Kah Hong T

Kah Hong T

Quantitative Research and Machine Learning

Convexity

1 minute read

Published:

Understand the intrinsic convexity of opportunities. Convex opportunities are those that produce increasing return per unit of marginal effort over time. Here, return and value are intentionally ambiguously defined, and they could be any notion ranging from value to the customer, financial return, number of goods sold, new knowledge, etc. Identifying convex opportunities can be challenging; it can be due to a mix of both exogenous and endogenous variables including various technical, regulatory and market dynamics, state of technology, location, and point in time.

The most successful technological companies today are all engaged in convex dynamics: Meta, Amazon, and Google all have strong network effects that scale exponentially with each additional platform user. In other cases, like in the case of Stripe, with fixed cost invested in research and development their technological stack can accommodate exponentially more users while the cost base mainly consisting of engineering talent costs scales linearly. This means % margins grow exponentially. OpenAI benefits from a similar dynamic even more so when their cloud and compute resources are further scaled with their Microsoft partnership; but their training regimes are also bolstered over time by the creation of more content available for training on the internet. In such businesses, there is either a greater increment of value per unit user, a decreasing cost per unit user onboarded onto the platform, or both.

More thoughts on this to come.

You May Also Enjoy

Focus on Building

less than 1 minute read

Published:

Took a breather from the myriad changes happening the last few weeks. This year has been eventful indeed. Some principles to follow as I pursue a more definitely uncertain, indefinitely successful and infinitely interesting future.

OpenAI Reasoning Model o1

less than 1 minute read

Published:

OpenAI just released the o1 model that is focused on reasoning capabilities, generally substituting immediate next token generation with some reasoning steps introduced before some delayed token generation. This represents a significant 0-1 step-up in LLM capabilities, immediately surpassing beyond Gemini, Claude and other open-sourced models like Llama3 in mathematical and scientific reasoning. Most likely, this model does not represent a foundational model improvement, but rather entails introducing an orchestration layer informed by chain-of-thought reasoning, self-reflection and reinforcement learning. As opposed to zero-shot prompting using the base pretrained model, this means perhaps a factor more number of inference calls for each request. This has interesting implications on the output token cost structures of using AI models with reasoning.

AI Financial Analyst?

3 minute read

Published:

I have recently read several articles echoing a similar sentiment: due to the limitations of artificial intelligence, AI will not replace traditional financial analysts anytime soon. One author suggests that human intelligence is still decades ahead of AI in automated fundamental investing. However, despite the confidence seemingly portrayed in these articles, the rationales and empirical evidence provided to support this notion have been weak.

Competition as Validation

less than 1 minute read

Published:

Peter Thiel: Competition as Validation. It may be tempting to squeeze through the narrow door that everyone is trying to get into, but take a step back and instead look around the corner and explore the vast, empty gate that no one is paying attention to. It is unthinkable that in many perceivably esteemed career paths the degree of competition has become so fierce when the stakes are so low. The intention is not to insinuate a lifestyle lacking in competition, and thus lacking in ambition. While thriving to succeed in a highly competitive space may indicate some level of ambition, the reverse is not true.