that. it was built for ppl typing assumptions into cells, and on occasion they would do a nice side build, and it was good enough for what investing was in the past that we just kind of accepted it because it got us 95% of the way there, now i think that excel has become a bottleneck for what investors can do, so we should probably start to think of ways to get rid of that bottleneck.

other industries/fields that deal with predicting complex systems moved past this a long time ago. weather guys run different models and read the spread, epidemiologists and supply chain ppl simulate agents to see how stuff plays out, i actually don't know what they do if i am being honest, but i know that there is massive toolkit out there that finance bros are not tapping into. fundamental investing really hasn't touched any of that bc building it took a team of engineers, and all of the good engineers have gone into tech for the last decade.

w/ coding agents that should change right? if i were you I would literally rebuild excel from scratch, keep what you mentioned (auditable, transparent, easy to visualize), and then find a way to let fundamental analysts leverage all of the advanced forecasting techniques that have been developed over the last 20 years. i.e. how can you build a system to start to understand the all of the interacting paths (where the scale comes in), you can see where outcomes cluster, where they split and how the whole distribution compares to what the price implies, which is exactly what you are talking about with weighing insights. great investor's heuristics work alright, but they are becoming commoditized quickly now that they can be hardcoded. like in chess, everyone can see the entire board now. the edge is reading further ahead, and past a few moves u can't calculate by hand.

simple example. today most analysts probably model a restaurant chain as some drivers for same store sales hopefully they have a thorough unit economics build and they plug in data sources to get their estimates for the next 2-4 quarters, AI can do that in 5 minutes, but the analyst doesn't have an easy way to do much more than that. they should be able to model the actual business: every store, local demographics, compare that to demographics with other ramps, what is the spread of the ramp curve, how does timing effect the PV, local wages, how traffic reacts when menu prices go up (for this demographic how much of their wallet share does it eat into, etc.), how much a new store steals from the one down the road on a product by product basis, and with all of that information all stuff that is very easy to build, but you have to build a platform that lets analysts do audit, check, etc.

but who knows, maybe jane street just builds a massive model and trains it on every source of information that a l/s investor would ever use and any extra work to codify an investor's inductive bias is useless!

and maybe i was being too harsh on the talent point, you just need to actually want to try to do stuff like this, very easy with AI.

hope this is somewhat helpful and not just jibberish, always happy to chat/dms are open

[substrated @hardinmallow]

@ZeroDark9000 @FundamentEdge @simile_ai yes great point, this was the first thing that I thought when i saw Joon speak, pumped for what they can come up with

[substrated @hardinmallow]

what if AI is like that pill in that one american horror story episode where talented people take it and turn into prolific geniuses and non-talented people take it and turn into zombies, because I am pretty sure i am turning into a zombie

[substrated @hardinmallow]

woah instinct $0 to $10b to $0 in like 3 seconds

[substrated @hardinmallow]

@RihardJarc i think that literally every single person understands this

[substrated @hardinmallow]

can $AAPL just hurry up and finish this AI personal assistant race, the smart people need to be doing something else

[substrated @hardinmallow]

companies are blatantly buying any service that can possibly help them leverage AI, cough $PLTR $ACN cough, and now wannabe value guys are taking a victory lap on their "software/advisory was oversold" thesis, these stocks are 10x more fucked than they were in january people just can't manage risk properly

[substrated @hardinmallow]

going to the dentist tomorrow, told them to hold the novocain, so pumped

[substrated @hardinmallow]

woah the time-to-token-limit for all claude models is make them unusable, how long until anthropic finds a compute breakthrough and nukes the infra trade

[substrated @hardinmallow]

if g4 was any good google wouldn't be doing a limited rollout

[substrated @hardinmallow]

i think that a lot of the low hanging fruit for real multi-bagger arb probably happened over the last 6 months. I would argue that the multi baggers are more right place right time than super deep work, but who knows, really good ideas are obvious and so right now i think people are just hoping they can catch the obvious ones faster. my question is how long it will be until the average US analysts covers EM names, i think there are some major unlocks there

[substrated @hardinmallow]

@AnalystEgg @P_Remarks ODOT overhang probably needs to be resolved for the stock to work but agree very interesting at these levels, surprised crh, vmc, mlm have not taken a bite yet... mgmt does seem a bit asleep at the wheel

[substrated @hardinmallow]

whatever you do, do not ask a VC what they use AI for