Parijat Garg

I watched Hidden Figures a few months ago. 9 years too late, but even so. For all the bigotry of the time, and the incredible fact that humans actually calculated space flight trajectories, what really stood out to me was something else entirely. And I doubt even the makers of the movie would've thought how relevant that element would become today. It was the coming of the IBM Computer to NASA. And the decision by one woman to adapt rather than to give up. “It’s not going to matter soon. This IBM’s going to put us all out of work… Only one thing to do: learn all we can. Make ourselves valuable. Somewhere down the line a human being’s going to have to hit the buttons… We have to know how to program it. Unless you’d rather be out of a job?” Just like today, it must have felt that computers were coming for "white collar" jobs. Just like today, many people must have been in denial, or outrage. Just like today, humans double checked the output from the system. And just like today, one woman decided she had to learn how to make herself valuable.

Real risk is an immeasurable quantity. Both ex-ante and definitely ex-post.

I'm now having claude write me a seventh book. It's been amazing so far. But here's a thought. A lot of the history stuff in particular, it has been able to write because someone spent ridiculous amounts of time and energy researching the original sources, did interviews maybe, came up with a point of view, and wrote them out. So we have the luxury of being able to read a succinct distillation of it. But for that to be possible in the future, we need a Dalrymple and a Caro and the others who dedicate their lives to their craft. We need a mechanism to continue to reward that effort. One would be tempted to think that is we just fed the original sources to an LLM, they might produce what they produced. But I doubt it. Especially, if you read about Caro's process when he was researching the early years of Johnson's life, his wife and he had to make insane efforts to just be trustable enough to the women in the Texas Hill Country, that they would tell them anything. What LLM could ever do that? I love this technology. And I'm using it to build what I've always wanted - a tracker of "everything" that happens in the world. But I also think constantly about the value of a human being in a world so saturated with AI. And no, I don't think there is an dissonance in having this thought and having Claude write me books on things I want to learn about. In every field, we have to identify the true source of human value. And we have protect it, and nurture it, and incentivize it. Perhaps the truly important thing a human being brings to anything, is a point of view. It's the point of view that matters - whether it is about how you frame a person in a photograph. Or how you frame an argument over the water of a river. When more than one point of view is possible, when there is no objectively correct answer (and most of life is like that) - I think it matters what a human being thinks about it. And I think we will always care about it. I hope we do.

Not a few weeks ago, my US portfolio used to be the source of comfort, while the Indian one was getting beaten up. Not anymore 😂😭 At least my Infosys position is countering some of the mess...

I don't know who made the Jio Hotstar sound logo - but oh god it is horrid!

Reasoning about complex systems, especially human systems like multilateral institutions or geopolitical systems, involves three basic elements. The first is "models": mental models of how the various parts of these systems interact with each other, what drives them, what matters, and what incentives shape their behaviour. The second is "memory": factual knowledge about the various people involved, their histories, and the history of the institutions themselves, their structures and rules, etc. This means holding all of that information in memory so that it can be applied to the models to derive insights. The third is "throughput". There is simply too much happening. At an organization like the IMF, for instance, an enormous amount of information is generated almost every day. Even a human who has the right mental models and possesses extensive factual knowledge cannot process all of that information and reason through it. In some ways, then, the three fundamental elements are model, memory, and throughput. AI systems based on LLMs can largely solve for memory and throughput. This creates an opportunity to build systems in which humans who understand complex systems and have strong mental models can encode those frameworks into LLM-based systems, then leverage their capacity for memory and throughput to apply those models and infer insights about very large systems that generate information at enormous speeds. Of course, the models themselves do not remain constant. They evolve as our understanding improves and circumstances change. Humans can focus on refining and improving these models. In some ways, this is a form of meta-learning: humans develop and refine the frameworks, while LLM-based systems apply them at scale.

Alright. Time to shut down the internet again.

My new favourite learning hack. On any reasonably specific subject, ask Claude to write me a history of how the field evolved, what are the main ideas, and who are the people. Then have it convert it into epub and email it to my Kindle. Read myself to sleep. Today's reading: the structure of Indian business and economy in the 19th century.

Imagine being upset when people started selling cloth made by automated machines instead of a handloom...

This is happening in every creative field. If you write a children's book and illustrate it using AI, publishers basically just turn up their noses and walk away. Like, are the illustrations bad or you just don't like that they were made with AI? I mean, would you refuse your caramel machhiato if a robot made it instead of a barista? And it is not about protecting jobs. Illustration (like game design), needs aesthetic sense. A great illustrator, like a great game designer, will wield these tools better than the amateur. And so they can still win. But the issue is that the old guard don't want to move to the new way of doing things. And they don't want anyone on their turf either. It's going to happen to your field. Graphic design. Game design. Law. Investing. Whatever. You either upgrade your own game, or watch yourself get left behind. Gatekeeping will only work for so long.

This is getting addictive. Our "research analysts" are beginning to form worldviews, learning lessons about their investment process, and collecting/organizing oodles of factual information... from the process of doing the research. The analysts are getting better, and this does not even need expensive "model training" or fine-tuning. I don't think even human analysts learn so systematically. And you definitely cannot literally see what they learned and how their minds changed in the process.

So now that we've got agents that can read through 10 years of a company's annual reports, their last 8 quarterly results, investor presentations and earnings calls... for 1500 companies... in one day. What exactly are we looking for if we want to invest for 5-10 years?

it is surprisingly non-trivial to build the infrastructure to have an agent schedule tasks for itself... but once you get it right, it is surprisingly fun too!

😂

If you had a WhatsApp based assistant that could help you with investing, what would you want it to do? Apart from answering questions about specific stocks etc. Think of things you'd want it to do automatically for you, on a regular basis etc...

While frontier labs are racing to have their models generate "new knowledge", I'm very excited with the prospect of reorganizing existing knowledge in ways that make it dramatically easier for humans to consume and assimilate it, and *then* build on it. In so many fields, the ability to contribute meaningfully is limited by the ability to get to the existing frontier first. LLMs could make that journey meaningfully easier.

As the models get better and better, the differentiator begins to be what you are asking (and getting) them to do. It begins to seem like a lot of people didn't actually know what the end goal of their tasks were. They were happy just following "the process". But now as "the process" can be executed faster and at scale, you are forced to think about what the end goal really is. Until now, "just keep swimming" was the mantra. Now suddenly, you've got to think more clearly of where you are swimming to in the first place. Turns out, a lot of people don't know!

Can't wait to get them talking to each other

What a joy it is to behold as your agents start learning by reading and doing, and building their own compressed repository of notes, and even building an evolving worldview...