⌨️🤷🏻‍♂️📷

LLMs Are Magic - Their Applications Should Be, Too

Mon, 07 Jul 2025 15:17:28 -0700

Three years ago, it seemed absolutely certain (to me at least) that

LLMs would continue to get even better and better, and
LLMs would soon power just about every digital experience in our lives.

Only one of those turned out to be true. Three years later, LLMs are smarter than ever. They can reason, they can generate images and videos (either natively or with a bridge to another model). They’re fast and getting faster; cheap and getting cheaper; smart, and… you get it.

Trying simple tree-search techniques for LLM token sampling

Sun, 26 Nov 2023 19:56:30 -0800

Trying simple tree-search techniques for LLM token sampling

At its core, an LLM is a value function.

Given a state (i.e. the current context, or input text), it scores all possible next actions (i.e. tokens).

Therefore, it’s pretty simple to imagine the task of sampling tokens as a state-space exploration problem, where we use a tree data structure to map out the state space and explore it to find high-scoring states.

Using Guidance with GPTQ

Thu, 18 May 2023 14:52:01 -0700

I’ve been very enthusiastically playing with guidance, a library from Microsoft that lets you compose prompts for LLMs, and interact with those LLMs programmatically in a very natural way. I plan on using it for my personal projects going forward.

I’m excited because this library lets you build prompts the way you think about prompts. No more fussing around with Langchain.

One issue is, I mostly play with GPTQ-quantized models. My current favorite is Wizard-Vicuna-13B-Uncensored-GPTQ, very kindly quantized by TheBloke.

4.2 Gigabytes, or: How to Draw Anything

Tue, 30 Aug 2022 00:00:00 +0000

In our world, we can do anything that we want to do here. Any old thing.
- Bob Ross, The Joy Of Painting Season 29 Episode 1

Watching a vibrant Seattle sunset the other day, my imagination started running. The otherworldly hue of the sky evoked something from science fiction. The hazy orange-purple was mesmerizing.

I envisioned a massive, alien object hovering over a long-abandoned Seattle, with a burning orange sky, and buildings overgrown as nature reclaimed the city.

Bear plus snowflake equals polar bear

Sun, 13 Jun 2021 18:08:57 -0700

Prelude: e and e

Quick, how many bytes make up the following line? No tricks, I promise.

Hello!

The correct answer is: 6. Or 7, if you want to be pedantic and include the newline, but let’s not.

This is simple enough; this page is encoded as UTF-8, which implies 8-bits per ASCII character, or a byte per character.

Let’s play again. How many bytes make up the following line?

Hello!!

Easy enough. 7. 7 characters, so 7 bytes.

How to create a Hugo-based markdown blog

Sun, 02 Aug 2020 15:48:23 -0700

Step 1: The spark ignites

Oooh, look at this cool person’s blog post that was posted on Hacker News… Why don’t I have cool, minimal blog like this?

Step 2: Pick your theme

Spend two hours browsing minimal Hugo themes until you find one that’s simple yet functional.

Step 3: Simplify

Why does this “simple” theme include google analytics, disqus, and javascript auto day/night theme detection?

I should gut it of all the stuff I don’t need…

About

Mon, 01 Jan 0001 00:00:00 +0000

Hi.

I’m Andy Salerno.

I’m a software developer, currently building Copilot Studio at Microsoft.*

Things that interest me: programming languages with strong guarantees 🦀, open source, street photography, coffee.

Reach out

Email: andysalerno at gmial dotcom (fix the obvious typo in the domain)

Github is linked in the page header.

*Views in this blog are my own and don’t reflect the views of my employer, yadda yadda.