Machine learning – Made by Mikal

Coding agents and deep learning

Post author:mikal
Post published:March 1, 2026
Post category:Machine learning

Andy sent me this post about how Carl Kolon uses and thinks about coding agents. The post is pretty interesting, and when I went to reply to Andy I realized that at seven paragraphs perhaps the reply is better suited to here than a signal message. Also, because that is more friendly than replying to messages with links to your blog to drive traffic and the adoration of your readers? This is of course a forward looking statement for me, I do indeed hope one day to have a reader here but baby steps. Similarly to Carl I certainly started using LLMs as "smarter search", cutting and pasting queries into claude.ai, and waiting patiently for an answer. I think this should make Google very worried, especially as its so good at finding answers. The improved performance over a "raw Google" is largely because of persistence -- the LLM doesn't perform single search, it will keep searching until it finds the answer. The decreasing levels of supervision Carl talks about is what I am talking about when I talk about prompts and planning, which has been my favourite LLM topic for the last few weeks. Increasingly I am writing a plan for…

I think I just experienced my first hallucinated security vulnerability

Post author:mikal
Post published:February 17, 2026
Post category:Linux Machine learning OpenStack Security

Yesterday my time I experienced what I think was my first LLM hallucinating a responsible security disclosure. Honestly it was no curl situation, but I think it was still interesting. The bug is on launchpad.net if you're interested in taking a look. I think in total I spent a couple of hours on the whole thing, with the hardest bit being trying to understand what the author was claiming. Fundamentally they had conflated being able to change the state of memory and other hardware inside their virtual machine with changing the state of those things for the hypervisor. They did not seem to understand that the video memory of the guest was not the video memory of the host for example. That said, I tried to be nice and I hope my replies were perhaps a little useful to them.

Ancient code, mental health, and AI tooling

Post author:mikal
Post published:February 13, 2026
Post category:Machine learning Pngtools

In the early 2000s I was in my mid 20s, working a dead end job as a Windows programmer, and had two very young kids who were not super good at sleeping. I had worked as what I would now call a systems programmer for the Australian patents and trade marks office for a few years in the late 1990s doing low level image manipulation code -- we had a for the time quite impressive database of scanned images of patents and trademarks, and sometimes we need to do things like turn them into PDFs or import a weird made up image format from the Japanese patents office. So when you combined those things -- previous experience in a field I found interesting, a job I did not currently find interesting, and a lot of spare time very early in the morning because the kids wouldn't sleep but my wife really did need a rest -- you end up with a Michael who spent a lot of time writing image manipulation code on his own time. Even back then I was pretty into Open Source, so I released what I think was probably the first Open Source PDF generation library,…

Talking to people about AI is hard

Post author:mikal
Post published:January 5, 2026
Post category:Machine learning

I am not entirely sure what makes the AI debate so polarizing, although I suspect it has something to do with feeling threatened by a changing landscape. What I can say with certainty is that I find having a nuanced conversation about AI with people often difficult. It seems to me that people fall into two polar opposite camps — those who thing AI is completely great and that we should fire all the programmers and creatives; and those who think that AI is all bad and we should go backwards in time to a place before it existed.

(more…)

Tensorflow, ROCm, and version mismatches

Post author:mikal
Post published:September 21, 2025
Post category:Machine learning Python

I’ve fallen into this pattern where I do an hour or so of self-directed learning in the mornings before going to work. Until recently it was an excellent CMU course on the design of SQL database systems, which I’ve mentioned previously here. I’ve finished that, so I thought I would do something shorter and fun as a break before finding another course to do. I chose The freeCodeCamp.org hot dog or not hot dog tensorflow course. 90 minutes seemed achievable, and I too wish to know if an object in front of me is a hot dog or not.

(more…)