Alex's notes

A heuristic that’s helpe me decide if a given problem should be given completely to AI or use a human-in-the-loop model is “transaction symetry”:

I was watching an interview with a exotic car salesman when he said something that struck me.

AI agent security is application security, network security PLUS controls for prompt injection (the lethal trifecta) and LLM fallibility (trusted/untrusted input).

AI agents are able to do good things and bad things.

OpenAI released a macOS Codex desktop app (I’m pretty sure Codex was CLI only until now).

Install claude-code-router Create a config file in /.claude-code-router/config.json Run ccr start Run Claude Code ccr code Config for LM Studio { "Providers": [ { "name": "openai", "api_base_url": "http://<lm-studio-host-here>/v1/chat/completions", "api_k…

Runbook for creating a new dokku app on aws.

My current list of best large language models for local AI:

Lately I’ve been de-node-ifying my personal projects.

At time of writing, AI needs humans to do anything useful and there is a big difference between the best and worst employees at using AI.

RustPython is a python interpreter written in rust.

An increasingly important part of the AI stack is running untrusted code in a sandbox.

OpenAI announced Apps SDK which allows developers to integrate apps into ChatGPT.

When giving LLMs the capability to access private data, view untrusted content, and externally communicate, bad actors can trick AI agents into leaking private data via prompt injection.

Annoyingly, Fathom does not provide a way of downloading all meeting videos from an account.

I build personal infrastructure around the things I do constantly, refined for my workflow (quirks included), with built-in privacy, and for fun.

Only show error message details when the user can recover from it.

I’ve been spending more time with local large language models.

It’s been awhile since I’ve published anything.

Some initial impressions on the AI-first web browser from Perplexity, Comet:

Just like AI coding agents are forking VSCode so that agents have more control over the development environment, AI companies are forking Chromium so that agents have more control over the browser environment.

Meetings are not useful unless they lead to an outcome or clearly defined tasks needed to reach an outcome.

Today’s applications provide closed systems that can not be modified or extended to suit the needs of the user.

Here’s how I’ve seen DB connection/checks in a load balancer healthcheck become bad idea:

Maybe you want to be nice.

A surprisingly useful technique for getting the most out of AI tools is to prompt it to ask you questions that would help it to complete the task you want.

It’s still early days for building user applications with AI but we are starting to see things progress as the industry progresses from LLM workflow to AI agents.

Based on a data from Gainsight the average ratio of customer success managers to customers is 144 for low-touch businesses with less than $10k in annual contract value (ACV).

I kept getting an opaque rust compile errors when writing an axum API handler.

With the capabilities of large language models getting more useful for real work, the pressure is on to incorporate them everywhere.

Typed languages should be the best fit for useful AI agents.

Apple keeps making it more and more difficult to “install” a PWA.

I’ve been testing out goose an AI agent for writing code that runs on your machine instead of as an IDE co-pilot.

When I started doing founder-led sales, I thought it was my job to find a way to sell to anyone who got in touch and scheduled a meeting.

Ironically, one of the most efficient strategies for building with AI is to wait for better models.

The Plotter notebook from Designphil is a minimal ring binder for planning and writing.

An incomplete list of AI-powered web browser automation and AI agent projects.

I started trying out n8n for setting up automated workflows.

This is my personal list of ink reviews based on my usage and preferences.

When a process requires human judgement for an unknown number of possible decisions, automation is not possible.

Here’s a collection of things that have helped me be more productive.

Higher temperatures tell LLMs when generating a completion to not always use the highest probability next token.

Workflows are a sequence of actions with LLMs to process input into a desired output.

Customer satisfaction (CSAT) surveys measure how much people like a product or service.

Notes from reading Prompt Engineering for LLMs by John Berryman and Albert Ziegler.

A lab notebook is where research scientists keep track of their experiments so that they can be reproduced and verified.

Rather than converting to text at every step in a chain of thought process with large language models to solve a complex problem, new research suggests that reasoning can happen in a latent space using the internal representation of the model.

As large language models proliferate into every service and ultimately replaces business logic, we will be left with the horrible burden of maintaining mush.

Compiling rust dependencies every time a docker image is built can take a very long time.

Satya from Microsoft talks about how orchestrating between business applications is the next step for artificial intelligence which will replace business logic with AI.