Anthropic is onto something interesting with Artifacts

A few weeks ago, Anthropic released a new feature to their Claude chatbot — Artifacts. It was announced together with Claude 3.5 Sonnet. You can find more details in the section about the feature. It's …

Read more

In-browser LLM on Chrome Canary

The nightly version of Google Chrome ships with an on-device LLM. The model it runs is Gemini Nano. Some documentation about this can be found here.

It's not surprising to be honest, that tiny LLMs …

Read more

The Pragmatic Programmer Review

The Pragmatic Programmer 20th Anniversary Edition book cover

The Pragmatic Programmer 20th Anniversary Edition, by David Thomas and Andrew Hunt, offers high-level, actionable advice for software engineers. Unlike many technical books that focus on specific coding practices or languages, it focuses on cultivating …

Read more

Thoughts about Anthropic / OpenAI work on interpretability in LLMs

A few weeks ago, Anthropic came out with this paper, showing how they found interpretable features in one of their models. This means that they could see which features activate when the LLM generates, and …

Read more

Mozilla is experimenting with local alt text generation in Firefox

As the title suggests, Mozilla is experimenting with using on-device Generative AI to generate alt text. Here are some meandering thoughts.

This raises the question whether it makes sense to invest in AI generated alt-text …

Read more

Docker, compose, secrets, and environment variables

A dockerised app I'm working on needs some secrets at build time. In my case, it needs to pip install from a private repository. Note that this is a generic problem, luckily with quite an …

Read more

LLM synthetic data in bash

The llm command line tool by Simon Willison lets you interact with LLMs from the command line. I've been using it recently, and while it's nice and convenient, I wasn't doing anythin I couldn't do …

Read more

Andrew Ng's advice for iteratively writing LLM prompts

When building complex workflows, I see developers getting good results with this process:

  • Write quick, simple prompts and see how it does.
  • Based on where the output falls short, flesh out the prompt iteratively. This …
Read more

OpenAI Model Spec

OpenAI has recently published a document titled 'Model Spec'. It defines an ethics and norms framework models should follow. A few things that stood out:

  • It surprisingly resembles Isaac Asimov's "Three Laws of Robotics", with …
Read more

Deployment tech for the hobbyist, again

TL;DR: Try Dokku if you want to run multiple toy web apps on a single server and keep the cost at minimum.

Dokku logo with name

Let's talk, yet again, about deploying toy web apps. I wrote about …

Read more