We love to optimize Technical people love to optimize every little thing. For example, I am always thinking about how to use my time, money, and knowledge in the most effective way to the point of getting upset when I do things suboptimally. However, everyone optimizes for different things. When I had my first software engineering job, I was learning to use vim and I downloaded a bunch of different plugins to make the app prettier and have more functionality. Then, when I finally had my setup perfect with 10 different plugins downloaded, I saw that my boss was using stock vim! How can you get anything done without neotree?? He viewed vim as simply a means to an end. How much do plugins really help productivity? Was I really optimizing for the right thing by using my time to learn these plugins? ...

Debugging prod after shipping on Friday Your job as a software engineer Every line of production code is a bet on speed, safety, and correctness. When you ship software, you are balancing: Getting it done fast -> shipping faster means the feature appears earlier for customers and ultimately the business makes more money Not introducing bugs -> bugs can cause outages or other errors that lead to losses The problem is that anyone that has worked with software knows that no matter what, you cannot guarantee that your shipped software will not have bugs. In fact, your software will always have bugs. No matter how many “lgtms” you get on your pull request, bugs always find a way of creeping in. ...

How to learn I did competitive gymnastics growing up, and the training was chaos…in a good way. You had to be comfortable flying through the air, twisting, flipping, and hanging upside down. Even one skill, like a giant (spinning in full circles around the bar), had a hundred little steps. You first learn it with your hands strapped to the bar, feeling the motion through your body. Then a coach talks you through it while spotting you, combining physical touch with verbal cues. The whole time, you are watching other gymnasts do it and trying to mimic them, visualizing the move as they do it. You feel how your muscles should move, hear corrections in real time, and see it all come together. It wasn’t just doing: it was watching, feeling, hearing, and imagining. ...

I have recently seen a lot of rhetoric about this study that shows that developers take longer to code with AI compared to without AI. However, speed isn’t the whole story. The real benefit is that AI finishes your thoughts. You start writing or coding and it carries the thread to completion. That means less mental juggling, less half-baked ideas, less cognitive strain. AI is not just an accelerator, it is a way to reduce the cost of thinking. That is what actually makes you more productive.

Recently, when I was at the airport looking to get a ride share back home I checked the price of both Uber and Lyft. I noticed that an Uber was 25$ and a Lyft was 39$…that is a price difference of 14$, or more than half the price of the Uber! The Lyft ride was 56% more expensive… How can two multibillion publicly traded companies have such different pricing? Both companies have entire teams devoted to determining pricing, yet they are still so off. ...

I remember watching a video a while ago where a pro card counter answers questions about gambling. In one of his answers, he said that what makes someone a true gambler is that the joy of winning a bet is more than the sadness from losing a bet. This has stuck with me since I watched the video over a year ago. Since then, I sometimes think about why people gamble and what makes some people more prone to gambling. I play poker with my friends weekly, and in my group, there are many different kinds of gamblers. Some people are there for fun, some care more about making money, and some just like the thrill of gambling. ...

With holiday season in full swing, people may be more enticed to give to charity. However, there is a question of how to optimize your charitable donations. Specifically, I would like to optimize charitable donations for maximum impact. It seems to me that there are two parts to this equation: ...

Why home runs? Some of the best moments in baseball games are home runs. Something about hitting the ball out of the park is satisfying. Since baseball season just started, I wanted to model a part of the game. I decided to model home runs since they are pretty rare events but should still be able to be accurately predicted. When I say accurately predicted, I mean that we can accurately predict the probability of a player hitting a home run. At the end of the day, we can only compare to the true data…but we have many years of baseball data (baseball is nice since there are very detailed statistics and metrics for each player for 100+ years). ...

Introduction The goal of this article is to explore the latency of different OpenAI models. When using AI models in production, latency is an important factor to consider. Comparing Model Architectures First, I test the latency for different OpenAI models. I test the following models: gpt-4, gpt-4-0613, gpt-3.5-turbo, gpt-3.5-turbo-0613, gpt-3.5-turbo-16k, gpt-3.5-turbo-16k-0613, text-davinci-003, text-davinci-002, text-davinci-001, text-curie-001, text-babbage-001, text-ada-001, davinci-002, babbage-002, davinci, curie, babbage, and ada. These are all the OpenAI models that are available for inference through the chat and completions endpoints. The models can be divided into chat models, instruct models, and base models. Chat models are gpt-4 and gpt-3.5 and are LLMs that are optimized for chat. Instruct models are models that are trained with reinforcement learning through human feedback to follow instructions [1]. ...

Disclaimer: This article mentions https://terra-cotta.ai/, an LLM experimentation platform I am building Introduction ChatGPT, Bard, and other large language models (LLMs) are very useful for a wide variety of tasks from writing code to answering complex questions to aiding with education. However, these models are ultimately limited by the data that they are trained on. Also, these models are trained to be able to answer a wide variety of questions which may not be sufficient for domain-specific questions. Fine-tuning is essential in order to make these models accurately answer domain-specific questions and be useful for difficult tasks. Furthermore, fine-tuning may be cheaper for inference. ...