We keep trying to get LLMs to do math. We want them to count the number of “rs” in strawberry, to perform algebraic reasoning, do multiplication, and to solve math theorems. A recent experiment particularly piqued my interest. Researchers used OpenAI’s new 4o model to solve multiplication problems by using the prompt: Calculate the product of x...
In the 1800s, before serfdom was abolished in the Russian empire, landowners paid taxes based on how many serfs they had. A census was conducted every few years by government employees traveling across the empire and doing counts; a manual map-reduce of epic proportions. If a person was dead, it would often be years before the government cleared...
This is a near-transcript of the talk I gave at PyCon Italia 2024 in May in Florence. Introduction Buongiorno PyconIt, grazie per avermi invitata a parlare! Avrei voluta fare tutto il discorso in italiano, ma lo sto ancora imparando. Per adesso posso parlare soltanto di gelato o colori. Perché non so ancora dire, “don’t worry about LLMs”, il...
Jakob’s Law of UX goes something like this. I, as a user online, spend my time on many sites. As such, when I come to your site, I am already used to the way the other sites work, and I don’t want to learn new paradigms. Some also call these preconceived notions user mental models or affordances. I like to call it the user-site contract. For...
We are now in a very weird liminal space in information retrieval for consumers, particularly those attuned to trends in search and working on the bleeding edge of LLMs. On the one hand, we have the fall of old companies. Broadcast-based centralized social media, which steadily served as a newsfeed and realtime search for a small, vocal minority,...
I, like many developers who have worked on high-scale, low-latency web services over the last fifteen years, have an intimate relationship with Redis. At any new job, when you ask where the data is, and someone points you to a server address with port 6379, you know you will meet an good, reliable friend there. When you shell into the redis box...
edit: Huge thanks to Vamsi for digging into how to translate this into English! I recently came across a really great Soviet video from 1971 called “Myself and Others” (unfortunately only in Russian so far) where the creators examine how people react psychologically to different situations and how we see ourselves in light of a group. In the...
Table of Contents How We Use LLM Artifacts What is a machine learning model Starting with a simple model Writing the model code Instantiating the model object Serializing our objects What is a file How does PyTorch write objects to files? How Pickle works From pickle to safetensors How safetensors works Checkpoint files GGML Finally, GGUF...
Image with some help from Dingboard. In 2023, I wrote two pieces on machine learning engineering for The Pragmatic Programmer. (Part 1 and Part 2). However, since I started working with LLMs recently, neural architectures have changed some of those assumptions. To be clear, most of machine learning in production is still not related to large...
Viberary is a side project that I worked on in 2023, which does semantic search for books by vibe. It was hosted at [viberary.pizza.] I’m shutting down the running app and putting the codebase in maintenance mode because: A lot of what I want to continue to do there (i.e. changing embedding models, modifying training data) involves building out...
Favorite books of 2023 This year, I managed to read more than last year, but I was still pretty caught up in technical learning and unfortunately didn’t reach the fiction-non fiction balance I wanted (I always try to read more fiction than non-fiction.) Demon Copperhead by Barbara Kingsolver - By far my favorite book of the year. The premise is,...
I saw this tweet over the weekend and wanted to dive into the fundamental question behind this: Given this potential error, why do we use conditional imports at all, or, more specifically, when might we use this pattern? The TL;DR is that we use this pattern to hedge between the differences in typechecking enforced by mypy and typechecking as it...
This is the keynote I prepared for PyData Amsterdam 2023. The TL;DR is that we must understand the historical context of our engineering decisions if we are to be successful in this brave new LLM world. The text here isn’t exactly what I said, it was my notes ahead of time. My slide template is by Barbara Asboth, who also did the templates for...
Every day I open my LinkedIn and Twitter (and Mastodon and Bluesky and Threads….) and am innundated with the same messages: LLMs are sent to us from above, they make everyone’s life easier, we are quantizing and pruning, going faster, getting smaller, they will change education, they will write our poetry, they will outlive us all and overthrow...
“The beginning of wisdom is the ability to call things by their right names. " - Confucius. As a writer, I’ve always been fascinated with names. How people get their names, what they mean, whether they like them or not. When I was twelve, I bought a baby name book and spent hours poring through the various sections trying to decide on names for...
I work in machine learning and read about it a lot, but ChatGPT still feels like it came out of nowhere. So I’ve been trying to understand the hype. I’m interested in what its impact is on the ML systems I’ll be building over the next ten years. And, as a writer and Extremely Online Person, I’m thinking about how it could change how I create and...
This is part of a series of posts on building Viberary, a semantic search/recommendation engine for vibes and what happens when you have unlimited time to chase rabbit holes in side projects. I’m still in the early stages of this project and doing data analysis on the input data, a dump of 10GB from Goodreads in JSON. See the input data sample...
I did a LOT in 2022. Way too much, especially in December when I was still a few months into a new job, running a conference, wrapping up a compsci class, and trying to plan a family trip to Argentina. Note, none of this does includes my other main hobby - actively parenting two small children. As soon as Normconf was over, I got so sick I was in...
Leggendo Wohpe You can buy Wohpe on iBooks, for now the ebook here, on Kobo here, and soon on Kindle. There are very, very few people who are both excellent engineers and excellent communicators; so rare, in fact, that I can count them on one hand: Paul Ford, whose writing about technology was my first hint that technology is something you could...
Over winter break in 2022, my husband, my oldest daughter and I went to Argentina to see the country and visit my friend. We spent four days in Buenos Aires, the capital, and four in Bariloche, a small resort town that nestles the Andes foothills in Patagonia on the border with Chile. Our Airbnb in Palermo, Buenos Aires I wouldn’t describe the...
On December 15, 2022, the first and only Normconf, the tech conference about all the stuff that matters in data and machine learning but doesn’t get the spotlight, happened. **All the talks, the lightning talks, and hallway track talks are here. ** The event was a 15-hour long event split into three sessions hosted by 2 MCs, free and streamed on...
Composition X, Kandinsky Recently, I’ve come to the realization that much of what we do in modern software development is not true software engineering. We spend the majority of our days trying to configure OpenSprocket 2.3.1 to work with NeoGidgetPro5, both of which were developed by two different third-party vendors and available as only as...