James Hall

What happens when AI models take Open Source code and documentation without attribution? Is this approach destroying Open Source as we know it? Me and @mrjameshall.bsky.social dig into these questions in last weeks episode of Off Script:

heypresents.com/podcast/57

What's better than a Friday afternoon? A Friday afternoon when you're chatting to this wonderful lot about running a conference.

cc @philhawksworth.dev @mrjameshall.bsky.social

It’s always a good sign when the venue tells you you’re playing Daft Punk too loud and it’s affecting the screening of Minecraft next door.

*turns it up louder*

A screen at the front of a stage at a conference showing live captioning. It reads, “[Hope you had a nice lunch… captioned took a nap]”

Live-captioning a tech conference from another timezone must be knackering!

Happy to see this from our wonderful stenographer at #AllDayHey.

Carl Sagan explains how the ancient Greeks knew the Earth was round and calculated its circumference over 2,000 years ago.

How to plan and run a discovery workshop What is a discovery workshop and how can they be used to work through problems in your business? In this episode, James and Josh discuss how they've run them in the past and offer some tips and tricks...

I learnt so much from this chat with @mrjameshall.bsky.social on running discovery workshops, if you're looking to host one it's definitely worth a listen.

Spammers are better at SPF, DKIM, and DMARC than everyone else. Which makes them entirely useless signals to use to block mail.

Great work everyone 😂

toad.social/@grumpybozo/...

Getting your head around sampling (and constrained sampling - like used for Structured Outputs) is a good area to get stuck into.

- Built a basic UI on top of this to start exploring.

Some interesting advantages to having greater insight into underlying probabilities is knowing when it might be wrong, debugging models issues, and for learning purposes.

What I've done:
- Pulled Llama3.3 locally from HuggingFace, converted it to GGUF for llama.cpp.
- Quantized it so it doesn't crash my M3 Mac.
- Started a llama.cpp server, and use the raw (non-OpenAI-compatible) API to delve into the underlying logprob data.

Another weekend project, I've built another visualisation.

Commercial LLMs often don't expose their logits/logprobs. This is the chance that the next token/word appears.

Add your own words or compare predefined sets.

I'm working on a way to visualise the famous king - man + woman = queen example. Any other ideas?

I've made a tool that lets you explore Text Embedding Vectors (an important step for AI tools which use Retrieval Augmented Generation) in a 3D space.

You can choose between different dimension reductions, which helps you compare features.