Situational Awareness, run by 24 year old former OpenAI researcher Leopold Aschenbrenner was forced to liquidate its public equity portfolio to cover losses after a recent selloff in AI stocks put the fund at risk of insolvency.

The fund which was reportedly up as much as 439% by the end of June, dropped 67% within a few weeks, triggering a margin call, forcing them to sell the majority of its public equity book, to Citadel, the hedge fund led by Ken Griffin. Non-finance folk might recall the firm from the Gamestop saga of 2021.

The public hasn’t been kind to Leopold. Many called him an idiot, ignorant to financial markets, dismissing his prior success as luck etc. This is coming almost exclusively from people who never met him. It’s true he’s new to managing money but he graduated from Columbia at 19, worked at FTX (before it blew up), OpenAI (before he got fired) and is connected with many of the key figures at the AI frontier labs. I doubt his critics can boast a similar track record by the age of 25, or at all.

Leopold is the latest in a long line of fund managers that posted stunning returns for a while before their investment strategy succumbed to the wrath of Mr. Market. Whether it’s Bill Hwang, Cathie Wood, Sam Bankman-Fried or Bernie Madeoff, every few years, the financial press crowns a new Warren Buffett, only for their funds to suffer devastating losses. In some cases blowing up the fund and/or sending the manager to jail. It’s almost a guaranteed kiss of death at this point, like the Forbes 30 Under 30 list.

Nobody has accused Situational Awareness of doing anything shady or illegal but there are a few lessons that can be learned from this episode, some of which were predictable.

Margin Calls & The Situational Awareness Strategy

“Illiquidity, leverage and concentration. Each are a problem but manageable in isolation. If you have two of them, uh oh. If you have all three of them then you are whistling past the graveyard.” — Steve Cohen (sourced from Marc Rubinstein at Net Interest)

Why was the Situational Awareness blow up predictable? They were heavily concentrated into AI Infrastructure names with meaningful leverage. Beyond that, the portfolio was designed to perform well during a looser credit cycle, once money began to tighten, it was inevitable that some of the positions would be negatively impacted. The US-Iran conflict has reduced the supply of energy which has been putting upward pressure on inflation and interest rates which shifts demand for near term cashflows.

The reason Leopold needed to sell his public book as quickly as he did, was because his investment portfolio was trading on margin, which meant he borrowed to place trades. It’s not that different from taking a mortgage.

You purchase a $1 million property with a $200,000 down payment and an $800,000 mortgage. You control a $1 million asset using only $200K of your own capital. If the property rises 20%, you pay back the $800K and your equity is worth $400K. If it falls 20%, the value of your equity is worth zero because the lender is still owed $800K.

The difference is that your bank doesn’t revalue your home daily. Whereas a prime broker continuously monitors the value and volatility of an investment portfolio. When the collateral drops too far, the broker will demand additional capital, reduce the amount it’s willing to extend or begin liquidating positions.

Situational Awareness reportedly borrowed $3 to $4 for every $1 of capital. In that case, a decline of 20% to 25% would wipe out all of the fund’s equity.

Situational Awareness was running a Long-Short (LS) portfolio buying companies exposed to AI infrastructure like Micron, Sandisk and Coreweave and selling short software companies such as Adobe, Salesforce and Microsoft. This strategy was working well throughout 2026, as many piled into the AI trade. However, in July, due to a number of factors, including concerns about higher inflation resulting from the US-Iran conflict, some of these high-flying names started to give back prior gains.

Any stock that can quickly rise, can fall just as fast. It’s normal for stocks trading based on the presumption of large growth relative to current earnings to ping-pong back and forth as market sentiment changes. The AI infrastructure piece of the Situational Awareness portfolio was trading based on the belief of large growth in future earnings, making it highly volatile. Meanwhile, the short side of the portfolio, were mega cap established companies that are comparatively less volatile. They were trading down because of concerns that AI was going to be disruptive to their business models in the medium-long term.

Let that sink in. The fund was long companies whose current earnings could only justify their valuations if the future arrived quickly. At the same time, it was short companies whose current earnings would collapse if the future arrived quickly.

It’s the same bet.

When inflation expectations and interest rates rise, investors become less willing to pay for profits expected far into the future. The infrastructure stocks fell. Meanwhile, software companies continued reporting real revenue, earnings and showing no collapse in their business models. Their prices were rising against the fund as well.

Leopold could not have predicted all the developments of the Iran conflict or the precise moment investors would turn against AI stocks. Though he should have realized that a highly leveraged and correlated portfolio would eventually experience a period when several positions moved against it simultaneously.

His call about the scale of investment required to build artificial intelligence could still prove correct. Unfortunately, when leverage is involved, being right isn’t enough. You have to remain solvent long enough to collect the payoff. Leopold made several unforced errors, which threatened his staying power. Once it was too late, the sharks knew he was ripe for the picking.

Ken Griffin Plays Hardball

Ken Griffin has been a major player in the hedge fund world since before Leopold was born. One of Griffin’s favourite books is Hardball: Are You Playing to Play or Playing to Win? This is telling to the way he sees business.

Most people in tech subscribe to the Peter Thiel view on competition from his book Zero to One: Competition is for losers. Thiel isn’t anti-competition but he observes that competition benefits consumers but is often destructive to shareholder value. The most valuable businesses in history have exercised some form of monopoly therefore, founders should strive to build one ideally by creating such a strong technological and distributive advantage nobody can threaten them.

After reading Hardball, you might wonder if Thiel wrote Zero to One as his response. Hardball accepts competition as an inevitability. It doesn’t discuss monopoly businesses, instead looking at companies that successfully defeated competitors by what the authors call playing Hardball. There are a few ways a company can play Hardball, but they all require being hyper aware of what your competitors are doing, so you can figure out where you can attack them to take market share. These can be larger companies trying to shake off upstarts or smaller companies going after the market leader. Being a Hardball leader and company, means you exploit every legal advantage you have at your disposal. This means hitting the competition, usually with indirect attacks, that put them under pressure, while being ready to fend off any of their retaliations. Citadel was built for the Hardball world.

It’s not clear how personally involved Griffin was in the purchase of Situational Awareness’s equity portfolio. However, the transaction reflects the Hardball culture he created. Leopold had spent the previous two years attracting enormous attention to himself and his strategy. His returns inspired an army of professional and retail copycats to pile into the same AI infrastructure companies. The more crowded the trade became, the easier it was for Citadel and other sophisticated firms to understand what would happen if those stocks began falling.

Situational Awareness had large, highly visible positions, meaningful leverage and a relatively inexperienced manager. Citadel had the size advantage and decades of experience profiting from distressed situations. You could not design a more attractive hardball opportunity. When the margin call came, only a handful of funds could take such a big position on such short notice. In came Citadel buying the book at a 10% discount, instantly printing an easy couple of billion dollars.

This is hardly the first time Citadel has mobilized quickly during a crisis. When Enron collapsed in 2001, Griffin chartered a plane and sent executives to recruit its energy traders. Situational Awareness was merely the latest distressed competitor to discover that Griffin is rarely sentimental when markets become chaotic.

Shortly before Citadel purchased the portfolio, Citadel Securities had publicly predicted that the Federal Reserve might surprise markets with an interest-rate increase. Higher rates would have been particularly damaging to the long-duration AI stocks Situational Awareness owned. Griffin wouldn't purposefully try to hurt an exposed competitor while they were already down would he?

(If you’ve made it this far, please settle a debate for me in the comments: Is a Beef Wellington considered a sandwich? (Yes/No and Why?))

Leopold is a new investor, he might be smart and right in the long term about his AI thesis but he structured his fund in such a way that it didn’t need a Black Swan event for him to get margin called. They claim to have hedges in place but the problem with hedging, is you can almost never perfectly hedge against every risk, if you do, it will come at the expense of returns. You can’t be doing 400% returns in 6 months while being perfectly hedged.

Trading on margin is commonplace, it’s done by most hedge funds and many retail investors, so is getting margin called. Over a million South Koreans got margin called in July alone. What’s notable is Leo flirted with liquidation while the S&P 500 is just off its all-time high. This wasn’t a market panic, this was sloppy fund management.

Situational Awareness was concentrated, leveraged and reliant on loose credit conditions to succeed. A bad month was inevitable. Leopold also overlooked that when you are trying to beat the market, you aren’t just going up against a passive index, you are competing with every other fund manager out there. Situational Awareness was exposed, and a Hardball competitor would not let that slide. The hedge fund world is full of Hardball players.

Despite an embarrassing episode, Leopold Aschenbrenner isn’t finished. While he had to sell off his public book, Situational Awareness still has a $5B stake in Anthropic. Leopold remains connected to many of the key players from the AI labs. Silicon Valley and he’s still at the very start of his career. Ray Dalio and countless other investing legends had trades go south on them early in their career and they were only the better for it.

He might have lost some of the boy wonder aura but people have short memories. Masayoshi Son has lost his fortunate and rebuilt it multiple times. Aschenbrenner might have prematurely claimed to have been gifted with situational awareness, but that does not mean he can’t eventually become one of the best investors of his generation. It just takes more than a coronation from CNBC for it to be true.

Thank you for reading. If you liked this article please Like, Comment and Subscribe below.

Subscribe now

Full-blown #REF! errors everywhere. Similar experience to when you ask the MBA associate to update one row of actuals. Just without the tears or over applied cologne. Yet I was still impressed.

Since the release of OpenAI’s ChatGPT in late 2022, we’ve heard endless predictions about AI automating knowledge work. Copywriting. Translation. Coding. Now, with the arrival of Anthropic’s Claude for Excel, the finance function is next.

Don’t worry. Beloved Investment banks, hedge funds and private equity firms won’t go anywhere. So long as long as fees exist, they will too. What will change is the work performed by overworked and underdeodorized juniors.

For decades, junior finance professionals spent countless hours doing grunt work before shipping deliverables to clients. Not without a half dozen pls fixes from seniors first. With Claude, what previously would have taken days and weeks, can be condensed to hours and minutes.

This won’t immediately eliminate jobs but it will shift value from execution to judgment. Building pretty models and slides won’t be enough to keep you employed. Sector expertise, unique insight and intellectual rigor will come at a massive premium. More than it already does.

This is not unprecedented. Before Microsoft Excel analysts had to run DCFs by hand with calculators and paper. Yes paper. Excel did not destroy finance. It compressed time spent on calculations and raised expectations.

Claude for Excel will do the same.

After a week of testing it inside real models, I’ve seen what it can and can’t do. It’s not as good as an experienced finance professional that knows what they are doing, but it’s at least 80-90% of the way there. Will it ever get that final 1%? Only time will tell.

In the meantime, keep reading if you want to learn what Claude for Excel can do and why you shouldn’t hand the keys over to the model completely unsupervised yet.

If this week’s article does not interest you, please check out some other recent ones:

Automation Risk: Computers Changed Finance but Did Not Lead to Mass Unemployment

The introduction of software to finance didn’t reduce employment, it increased it.

If you take a narrow definition of finance, in 1960 there were approximately one million professionals working in commercial banking in the United States. In 2020 there were 1.4M.

Employment has oscillated up and down with the business cycle but that number peaked around 1990. There were around 1.6M then. However, if you expand the definition there are 6-7 million people working in financial services today, millions more than there were 35 years ago.

A similar number will still work in finance 35 years from now. Just the industry and the roles will look very different. For one thing, many back office and compliance tasks, can already be automated. These kinds of roles can represent half the positions within the financial services sector, and many were created since 1990. Largely because of increased reporting and compliance requirements, especially post 2008.

Somebody in 1990 after seeing Microsoft Excel for the first time could not have predicted this. Similarly, most people will point to this capacity to automate as a key driver leading jobs to disappear. People ignore banks fire people and cut divisions all the time, usually for reasons unrelated to technology.

Nonetheless, in the short term, I would expect financial institutions to hire less for existing positions, but over time new roles will be created.

In fact, I predict that within 12 months major financial institutions will begin hiring for teams focused on prediction markets. Given the explosive growth of Polymarket, Kalshi and others, it’s not a stretch to imagine bets might one day get similar coverage to stocks and bonds.

Don’t believe me? Mergers and acquisitions and low grade debt issuances (junk bonds) barely existed before the 1980s. They became bigger fee generators than equity research or public issuances. More recently, private credit has seen its AUM explode from $100B in 2007 to $3.5 Trillion in 2025.

Things can change quickly, so I’m not too worried about Jamie Dimon and the next generation of charismatic bank CEOs.

(If your intern isn’t this deep into the model, they aren’t getting a return offer)

Using Claude for Excel

You need a Claude Pro plan, free one won’t cut it I’m afraid.

Next you need the full version of Microsoft Excel, sorry Google sheets and Numbers users. You shouldn’t be using these to model anyway.

Once you download Claude for Excel, you just need to make sure it appears as an Add-in the top right on your Home ribbon. With that you can start prompting.

To test it, I asked it to build me a fully integrated three financial statement forecast. This is a common case question you get when interviewing for investment roles. Depending on the questions you can get anywhere from 30 minutes up to a few days or even a week to complete, sometimes with an accompanying presentation.

I just wanted something dynamic that balanced. It took only five minutes to complete the task. Without copying from an existing template, there’s no way a human could do it this fast.

At the peak of my powers, I was probably doing this in 20-30 minutes. If I wanted to add scenarios, value drivers and other fancy stuff maybe closer to 45 minutes if I didn’t care about formatting. Naturally most finance hardos use plug-ins like Thinkcell, Macabacus and other tools to help with formatting and manual work but the point remains, being able to do this in 5 minutes is wild.

Yet I was still skeptical, I tried a few other tools in the past year that could also build models in minutes. They worked fine until you gave them real data that wasn’t neatly organized, then the hallucinations would start.

I added a few tabs of raw data pulled from Quickbooks (financials), Ramp (spend data), Hubspot & Stripe (sales & payments), with minimal enrichment. With this, I asked it to build a 12 month forecast across the 3 financial statements. I had done the same exercise a few months prior and it was weeks of work on and off.

Within 15 minutes, during which time I could make a cup of tea and fold some laundry, Claude produced a forecast that was pretty good. During this time it used other tools such as Python to analyze the large quantities of data. The net effect was it outputted something around 10% of what I had forecasted for most values. Keep in mind, it didn’t have context beyond the raw data, which limits effectiveness.

For example, it overstated some expenses because it couldn’t go into Ramp and see that the latest invoice was quarterly not a single month. It also couldn’t have known certain details around accounting adjustments or reversals.

Producing this from a few simple prompts was as good as you could have hoped for. I tried to add a few corrections to get it closer to what I had done, but this started to push the limits of Claude. The formulas weren’t as dynamic as I would have liked. I started to see a bunch of hardcodes and manual in cell calculations, something only noobs would do.

As I fixed some sections, others stopped making sense. I would decrease the projected revenues, but gross margins and sales commissions weren’t changing. Upon closer examination, it’s because these sections weren’t tied to schedules or driven off their own assumptions. Another intern move.

I asked it to make everything dynamic and build out supporting schedules, it fixed some of these errors. After an hour of this, the model wasn’t near finished but it was much further than if I was building it normally.

Next I asked it to build me some dashboards and charts, which it did fast, but the charts were the basic excel pre-built ones and while the dashboards were nicely formatted there was some unformatted data below the tables. When I asked it to clean it up, Claude just hid those rows. Not what I was looking for.

To further test it, I gave it part of a model I built. It was super granular, about 600 rows deep with a lot of formulas, referencing other tabs and multiple scenarios. It was a little bit messy because there were a number of schedules I had built but later decided not to use. I asked Claude to remove rows that were no longer in use, or tied to other rows in the tab.

This is where all hell broke loose.

It took 10 minutes to review the tab, then began to do the cleanup as instructed. It stopped to ask questions along the way and things seemed to be going fine. Then suddenly I saw the #REF!s appear. For the uneducated, a #REF! error happens when a formula is trying to reference a cell that doesn’t exist or you’re using circular calculations. This usually appears when you delete something you shouldn’t, which Claude was in the process of doing.

@fentifriedchickenMBA Interns: the most competently incompetent! #Working #Corporate

Tiktok failed to load.

Enable 3rd party cookies or use another browser

I pointed these out and it took a few minutes to fix them but it created a few more in the process. When I pointed these out Claude started spazzing. It said I might have to refresh the window.

At the moment I noticed in my company Slack page our Head of Operations was asking who breached their limit on Claude? This was a regular occurrence across the team, but usually from devs using Claude Code. I had breached my limit in two hours.

This ended my experiment for now but it was enough to demonstrate what Claude for Excel could and cannot do. It’s great at doing the grunt work in setting up a model, but it’s not ready to run it unsupervised for now. Too often it forgets to ask itself the key question anybody who works with numbers or data for a living should wonder “Do these numbers make sense?”.

This is what separates good decision makers from bad ones. Anyone can see if numbers balance, but figuring out if the output is realistic is what matters. If you are going to make important choices based on what these numbers tell you, they need to. Before making an investment, an acquisition, hiring or firing spree, you need to be as sure as can be.

This is an instinct few junior finance employees have. It takes repetition, experience and a certain generalist skill set to sense when numbers are off. This is rare in most organizations, but it’s likely seen either at the executive level or people with hands on operational experience.

You can train Claude on as many models as you like, but it won’t build this instinct. People with this ability will become increasingly valuable and productive. As they offload more work to Claude, they can spend more time digging into validating key assumptions. Or more importantly, improving them.

This will lower the barrier to entry for areas of finance that were historically reserved for people coming out of investment banking, private equity or consulting. Modeling ability acted as a gatekeeper, which screened out potentially great people with strong business judgment.

With tools like Claude for Excel, that dynamic begins to change. The mechanical barrier weakens. Instead of evaluating somebody on their ability to build a model, the shift goes towards understanding how a business works, and what questions should be asked? The intellectuals among you might even ask if financial models will even be a thing for much longer. The answer is yes. Excel will be the last software to go. It’s the way we use it that will change.

The net effect is smart, hardworking people who understand economics, incentives and strategy can use AI to close the execution gap much faster than before. Modeling skills will be deemphasized and judgment becomes the moat.

Thank you for reading. If you liked this article please Subscribe below. I publish articles on a wide range of topics from business, books, current events or anything on my mind.

Subscribe now

Is it their paint brush collection? The loft in Soho? A middling drug addiction?

The ability to create something out of nothing?

These are (mostly) nice things, but to be an artist, you only need:

1. The willingness to create

2. The right taste to know what will resonate with people.

For most of human civilization even if somebody had these qualities, their chances of being an artist were low. Leisure time is a relatively new concept that only a few generations, beyond the very wealthy, could benefit from.

Even if you had free time, most art forms were not accessible, because of high costs and barriers to adoption. As expertise became more widespread and advances in technology (water colors, cheaper instruments, editing software, etc.) permitted more people to participate and discover, these art forms evolved.

In any domain, as more participants enter, competition ensues leading to better products. If more people participate in art creation, it will get better.

AI and Large Language Models (LLMs), will allow many more people to participate. While most people are focusing on how AI will change the future of work, not enough attention is being paid to how it will create a new artistic renaissance.

If you don’t read another word of this article, let this be your takeaway: LLMs will do more to lower barriers and encourage more participation than any previous innovation, which will increase the quantity and quality of art.

Find out why in this week’s article.

(Click here for a primer on how LLMs work, Why you should be using them and How it will change The Future of Work )

But first, make sure to hit the Subscribe button below to join the nearly 1,000 Subscribers that get articles like these, delivered directly to their inbox.

Subscribe now

If this week’s article does not interest you, please check out some other recent ones:

The Rise in Leisure/Free Time

Leisure is a new concept. As late as the early 20th century, the average life expectancy was between 50-60 years old, meaning you didn’t really retire. You worked until you died if you weren’t killed sooner in battle, nicking your finger on a rusty spoon or during a hunting accident at the hand of a wild boar.

Those that didn’t die young, had very little time beyond work, household chores, childcare and sleep. As such, free time in most countries was really only available for the very wealthy, which numbered very few.

This lack of leisure time meant art creation was limited to full time professional artists, which there also weren’t many of. Even back in those days, the average artist was probably a starving one, and not the kind that secretly gets their rent and credit card paid off by their patrons, also known as mom & dad.

There was only tens of thousands professional artists in Western Europe during the 16th century. This didn’t meaningfully change until the late 19th and early 20th century, when this moved closer to 500,000 - 1,000,000. The big jump really happened in the late 20th century, due to the emergence of modern creative industries (film, radio, recorded music, television) which caused an exponential rise. In the U.S. alone, the number of professional artists jumped from ~750,000 in 1970 to ~1.7 million by 1990. In 2025, there are ~10 million professional artists in North America and Europe.

When you consider amateur and hobbyists, this number is way higher. According to the National Endowment of the Arts, in 2022, >52% of U.S. adults reported personally creating or performing some type of art (painting, music, writing or crafts) in the past year. It was a similar rate in Europe. This only includes one instance of art creation within a year, so it’s not much but it does demonstrate, many people at least occasionally participate in the art creation process. Due to the increase in life expectancy, leisure time and relative wealth, all of this makes sense.

Now you might be thinking, “Ok Ben, you’ve convinced me more people participate in art than they did 100 years ago. Good job. So what?”. This matters because, if more people have the time, willingness and tools to create art, more art will follow.

New Artist Formula = Taste + Willingness

“If you had one shot, or one opportunity. To seize everything you ever wanted in one moment. Would you capture it or just let it slip?” — Marshall Mathers (Eminem)

Not long ago, many artistic fields were structurally inaccessible. Few had access to professional equipment, teachers, or the opportunity to showcase their talent. Despite what older people say, art today is better than it ever was. Of course, there were exceptional artists from every century that would rival anyone today but never before has there been such variety of works and the ability to more easily discover them. Citizen Kane and Casablanca are incredible films, among the best ever, but compare other works from that era against the last 40 years of cinema, and it’s clear modern films are better. This is because of technology.

Before Oppenheimer, Tenet, Inception or The Dark Knight, Christopher Nolan spent most of the 1990’s working a full-time job at a corporate video production company, shooting training videos and industrial films. He would free lance doing camera work, editing, and small-scale production gigs.

On weekends with a group of friends and no money, he would shoot films on a 16mm camera and DIY lighting, this produced Following (1998), his first feature which cost about $6,000 to make. It managed to get into film festivals where it impressed critics and potential investors. On the heels of this success, he pitched his next project, Memento. The film got financed and made which cementing Nolan’s place in Hollywood and it is considered one of the best films of its era.

I can write an entire article on Christopher Nolan films, and I’ve come close but I use this anecdote to point out, that Nolan was only able to make Memento because he had made Following. Without the equipment, sound and editing tools available to him at the time, he wouldn’t have been able to. Nolan was able to showcase his talent because the technology made it possible to create an inexpensive proof of concept.

Technology reduces the friction standing between the idea inside an artists mind and the canvas. Despite the constraints, Nolan had to work with, he could still make a good film. As film studios increasingly gave him more budget and creative control, he could attempt more ambitious projects, which resulted in masterpieces.

Now consider this. What if anybody could produce a film, song or work of art without needing the resources of a film studio, music label or publishing house? If everybody had access to a world-class teacher/mentor that was available 24/7 at a negligible cost? Do you think we would have more Christopher Nolan level talents, or less?

Why AI will create great art

“Are you ever going to come back from the dark side and shoot on film again?” Christopher Nolan

“When you start writing your scripts in pencil.” Steven Soderbergh

The Gutenberg press, developed in 1440, didn’t just invent mass printing, it enabled mass literacy. By dropping the cost of books by over 90%, it made reading and writing accessible to the public. Within a century, the number of literate people in Europe increased nearly a hundredfold.

Stories existed before then, of course but outside wealthy elites, books were rare, expensive, and regionally locked. The printing press removed that friction and reshaped science, literature and culture forever.

AI can have a similar effect. For starters, it gives anyone access to knowledge, guidance, and feedback once reserved for formal training or elite mentorship. Ask ChatGPT how to become a better painter, and it will give you instructions on par if you went to a beginner or intermediate art class.

If you tried this for other disciplines, I’m sure you would get similar results but beyond access to information and getting feedback, AI goes even further. If you haven’t played with Sora, Midjourney or some of these other photo or video generation tools, take a look. Today their functionalities are still somewhat limited, but soon enough it will be realistic to believe you can create an entire feature length film, similar to Following, just using prompts.

With zero budget, you could create a proof of concept for the film you want to make. If you were a talented enough, maybe you could create an entire film, good enough for distribution this way. Imagine how many more people could participate in the film industry if they didn’t need to worry about costs, location, acting talent, lighting etc. This wouldn’t replace the epic films done by professionals such as Nolan, Tarantino and others but before long, I expect even Hollywood directors will be using these tools to shoot stunts too expensive or difficult to film in real life.

If you are still reading this, I imagine you might disagree with some of this. That’s fine. We are still in the early days of LLMs and AI adoption. You might argue that AI will never produce great art because it doesn’t feel. It’s not conscious or sentient. Perhaps you are right, that feelings are necessary and you need to live through heartbreak or obsession to make something beautiful. Rick Rubin and others might say as much. Many artists are emotional, erratic, deeply sensitive people.

This might have been a good argument but the inability for AI to feel is irrelevant. It’s the person prompting the AI that matters. You can give any person a camera, it doesn’t mean they will take a good photo or shoot a decent film. If a person with taste and the desire to create, uses AI and it will only enhance their work.

Are you still skeptical? Ok. Let’s assume you don’t believe that modern film editing software, sound & beat packages, autotune software, spellcheck and other tools have helped artists create better art. Think about how the internet reshaped the way we discover art. Substack didn’t exist a decade ago, now you’re reading this newsletter here. Justin Bieber, Billie Eilish, Post Malone and countless other artists were discovered on platforms such as YouTube, MySpace, Spotify and SoundCloud. Some probably could have made it without them, but I expect many wouldn’t have.

This is why the notion that AI cannot create art is a flawed premise. All modern art is created or supported by technology and not only have these tools democratized these art forms, they have increased the quality. AI will do more to reduce friction, permitting anybody to create art. Which is why the next generation of great artists, will be those that learn to combine their unique taste and creativity, with the possibilities enabled by AI.

Thank you for reading. If you liked this article please Subscribe below. Each week I publish an article on a wide range of topics from business, books, current events or anything on my mind.

Subscribe now

This probably feels like ancient history. It’s not. In the year 2000, less than half of American adults were online. It wasn’t until 2018 that global internet usage crossed the 50 percent mark. So if you’ve been using the internet regularly for 20 years, you were an early adopter.

Even crazier to think that most of the apps, habits, and workflows we consider as the default only became mainstream in the last decade. If ride-sharing, social media, streaming, and cloud storage, were alive they still couldn’t legally drink in the US. Despite this, odds are you're reading this on your phone right now either through a browser or the Substack app.

Technology adoption works the same way as bankruptcy. Slowly at first, then suddenly all at once. This is why the internet has massively changed the way the average person lives over the past few decades; unless you are an Amish person on Rumspringa, I probably don’t need to sell you on the benefits this technology offers.

Where you might need a bit more convincing is why you should be using Large Language Models (LLMs) (click here to learn how they work). They are still early in their adoption but their impact is expected to be at least on par with the internet. This is why not using ChatGPT, Gemini, or another LLM on a daily basis, is like ignoring the web in 2000. You can continue to live your life without it, but you are sacrificing meaningful benefits that extend beyond party tricks.

You will not unlock them from Google search or the old ways of the internet. Pretending like this new technology will come and go like NFTs or AR/VR , will make you like the Boomers that never learnt how computers worked so they became reliant on broccoli haired Gen Z interns to save file as PDFs.

This is why if you are not already using LLMs on a daily basis in 2025, you are slowly making yourself unemployable and less productive. Start using LLMs or your professional life will suffer.

Find out why in this week’s article.

But first, make sure to hit the Subscribe button below to join the nearly 1,000 Subscribers that get articles like these, delivered directly to their inbox.

Subscribe now

If this week’s article does not interest you, please check out some other recent ones:

The Current State of LLM Adoption

LLM adoption is much more widespread than you think.

In February 2025, ChatGPT had over 400 million weekly active users (WAUs), this jumped to 800 million by April. I could not find any evidence that another product has ever managed a similar feat that quickly. As a frame of reference, companies such as X/Twitter, Spotify and Snapchat have between 500 - 750 million monthly active users (MAUs). For most apps a strong ratio between WAUs to MAUs is generally between 20-25% (4:1 or 5:1), but for high engagement platforms like WhatsApp, Instagram, Youtube etc. this can range from 50-70%. Therefore it’s likely that ChatGPT’s monthly active user number is starting to approach Web2.0 consumer tech giants such as Youtube (2.5 billion MAUs), Instagram (2.4 billion) and TikTok (1.69 billion).

You might be wondering, who are these users? Is it just young people cheating on their homework? Surprisingly, according to some surveys, LLMs are seeing similar adoption rates from all age groups with the exception of 55+. ChatGPT as the first mover naturally has the highest usage rate, but the other most used products are owned by the big tech companies such as Gemini (Google), Copilot (Microsoft) and Llama (META). Only the power users are working with models from Anthropic, Perplexity or models from another pureplay LLM company.

This is not all that surprising, given how much mainstream press coverage Open AI has gotten while Google had early difficulties with Gemini. For the people using basic chatbot functions, the model performance doesn’t seem to matter as much so they are happy trusting Open AI or one of the big tech companies. Since these products have memory, they should get increasingly sticky and more useful the longer they are used.

On the other hand, those using LLMs to automate workflows or for specific niche tasks, will be more particular. Despite ChatGPTs advantage, this can fade quickly. During the early days of the internet, Mosaic then Netscape went from 100% market share to being replaced by Internet Explorer (Anti-Trust Bill’s prime) in about 4-5 years.

Speaking of the internet, it took almost a decade for nearly half of American adults to get online while relying on broadband and needing to download AOL or some other software to access it. In only about 2.5 years, ChatGPT and LLMs have already achieved a similar penetration rate with much smaller hurdles for continued adoption. Even beyond the consumer web, seemingly every company needs to come up with a response to the question “What is your AI strategy?”. The fine folks at McKinsey report 78% of organizations claim to be using AI in at least one business function, up from 72% in 2024 and 55% in 2023.

Hopefully this section was enough to convince you that LLMs are no longer a niche thing for young people afraid of social interaction. More people are using these tools every day so choosing to ignore them will put you at a disadvantage.

How LLMs can help your professional life

Unless you are very inefficient at your job, LLMs won’t help you accomplish 10X more work in half the time, as promised by some AI efficiency gurus. This does not mean they aren’t useful though.

Salespeople working in B2B SaaS make all kinds of grandiose claims their products will never be able to back up, but companies still spend trillions of dollars on these tools because they solve problems. Seeing as the average while collar professional spends about half their work week writing emails, working on memos/presentations or doing research, even slight efficiency gains can translate to major time savings.

This is why the low hanging fruit for LLM adoption starts with producing text based content. With a few prompts, you can producr great how to guides, FAQs, emails, internal memos, presentations or other time consuming internal documents. Anyone that has needed to create a playbook for new hires understands the pain of building these from scratch. Realistically, for many requests, asking ChatGPT to work on the first draft will get ~80% of the job done right there.

Although it’s usually the remaining 20% that takes the most time, it’s helpful to offload low value added tasks and focus on the difficult stuff. It’s like having an eager intern at your beck and call day and night. Except this intern doesn’t need to get paid or occasionally throw up on the CEOs shoes. This is really how you should think about these models right now. They are great for basic, low consequential stuff but as you start to use it for more important tasks, a higher level of scrutiny is needed.

I’ve been regularly testing ChatGPT over the past year for my own workflows and I can say, the product has really come a long way, but it isn’t ready to replace me (yet). When I would pose it basic cap table questions, it would hallucinate responses or confidently tell me things that were completely wrong. I would ask things such as “If we raised another $50M in a round of financing, valuing the company at x millions of dollars, what would be the new price per share?” it would return a response that would look accurate. Yet when I followed the methodology, it was immediately clear they weren’t using the correct formulas or making up new assumptions.

When I tried again recently, it did a better job but you should still definitely not trust a chatbot for something like this unless you are already an expert in this domain. The false sense of confidence it gives is deceptive, so need to know enough about a topic to spot when it makes mistakes, or you will follow bad advice and get into trouble.

Getting answers with high confidence but inconsistent accuracy is the most dangerous combination, which is why you need to ask yourself if these answers make sense and regularly audit responses.

Where are all the developers?

Outside of drafting text for white collar workers, another area that people believe is being hyper charged is software development. In recent weeks, Google and Microsoft have claimed AI writes somewhere between 20-30% of new code. They are embracing this since these companies are under pressure to accelerate new product releases with fewer developers. This is consistent with the many layoff and downsizing initiatives we’ve seen in Big Tech these past few years as they attempt to show the stock market they can continue to grow quickly with good margins.

Even early stage startups have been forced to embrace capital efficiency and operate with fewer employees now that there has been a major pull back in venture capital activity (outside of buzzy AI companies). Whether this can be attributed to efficiency gains from LLMs or not, there are now >200% fewer software development job postings on Indeed in the US compared to at the peak of early 2022. Looking back to February 2020, there are now 36% percent fewer roles compared to then.

If you ask Salesforce CEO Marc Benioff, it’s definitely because of AI tools that are boosting engineers productivity. In December 2024, Benioff said these tools were boosting engineering productivity by over 30%, prompting them to freeze hiring for software engineers. More recently, Shopify CEO Tobias Lutke issued a leaked company-wide memo emphasizing the higher use of AI in company operations. The use of AI is now a “fundamental expectation for all employees”, including Shopify execs. Lutke mandated that before requesting additional headcount, teams would have to demonstrate that they could not achieve the same outcome with AI.

What we have learned so far is that using AI well is a skill that needs to be carefully learned by… using it a lot. We also learned that, as opposed to most tools, AI acts as a multiplier. For the first time, we see the tools become 10X themselves. I’ve seen many of these people approach implausible tasks, ones we wouldn’t even have chosen to tackle before, with reflexive and brilliant usage of AI to get 100X the work done.

— Tobias Lutke April 7 2025 (You can read the full memo here)

This was directed at all employees of Shopify, not just the software developers. While Tobias was delivering this memo, another CEO was making a similar warning to his employees. Micha Kaufman, the CEO of Fiverr, believes AI is coming for jobs, including his own. It is clearly the time to start learning how these tools can help you do your job better and faster.

Conclusion

Using LLMs today feels a lot like relying on Wikipedia in the early days. For certain topics, it was a game changer (fire). For others, complete nonsense. Even if it occasionally got things wrong, it was the fastest way to get answers. With basic judgment and a few minutes to cross-check, it could save you hours.

That’s where AI is right now. Great for for low-stakes, high-volume tasks like summarizing meetings, drafting memos, or formatting. This alone can save most people hundreds of hours a year. For more critical tasks, we are not yet at the point where we can rely on autonomous agents to handle execution end to end. For the time being AI should continue to play the role of helpful assistant. So for now, continue to regard it as an eager and bright junior employee, who knows a lot of helpful facts and skills but struggles at time to be forthright when they aren’t sure about something.

Soon enough though, like a junior employee, they will be useful enough to handle things alone. This is why knowing how to prompt, while important is not the only skill you need to unlock major efficiency gains from AI. You need to be able to exercise judgement, for when you can automate things, when you can trust the answers you get and when some responses appear off. If you are sufficiently advanced in your domain and motivated, adding this will help you reach even another level.

This is why forward thinking companies are no longer sitting around a board room discussing these things conceptually, they are being used or tested in offices today. The time to start using LLMs is right now. Those that started using the internet early, gained an advantage over laggards. Don’t be a laggard. You have enough time to learn how these tools work, and not have to rely on your Zyn addicted junior. Instead you can soon have an army of autonomous agents at your disposal that can replace most of work of your entry level hires, without the attitude (or smell) and most importantly, finally create lasting and enduring shareholder value.

Thank you for reading. If you liked this article please Subscribe below. Each week I publish an article on a wide range of topics from business, books, current events or anything on my mind.

Subscribe now

Upon reflection, I saw a strange parallel. The Trial depicts a society where individual liberties have been replaced with an ever-growing bureaucracy, designed to further perpetuate itself, at the expense of everyone else. The Technological Republic, on the other hand, argues that this is what our society will become if we do not stand up for individual liberties, which are among the hallmarks of Western civilization. The way we defend these liberties, is technological and military superiority.

Translation: He wants you to buy more Palantir stock. If you are a short seller, he especially wants you to put down the coc@a!ne and reverse your position. Karp is a colorful character. Since Palantir has become a stock market darling, he has become increasingly outspoken. One of the topics he has regularly discussed in public is his belief that Silicon Valley companies should focus more on solving problems related to national security instead of launching another photo-sharing app.

(not AI, he did this during a live interview)

This article will give a short primer on The Trial and what Kafkaesque means. Discuss what Alex Karp wrote about in The Technological Republic, and contemplate whether his views and actions would be consistent with a Kafka protagonist or other great characters of literature. If you haven’t already clicked off, buckle up, buckaroo!

But first, make sure to hit the Subscribe button below to join the nearly 1,000 Subscribers that get articles like these, delivered directly to their inbox.

Subscribe now

If this week’s article does not interest you, please check out some other recent ones:

The Trial & A short intro to Kafka

Franz Kafka’s The Trial is one of the most unsettling novels you’ll ever read. Josef K., is arrested and prosecuted without ever being told what crime he committed. He becomes the latest victim of a nightmarish legal system, where guilt is assumed and justice is just an abstract concept without a functional reality. While it’s a darkly funny, it lingers with you as the weight of the implications settles in mostly because of how prophetic it turned out to be.

The Trial was published in 1925, shortly before the rise of the National Socialist (Nazi) party in Germany. Soon, Kafka’s absurd bureaucracy would be mirrored in real-world horrors, where Jews, like Kafka would be arrested without accusation, due process, or access to a legitimate court. The legal machinery that devoured Josef K. in fiction became horrifyingly real, ultimately culminating in the Holocaust, where millions of Jews and other targeted groups were systematically exterminated. Kafka, who died in 1924 from tuberculosis, would not live to see these events unfold or he likely would have perished in a concentration camp, like most of his family.

The Trial is not about hateful political ideologies but rather the crushing weight of excessive bureaucracy. A system that no longer serves people but instead perpetuates itself for its own sake. I wrote about this earlier this year in How Society Uses Bureaucracy to Keep You Safe but Miserable, before I started The Trial. I took a different approach but the bottom line’s the same: if bureaucracy is not actively kept in check, it will become overgrown and eventually erode everyone’s freedom.

Beyond bureaucracy, Kafka’s novels discuss themes such as dehumanization, authority and existential isolation. His work is so iconic, that people refer to things as Kafkaesque when a situation is surreal, oppressive, and/or absurd. The term is often used to describe when individuals are caught in nightmarish bureaucratic loops, against nameless/faceless illogical agents acting on behalf of an uncaring system. In Kafka’s novels, the protagonists are generally the victims of these bureaucracy, yet still choose to fight or reason against their oppressors. These characters often exhibit alienation, guilt, and deep self-doubt, making them appear depressed or overly self-critical, as though they are fundamentally flawed.

If you contrast this with another great fiction writer of the 20th century, Fyodor Dostoevsky, his characters often go through major moral and psychological turmoil, and encounter existential questions. They differ from Kafka protagonists in that they often perceive themselves as more enlightened, morally superior, or "better" than the people around them. Therefore depending on how people perceive themselves, we can decide if they are a Dostoevsky or Kafka character.

We can all find ourselves in Kafkaesque scenarios; anyone that has ever gone to renew their drivers license or apply for a permit, will attest to this. Yet when faced with this scenario, do you accept your fate or try to be like Josef K? Alex Karp sees the West approaching this.

Alex Karp & Palantir

Before co-founding Palantir 20 years ago, Alex Karp did not fit the bill of the typical Silicon Valley Founder at the time. Instead of coding late into the night on side projects, he was busy earning his law degree from Stanford and a PhD from Goethe University. Karp started his career at the Sigmund Freud Institute, an institute focused on psychoanalysis, as per its namesake. He ran a hedge fund for a couple of years before joining Palantir with his former Stanford classmate, Peter Thiel in 2004.

Palantir Technologies is a software company that specializes in data analytics. They work with many large companies but they are primarily known for partnering with governments and military organizations. They first rose to prominence when reports suggested their work aided the operation that led to the death of Osama bin Laden. Shortly after, documents leaked by Edward Snowden indicated that Palantir played a role in supporting the National Security Agency's (NSA) surveillance programs, which were designed to monitor and analyze global internet traffic. Despite questions about these programs, nothing hindered Palantir’s growth.

In 2024, Palantir generated ~$3B of revenue, while growing at a rate of around 30%; they currently have a market cap of around $200B, which is in line with Salesforce and Adobe that generate around 10 times more revenue. The market is currently rewarding Palantir with a very high multiple, which means investors expect sustained future growth.

This can be partially explained by the many ongoing military conflicts providing a boon for military and defense stocks. The United States alone spends nearly a Trillion dollars per year on military and defense, meaning there is plenty of money up for grabs for those that can secure those sweet government contracts. Possibly inspired by the success of Palantir, there has been an emergence of newer startups entering the space. VCs are suddenly very interested in funding defense-tech companies, investing $3B in 20241. This contrasts sharply with when Palantir started; Karp said no VCs were interested in the sector, forcing them to raise funds from the CIA and Thiel’s Founders Fund. This is why, two decades later Karp can confidently take a victory lap, since those investments paid off handsomely and Palantir is well positioned as a leader in a space that is suddenly the hot thing.

The Technological Republic: Hard Power, Soft Belief, and the Future of the West

Alex Karp through The Technological Republic posits the West's technological leadership is endangered by a culture of complacency and misaligned priorities. He believes Silicon Valley has deviated from its historical collaboration with the government on significant technological advancements, such as the internet. Instead founders have been focusing on photo sharing apps and other trivial products. This is consistent with Thiel’s famous quote:

Karp sees this shift as part of a broader societal trend where intellectual fragility and a lack of ambition have infected academia, politics, and business, making this generation weak and less equipped to address, real pressing challenges. He’s certainly not the only person to lodge similar complaints about Silicon Valley innovations failing to address more important issues. For Karp, ensuring Western Democracies, remain the most technologically advanced is the most critical issue of the day. This is needed to maintain hard power, and soft belief.

If the US does not win the emerging artificial intelligence race, it will cede its superiority to rivals such as China, which has a different set of values. This might sound obvious, but as Karp sees it, many people in Silicon Valley and major institutions, are unwilling to say the quiet part out loud, America/the West needs to win. They are afraid of alienating anyone, so they choose not to say or stand for anything.

When Karp sees employees at big tech firms such as Google and Microsoft unwilling to work with the US military, but having no qualms working with the Communist Party of China (CCP), he sees this as an example of people wishing to benefit from the comforts of Western civilization without doing anything to protect it. Karp does not believe you can fence sit. You either support a cause, or you support its opposition. If you don’t support America and its allies, then you are for their enemies. Most people try to pretend these dilemmas do not exist, but in a competitive situation, opting out, is tacit support for somebody. This is effectively the trolley problem.

Many in positions of power feel unable to express their perspectives on issues of consequence, without potentially damaging their institution. However, not offering an opinion, is sometimes worse than taking a position. Karp highlights how this led to the ousting of former Ivy League Presidents Claudine Gay and Liz Magill; when they failed to denounce on-campus antisemitism. Karp doubts they are antisemitic or supported what they saw, but chose not to do anything because they felt unable to exercise their personal judgement, without in some way alienating a stakeholder of their university. In essence, they found themselves on the trolley and instead of choosing a track, they stepped out of the control room, hoping they would not be held responsible for the outcome. This approach failed them.

Karp on the other hand is not afraid. He has risked alienating customers and employees by taking public stances supporting Israel after the attacks of October 7th. He and Palantir were among the first Western companies to visit Ukraine after the invasion by Russia. Karp sees the situation as absurd. The West is in a great competition, yet many of its brightest technologists don’t want to help them win, which aids the competition. As CEO, Karp does not believe his job is to be a soulless agent serving only the system and shareholders. Therefore this begs the question, is Karp a Kafka or Dostoevsky character?

Conclusion: Who is Alex Karp?

Undoubtedly, Karp in his public appearances is giving off Dostoevsky main character energy. He does not follow the generic strait laced, publicly traded tech company CEO playbook. He’s brash, loud and unapologetically outspoken when sharing his views. On the spectrum he’s much closer to believing he’s the smartest person alive instead of being fundamentally flawed. While right now Palantir’s share price is on an incredible run, and he’s flying high, markets can be unpredictable and this can change almost overnight, just look at Zoom shares over the past 5 years.

Yet, you cannot overlook some absurd contradictions in Alex Karp’s rhetoric. He criticizes Silicon Valley for prioritizing trivial consumer apps over serious technological innovation, yet his co-founder, Peter Thiel, helped bankroll the very social media revolution that led to that shift. He argues that American tech companies have a moral duty to strengthen the West’s technological superiority, yet Palantir’s involvement in NSA surveillance raises questions about whether all "Western values" are being upheld equally. If Palantir and other American big tech companies win, how certain are we, we can still enjoy all of these freedoms2?

This paradox makes it difficult to place him cleanly in the Dostoevsky or Kafka archetype. If the West maintains its technological dominance, Karp will be seen as a visionary that saw what others didn’t and prevented the decay. If the West loses the AI arms race and America cedes control to adversaries, then Karp was a Kafka protagonist, that saw this happening but lost because too many in our society were afraid to standup for something. Perhaps it doesn’t matter whether we win this technological war or not, because one way or another, we will all be Josef K.

I hope you enjoyed today’s article. Please Subscribe or Share Some Feedback in the Comments

Subscribe now

Despite the hype, only a small percentage of people actually understand what an LLM is and how these models really work. Though not an expert, I decided the best way to share what I know on the topic is with the beloved 1998 film The Big Lebowski. For those who have not seen it, the film’s tension is confusion around mistaken identity. Prior knowledge of the Big Lebowski is unnecessary to read this article but, since it is an all time classic, I suggest you read the rest of this article, figure out how concerned you should be about LLMs, and then go watch it.

Subscribe now

Data and Compression

Let’s start with some basics. The models that power Artificial Intelligence are algorithms: mathematical representations designed to detect patterns learned by training on data sets. These data sets can contain all kinds of information in different formats. It can come already tagged/labeled by humans, so that the model can see the correct input and output example to be able to answer new but similar questions (i.e. Supervised Learning). The model can also be given lots of raw data and an objective to follow without explicit instructions how to accomplish the task (i.e. Unsupervised Learning). Hybrid models use both labeled and unlabeled data. In any case, primary factors impacting a model's effectiveness is the quantity and quality of the data on which the model is trained. Even with fine tuned parameters (the weights the model learns from the training data), without enough high quality data, any model will be useless. 

ChatGPT and other language-prompt interfaces are powered by LLMs. An LLM differs from traditional models because of their size (some contain upwards of 70 billion parameters) and their UI/UX design (which enables output based on conversational text prompts). LLMs are first trained on massive sets of internet data that are compressed into these many parameters. This forms the basis of the neural network1 that will try to predict the next word in a sequence. While this part is largely unsupervised, the models would not be very useful at this point. To help improve the performance, ChatGPT, Gemini, and other leading companies will have humans manually train and improve the models. This process is called Fine Tuning, which involves swapping out the compressed internet data in exchange for the human supervised responses.

Over time, if successful, the LLM should be able to provide accurate responses to increasingly difficult questions. This is the basic process to get LLMs in a useful state. Sounds simple right? Not so fast, there are a few more concepts you need to understand before we get into The Big Lebowski. The first of these are Word Vectors. 

Word Vectors and Association

A Word Vector represents words using a series of numbers. They appear as a set of coordinates (e.g. 0.0041, 0.1441, -0.9782, etc.), the closer the association between two words, the closer the numbers will be. This differs from human language, which does not reveal anything about the relationship with other words based on letters alone.

In English, two words that are closely associated are pizza and pasta. Though they start and finish with the same letters, you would not be able to determine any link between them. Otherwise logic would dictate a similar association from words such as panda, panorama, or propaganda2. In LLMs, Word Vectors numbers are not static; they change as the model determines a closer association between two words (making the numbers closer together). With this, LLMs can more easily predict the next word in a sentence based on the most commonly associated Word Vectors. As such, these vectors provide important information about the relationship between different words.

(Source: Hariom Gautam Medium Article Word Embedding: Basics)

A complicating factor here is that words often have multiple meanings and interpretations depending on their context and use. 

A word's denotation is the dictionary definition of the word, whereas its connotation is what the word implies. 

A word on its own will normally be interpreted somewhere close to its denotation. Whereas the situation, or words used before or after, can create an entirely different meaning because of the connotation. As an example, if somebody called out “Dude!” one would interpret that this person is calling out for an unspecified man. If they said “The Bowling Lane is ready for The Dude and his friends!”, you would realize that Dude is a specific person’s nickname. You need to be able to understand both the denotation and the connotation of words which is why context is vital for humans and LLMs when trying to understand and execute requests. 

While Word Vectors depict a relationship between words; on their own they have limited applications because they lack context. The lack of context is what drives most of the trouble The Dude encounters in the Big Lebowski. LLMs solve the missing context problem by using something called a “Transformer.”

Transformers

A transformer is a layer that adds information by extracting context present in the data set to help clarify the meaning of words used in a prompt. 

LLMs are built with many layers, the first few start by identifying the syntax of the sentence to iron out words with many meanings. The outer layers increasingly build upon the understanding of the previous layers to be able to understand precisely what the sentence is trying to say. The most recent LLM models now have tens of thousands of these transformers, which is why they can understand your prompts, even if you have typos, imperfect grammar, or incorrect sentence structure. 

(Source: Timothy B Lee and Sean Trott Substack Article LLMs explained)

Now to The Big Lebowski.

The film’s protagonist is named Jeffrey Lebowski. He does not, however, go by this name, instead he prefers to be called “The Dude.” This is something not known to the many characters he encounters throughout the film, as they seek to blackmail a different Jeffrey Lebowski (Big Lebowski), who they know as the wealthy husband of Bunny Lebowski. Without giving too much away, The Dude lives a very quiet and laid-back lifestyle that largely involves bowling with his friends and drinking White Russians3. He is not wealthy, he is unemployed and behind on his rent. 

One day, he comes home to find two thugs waiting in his apartment, to collect money on behalf of their boss, noted pornographer/loan shark Jackie Treehorn. They say they are there to collect money from Lebowski and, as his name is Lebowski and alleged his “wife” said he would be good for it, they want the money. 

The Dude immediately points out that his apartment does not look like a place a woman would live and he is not wearing a wedding ring. The Thugs realize their mistake and leave angrily, The Dude sits bewildered, processing the fact that he watched a grown man pee on his carpet, while another stuffed his head down a toilet.

What clearly happened here was that Jackie Treehorn gave his thugs the instruction to “Collect the money Bunny owes me from her husband, Jeffrey Lebowski.” 

If an LLM was processing this, the first few transformers would help identify that Jeffrey Lebowski was referring to a specific human person. 

It would have looked something like this: Word Prompts (What is Jeffrey Lebowski?) —> Translate Words into Numerical Word Vectors (0.3541, -0.9141, .2141 etc.) —> Run sequence through a transformer to provide additional context (Jeffrey Lebowski is a proper noun). 

Now that they established that Lebowski is a person and the rest of the prompt mandated they go retrieve money from said individual, all seemed simple. However, as the scene unfolded, this was clearly insufficient to accomplish Treehorn’s task. 

With the help of transformers, their probability of success would have been higher if they realized4:

1. Lebowski being married to Bunny 

2. Lebowski is very wealthy 

3. There are more than one Jeffrey Lebowskis in the Greater Los Angeles area 

4. The Big Lebowski is confined to a wheelchair

5. The Jeffrey Lebowski whose apartment they broke into goes by “The Dude”

With this information, they would not have confused The Dude and the Big Lebowski. 

Now that the first few layers of transformers have clarified that the person they need to ask for ransom is the wealthy Big Lebowski, there is the practical matter of how. Transformers need to take a given prompt then execute a response, which happens in two steps.

1. First they evaluate the words given and look around for other words within the network that have relevant context to help it make sense of them. This is the attention step.

2. Next, each word evaluates the information gathered in the previous steps in order to try to predict the next word. This is the feed-forward step.

With the right data set and a properly fine-tuned model, simply prompting Lebowski instead of The Dude would have given the model the clue that the prompter was looking for the Big Lebowski. This would have been figured out in the attention step, since the information that The Dude does not go by his legal name and is not wealthy would be available within the network. Even if the thugs could not access the information within the network, after receiving the instructions they could have realized that Treehorn meant the Big Lebowski (feed forward step). 

Note that Treehorn said he wanted the money that Bunny owed him and to get it from her husband. Nothing about the way The Dude lived demonstrates that he can afford ransom money. The thugs should have immediately realized somebody wealthy would live in a nicer house or area or at least with Bunny. By paying attention to all the information they were given and discovered up until The Dude arrived, they should have realized that they had the wrong guy. 

The Dude did not have enough information at the start of the film to understand what was happening. His neural network did not know Jacky Treehorn, the other Lebowskis, Karl Hungus and the Nihilists, or their relationship with one another. He was able to acquire more information and learn after each interaction with these different characters. When Karl Hungus and the Nihilists claim to kidnap Bunny and threaten to kill her unless her husband pays her ransom, The Dude believes them. The Nihilists knew she just went out of town without telling the Big Lebowski, so they took advantage of the situation. Like the Dude, they falsely assumed that the Big Lebowski had a lot of money, and would pay the ransom. When The Dude learns from Maude Lebowski that her father does not have any money, he realizes that the Big Lebowski was trying to take advantage of the situation to embezzle money from his foundation. Maude fulfilled the role of the transformer by helping The Dude make sense of the information, which eventually helped him realize what the Big Lebowski had done. His process of acquiring information, learning and adjusting based on newly acquired prompts and information, could be seen as being similar to an LLM.

LLMs need so much more data and context to be useful because, compared to humans, AI is still not that intelligent. Popularized by the late Daniel Kahneman, human thinking can be divided into two systems: 

• System 1, which is quick, instinctive thinking that can be quickly retrieved within a few seconds 

• System 2, is slower, more complex thinking that takes much more reflection. 

System 2 can allow inferences based on information or context that might not be immediately evident. Humans have the luxury of being able to use both systems whereas LLMs can only use System 1. To compensate, they need incredible amounts of parameters, transformers and other tools to provide useful results. Despite having incomplete information, Treehorn’s thugs did not take advantage of their System 2 thinking, which led to them to urinate on an innocent man’s rug.    

(Image Source)

To end, LLMs work by training on large data sets, and receive fine tuning with the help of manual human intervention. By using word vectors and transformers, LLMs help compensate for their lack of human intelligence via prompt identification and association. These models have gotten better outputs because of the increasing number of parameters and transformers in each new version, giving them more information and context. 

The Big Lebowski is the perfect film to show the difficulty machines (i.e thugs) have accomplishing simple requests when missing relevant information and context.

Well, that’s just like, my opinion man. What’s yours?

Subscribe now