This is the first post on this blog, and it has been rewritten twice. The original, from August 2023, said I was thrilled to embark on this new adventure and would be exploring machine learning, deep learning, natural language processing and computer vision. It committed to nothing and told you nothing.
Here is the version with content in it.
What this blog is
Notes on building things with AI, written by someone building them. I am Abhijoy Sarkar. I work on agent infrastructure and LLM security, and I build a lot of small things in public.
The rule I try to hold myself to now, which I did not for the first two years, is that a post should contain something I found out by doing it. A measurement, a failure, a number that surprised me. If a post could have been written by reading three other posts on the subject, it should not exist.
What is worth reading first
- A one-page policy matrix for agent tools. The permission model I use to stop untrusted text from reaching privileged tools, with the gate that enforces it and a repro of the attack it prevents.
- AI command automation. What happened when I tried to make a local model drive my computer, and the discovery that most of my “prompt engineering” was a parser I had not written.
- Building a mini language model. An RNN and a transformer trained on the same data, which finish in a dead heat, which is not what I expected and turns out to be the interesting part.
A note on the archive
Posts from 2023 and 2024 have been rewritten. They contained real errors: a wrong description of sliding-window attention, a claim about a benchmark that the paper in question never ran, an assertion that language models do not retain their training data, and, worst of all, a corporate case study that I presented as evidence and which never happened.
Each of those posts now opens with a note saying what was wrong and what changed. I would rather carry the corrections visibly than quietly replace the text, partly because the errors have a pattern worth naming: every one of them came from writing about work rather than from doing it, and from citing summaries instead of sources.
If you find another one, tell me.