This is writing that comes out of the work. We publish when something is finished, so the gaps are long.
An AI assistant looked at the raw numbers inside several small neural networks and, an MIT CSAIL blog post claims, made sense of them. It identified four games from weight data, wrote working game networks from scratch, and found and fixed a fault in a language model’s weights. The blog’s headline figure, a jump from 7.9% to 48.1% on a maths test, is mostly a looping fix. The editing has since been repeated with a different AI, though nothing is peer-reviewed.
What you say to your AI coding agent in what feels like a private exchange can end up public on GitHub, because the tools write commit messages, rule files and conversation logs into the project folder. We went looking and found 252 lines of developer rage, a shouting index that turns out to be a statistical draw, and 15,040 chat-history files sitting on public repositories, some as long as a short novel. The companion video is set to Grieg.
Anthropic filed the first S-1 from a frontier AI lab. Most of the $80.7 billion headline loss is a paper accounting entry. An 85% gross margin on serving looks comfortable, but it leaves out the training that stops models going stale, all of which the accounts classify as research. The note works out what the margin looks like under five definitions, puts Anthropic beside the Chinese labs and OpenAI’s reported figures, and follows the hundreds of billions in computing commitments the company cannot walk away from.
Two thirds of AI users on corporate devices were on personal accounts outside the company’s control. Most Finnish companies have no written rules for staff AI use, and the few good policies we found came from municipalities you have probably never heard of. The note sorts six kinds of information against three kinds of tool, picks out thirteen rules from published policies in Finland and four other countries, and covers what the law already requires.
Your documents are probably already in a US cloud. Adding AI through the same provider adds no new jurisdiction, but it does add new questions: what the provider can do with your data, whether the model runs in Europe, and what happens if you need to leave. We looked at the prices, the capability gap and the seven questions a company needs to answer about each route. The short version: the first step is business terms and costs almost nothing for staff already on a paid plan. The expensive part is control most companies don’t need.
Eleven worked cases are now published, each with the run behind it open to inspection. They have a section of their own.
Large language models produce answers without a reliable signal of how much trust those answers deserve. We are writing a practical Bayesian framework for calibrating uncertainty in LLM systems: when to act on an output, when to route it for human review, and how to measure whether the system is as well calibrated as it claims. It is written for practitioners.
Acerbi & Martin · in progress
When something is finished it will appear here. There is nothing to subscribe to, but if you would like to hear when a note is published, write to enquiries@hapax.fi and we will email you when one is. We use the address for that and nothing else.