Use Case

Contract Programming with AI

December 1, 2025

The AI revolution has allowed us to write enormous amounts of code extremely quickly. Writing boilerplate setup and boring repetitive glue-code is now a thing of the past. The downside is now that we can write so much, we find ourselves having to read it.

Gavin

More Stories

Data Workflows

When does GenAI work for business?

November 26, 2025

The approaches that have been most often successful take existing workflows and try to redesign them to incorporate generative AI to improve throughput. The key is that business processes have to be design to appropriately accommodate AI as AI is not capable to use as a drop-in replacement for people.

Gavin

Data Workflows

Visual Tokens and AI Communication

November 20, 2025

DeepSeek-OCR doesn’t translate pdfs into text. It translates PDFs into *tokens* which give a tokenisation of the visual representation. And what is even weirder, these visual tokens are more efficient representations than textual tokens. In fact, it requires about *half* as many tokens for a visual representation.

Gavin

Data Workflows

Agoric Computation

November 18, 2025

Our prediction is that the business model of products will be increasingly pressed towards API layers which take part in a well defined and specified market of agoric computation. Endpoints will tell you their cost, agents can decide whether or not to use them, and then can calculate a cost plus for their own activities based on current input commodity costs.

Gavin

LLM tips

Teaching AI to count

November 17, 2025

AIs are notoriously poor at math, and this isn’t restricted to complex arithmetic. They also tend to be very poor at counting. You can see this by asking one of the AIs directly to produce JSON with a certain number of elements. They will just often get the number of elements completely wrong.

Gavin

Models

Why do AIs lie?

November 14, 2025

Hallucinations are common when the model simply doesn’t know what it is talking about. More technically it is when the modal is epistemically uncertain (high entropy or high disagreement of plausible sequence continuations).

Gavin

AI Adoption

AI Adoption is Surging. 80% of Orgs See No Bottom-Line Value. Why is that?

November 6, 2025

Mckinsey's state of AI report shows that AI adoption is surging, but the reality is that 80% of companies see no bottom-line value and that those who are winning at the big players. This article looks at how startups and smaller enterprises can use AI to punch above their weight.

Oliver

Use Case

An idea for a self-service AI assistant builder

November 6, 2025

I'm a novice coder myself, but I have an idea for a self-service AI-assistant application built with BookWyrm. This article details the top-level overview of the app and how you could scale it for automation and workflows.

Oliver

Data Workflows

How to be a robot pychologist: LLMs require guidance

October 21, 2025

It's probably best to think of an AI as an idiot savant. It can type at a bazillion keys a second and is good at keeping several pages of code in its head at once. It has an encyclopaedic knowledge of libraries. It is *not*, however, very good at thinking 'out of the box' and it has an annoying tendency to make the sorts of coding errors that are pervasive from the vast resources of bad code that already exist in the wild.

Gavin

Data Workflows

AI Citations with big context models

October 8, 2025

Recently some of the big AI companies such as OpenAI and Grok have been releasing models with really enormous context windows. OpenAI's new 4.1 ('gpt-4.1-2025-04-14') has a 1 million token context window.

Gavin

Data Workflows

AI Driven Citation: Controlling Hallucinations With Concrete Sources

October 6, 2025

This process, which I call “citation under question” because it does a deep read through the entire document with a question in mind, taking notes about sections that answer the question, is precisely what the bookwyrm.ai citation endpoint is designed to solve.

Gavin