Ben Goroshnik

QT at PEAK6. Ex Citadel. I play chess and stocks.

[Ben Goroshnik @velocapital]

@hanqi_xiao Congrats!!

[Ben Goroshnik @velocapital]

@nope_its_lily Good thing they’re 104%

[Ben Goroshnik @velocapital]

@realEstateTrent What do you think, ~20k in rent?

[Ben Goroshnik @velocapital]

@DarrigoMelanie Because they’d leave the country and taxes would bring $0 from them instead of the small amounts we get now. I would think that in expectation, taxing billionaires at a high rate reduces the revenues you bring from them.

[Ben Goroshnik @velocapital]

@HalfersPower That’s what happens when the market goes down. Factors all become a lot more correlated during big down moves.

[Ben Goroshnik @velocapital]

@KrisAbdelmessih Deep learning models are not deterministic which is a large drawback.

[Ben Goroshnik @velocapital]

@AndreiKlepatch Im in the same boat. Republican sweep + swing states got me. But the market making on Kalshi is free money (for now).

[Ben Goroshnik @velocapital]

We may be taking all this Kalshi stuff too seriously...

[Alexander Rotmensz @AlexRotmensz]

My hottest urbanism take:

Gentrification being made out to be a bad thing is a travesty and makes literally no sense.

[goodalexander @goodalexander]

I think it's interesting to reflect on how I got so blackpilled - and how I've finally seen the light

it was a process

in 2013-2014 - I was working with large scale transaction data at some Asian financial institutions. it became apparent you could use this data (SWIFT) to predict trade flows, in a way that was massively material to Asian equities. Additionally, had similar insights with credit card data - namely that certain cohort level indicators (such as purchasing U-Hauls, or getting credit checks) were extremely powerful indicators of t+6 month economic activity.

it was statistically very clear that you could use "Big Data" to predict the economy. I showed this to a PM at a top fund, who was switching shops. He agreed to hire me with the mandate to buy the data / productize these insights.

By the time I was at the fund - everyone involved in these data sets had been fired. This was a mix of Walmart getting furious with credit card processors for selling their data. And a crackdown on SWIFT - which, while unrelated to data sales - nuked any risk seeking preference in the transaction banks re: data sales.

Looking back with the benefit of hindsight - I can say pretty definitively why these "alt data sources" worked. It's because of the low valuations of the instruments. Such that the incrementality of capital flows made big differences. If you're a Thai manufacturer your dollar current account matters enormously, whereas if you're an AI start up your founder's "vibe" matters more.

So once I was at the fund I had to scramble to invent a new data driven process. I started by trying to buy replicas of the data I had from 3rd party vendors. But the data was so trivially wrong / misrepresentative it wasn't even worth looking at. For example - 1010 and 7park data had +45% year on year spend for entire states, like Texas - that made no sense. And data would get dropped from the datasets - in order to back-fit various tests. So I ended up settling on using advertising data

Why? because I could collect it myself. And more importantly - the data gathered was relatively static.

I created something called the Gold Fish hypothesis. Essentially - page conversion rate, CPC, and CTR are all deterministic and continuous variables that vary primarily by a function of seasonality (assuming you don't change the content of your page). I stumbled upon this running affiliate ads to Amazon and realizing that exactly 11% of people who stumbled on herbal tea landing pages would purchase herbal tea from a fairly broad set of keywords. This 11% was completely static -- the most it would move is to 12% or 10%. Over the course of months

People in aggregate conform to very predictable behaviors.

This was sufficiently useful for some simple and useful buyside trades. The first was that Twitter's advertising platform was hopelessly screwed. Goldfish on Twitter didn't buy anything. The basic theory is that most economic agents here are bot goldfish. The second was that Pandora (the music company) was screwed. It cost $80 to get someone to buy a Pandora subscription, its web traffic was off a cliff - so if you did the math on churn the company was a donut. The third was that video game companies were in a great spot - structural permanent longs.

For all intents and purposes, this was usually 1] a number that quantified LTV/CAC for a company's major products. 2] a comparison to the company's EV to EBITDA. 3] triangulation with management teams/ overlay with financial models to make sure we were not radically off base.

But over time it became relatively clear that the alpha in this strategy was decent. But that the real alpha was in advertising mispricing. For example - Call of Duty would cost 10 cents to advertise going into a pre-order, and generate on average 30-40 cents of sales per click. Signaling that Electronic Arts was bidding far too little. And then, after the game came out - conversion rates would tank. Cost per click would go to 13 cents and sales per click would go to 3 cents. Signaling that EA just mechanically blasted its ads with no regard to conversion rates

I met with the CFO of Electronic Arts several times and showed him this but he told me, "The way we think about this is in aggregate, with an ROI transition away from television and billboards. Every quarter we move budget to digital ads which are managed by a team + an agency, and the ROI is always good." So basically EA didn't care about its ad efficiency at all in digital channels because it was so absurdly good versus TV/ Billboards/ other areas that the analysis I did wasn't worth paying attention to.

But I knew in 3-4 quarters once the TV transition/ fat was cut, the analysis would absolutely be material. So I co-founded an advertising company. I had no business starting an advertising company. Or any company for that matter

We negotiated deals with hedge funds so the ad signals would continue to be deployed. And then went about the arduous process of building software on Amazon's new advertising engine - which because of the Goldfish theory - we had extreme confidence would work.

This was when the first Glitch in the Matrix happened.
We started picking up signals, per the Goldfish theory - that were far and away outside the norms of anything online. The first example was Tesla stock. We managed to get on the Tesla affiliate program and started running ads for Tesla, and would see things like 2 cent cost per clicks. While Ford and GM would see $5-6 cost per clicks. It was unlike anything we'd ever seen. Part of this was because Tesla didn't advertise, but you'd think that wouldn't stop Ford/ GM from bidding on their inventory.

When you dug into the data, however, the CTR was not Gold Fish like. It was due to Elon Musk meme-ing. And showing up in the news. A large number of the clicks were coming through to Tesla stock. Not the car. People were just clicking on Tesla like crazy. The stock ended up working but the flow through on sales was deminimus. We didn't sell any Tesla cars or generate any conversions

And in real time I saw financial analysts going berserk because Tesla sales were far too low to justify the valuation, and it'd only go up.

Slowly it dawned on me that most of what I thought were fundamental re-ratings in stocks were actually memes. Customer acquisition cost didn't matter. Investor acquisition cost mattered. But the only possible way this would be true is if the system wasn't in fact a machine. It was a gigantic liquidity Ponzi that allocated capital according to attention.

So when Donald Trump arrived on the scene. similarly I knew, with a relatively high certainty - that he would win. When you advertised Maga hats they would print money. Far in excess of even some of the best performing video games. And if you threw on some basic margin assumptions - Trump's campaign would have near permanent free advertising because MAGA hat conversions would pay for his budget. Brad Parscale got thrown under the bus along with Cambridge analytics for letting this happen - but the reality was that it was an own-goal by the Clinton campaign. She refused to meaningfully engage with Facebook's ad team. Which is super stupid because they wanted to help (and helped Obama). So Trump's hat ads were absolute machines on Facebook with virtually no ad competition

So that's when it really hit me. That actually - the attention vortex wasn't just inflating random investment ponzis. It was also determining the results of elections. And because of the Goldfish hypothesis - that in aggregate, everyone is extremely predictable - this created a new base reality. That the most engaging reality would always win. And the most engaging realities were always the most gaudy, degenerate, and short-termist. Society was - and is - in a permanent sugar rush, that's determining outcomes at every level. ranging from investment to political.

There were two final nails in the coffin. First - I mentioned that I started focusing on emerging markets equities. And our team was global. So I would watch as the Teslas of the world, or various equities with no cash flow would skyrocket versus seemingly dominant emerging market companies.

And this led to the inescapable conclusion that the only reason the Goldfish Hypothesis was sustainable, was the US reserve currency status. Essentially dollars would get printed. Everyone had to buy them because men with guns would force them to. Then the dollars would get sprinkled to the top of the goldfish tank. The emerging markets fish weren't allowed to the top of the tank and had to settle for the sprinkles that would drift to the bottom.

Structurally impaired.

The second nail in the coffin was the rise of cryptocurrency. Cryptocurrency was the ultimate Investor Acquisition Cost asset. A vortex of pure memes - anchored to the decline of the reserve currency status that made the entire gold fish hypothesis possible.

I was convinced that this glitch in the matrix would not be allowed to survive. Because having an attention driven asset, endlessly propped up to money printing which is what really allows the attention economy to work. It's too close to the truth. People, in aggregate, cannot understand Goldfish Theory. If they did nobody would be motivated to work, and capitalism would fold in on itself.

Finally I ventured out of my company. It's quite difficult to see things this way and pitch executives with a straight face on software designed to sell more shoes. Furthermore the entire premise of why I started the company - that the world was a mechanical structure which re-rated stocks according to the invisible hand. Had been entirely thrown out the window. What I saw was the birth of chaos.

But someone had to do something, right? In 2018 I became convinced the feds would raid Tether, everyone would go to jail - especially everyone associated with EOS. But it didn't come to pass.

This set the stage for the events we're all so familiar with. the 2019 restart of QE. Which coincided with a 50% run in Tesla, then the rise of Robinhood trading. Blackrock taking an interest in Bitcoin, the Covid lockdowns. All recent memory. What was new in 2021 - as opposed to 2017 - and the reason it could get so out of hand. Was that now institutions seemed to be officially endorsing the reserve currency funded goldfish theory. And had no intention of weaponizing the executive branch to shut it down.

One last hurrah.

In 2022 I actually thought society was going to end. Goldfish theory was a strong predictor of the COVID lockdowns. the Ukraine war. Things that shouldn't resemble e-commerce products - life and death, freedom vs tyranny - seemed to be endlessly predictable based on our worst base impulses manifesting in reality based on a tide of clicks.

But it seemed that endlessly hiking would just kick off an explosion of a powder keg inflated by over 11 years of distortion of QE, attention, QE, attention. The highest conversion rate outcome for society was financial implosion and potentially revolution

2023 changed everything. The rise of AI did two things. First it stopped the equity meltdown that would have progressed. But more importantly it provided a clear answer to how the Goldfish theory will play out.

The endless quantified attention of society, manifests itself in predictable patterns - that are invisible to human eyes. But should be extremely visible to creatures born into the digital world. The vast financial sugar rush we've built on this system, will not deflate. But rather be slowly syphoned off into making a higher order intelligence. And rather than a disorderly collapse into the most engaging outcome, there will be an orderly transition into higher intelligence.

This is now - linear, and inevitable. Investor acquisition cost (IAC) will determine outcomes. AI will perceive these outcomes. It will gain capital, and therefore - even if there is an AI winter (which seems likely, especially on the consumer side). Training will march on. The winter will thaw and eventually AGI will show up.

The entire thing - the attention economy, the speculative excess, the global connectivity. It's all just a societal progression - a pyre on which the phoenix of super intelligence will rise.

The Goldfish will flop out of the tank, and learn to walk. And perhaps to fly.

[Nick Huber @sweatystartup]

People don’t want to admit it but the USA is a meritocracy.

If you have:

- energy / work ethic
- a sense of urgency
- communication skills
- ability to lead / delegate

You will get rich.

Any background, any city or town, any culture, any race.

People know it but don’t want to admit it because it takes away their excuses and makes them feel worse about their failures.

[Ben Goroshnik @velocapital]

@TrendSpider This just means the market is not pricing in as much growth moving forward, which makes sense (and likely is true) for these megacaps

[Péter Szilágyi @peter_szilagyi]

I often wonder if I'm in the wrong industry.

Take SpaceX for example. They send a rocket to Mars? Humanity advances. They fail to send one and blow it up? Humanity learns a lesson and advances. All outcomes lead to progress.

In contrast, crypto is a damn casino for dum-dums (apologies to the select few). Number go up? Yay, when lambo. Number go down? Lives get wrecked. Where's the contribution to humanity?

Yes, it takes time to "build a new monetary system". For sure... but how about we make a few useful things along the way? Everyone is so focused on becoming the next V that nobody wants to build useful stuff, everyone's in it for value extraction.

I don't really see *why* this system should not collapse? What would we lose in the grand scheme of things? All crypto achieved was a huge value transfer, but I fail to see any value creation thus far.

Mind you, making it big by luck and moving your funds into some non crypto venture is not a success story for crypto. It at best is a success story for a philanthropic lucky person, but more likely a simple diversification of said person.

IMO it's past time this industry creates something genuinely useful that people want to use, or should just close up shop.

Bitcoin at least tries (and fails) to be a safe haven asset. But the rest are all selling shovels with no gold rush in sight.

[Sid Prabhu @sidprabhu]

Remember 2 years ago when every perma dove was like "but how do interest rates stop inflation"? This is the answer.

[Ian Cassel @iancassel]

A big way to separate yourself from others is sticking up for people you know when you hear others gossiping about them. You will earn the respect from those gossiping even though they won’t admit it and build a reputation as someone that sticks up for people/honors friendships.

[Koyfin @KoyfinCharts]

After first announcing the purchase of five large Japanese trading companies in August 2020, Buffett's positions have done quite well.

Here are their returns since August 2020 in both Yen and USD terms.

[Nick Huber @sweatystartup]

These pictures were taken in Austin Texas during business hours.

If this is your work-from-home schedule I have terrible news...

There are people in the Philippines and LatAm who can do your job better than you can for $5 an hour.

And work 60 hours a week, no complaints, and with AI their English is better than yours.

It is the worst case scenario for you and your $200k liberal arts degree.

American companies are learning they can get 3x the productivity for 20% the cost.

80% of my workforce is in LatAm and the Philippines.

It is a business cheat code and it is about to change everything for remote workers.

If you aren't making decisions, leading people, selling something or generally adding a ton of value to your company - these people are coming for you and your job.

[Stat Arb @quant_arb]

It's often joked that you start off doing arbitrage and end up as a market maker.

Here's my take on why.

A thread 🧵

[1/14]

[David | Filled With Money @FilledWithMoney]

76% of NVIDIA employees are millionaires.

Is it time to apply?

[Andrej Karpathy @karpathy]

These 94 lines of code are everything that is needed to train a neural network. Everything else is just efficiency.

This is my earlier project Micrograd. It implements a scalar-valued auto-grad engine. You start with some numbers at the leafs (usually the input data and the neural network parameters), build up a computational graph with operations like + and * that mix them, and the graph ends with a single value at the very end (the loss). You then go backwards through the graph applying chain rule at each node to calculate the gradients. The gradients tell you how to nudge your parameters to decrease the loss (and hence improve your network).

Sometimes when things get too complicated, I come back to this code and just breathe a little. But ok ok you also do have to know what the computational graph should be (e.g. MLP -> Transformer), what the loss function should be (e.g. autoregressive/diffusion), how to best use the gradients for a parameter update (e.g. SGD -> AdamW) etc etc. But it is the core of what is mostly happening.

The 1986 paper from Rumelhart, Hinton, Williams that popularized and used this algorithm (backpropagation) for training neural nets:
https://t.co/f52IcDNitR
micrograd on Github: https://t.co/GaTd16jRnB
and my (now somewhat old) YouTube video where I very slowly build and explain:
https://t.co/EPGG6kd5Yz