Notes
Notes from my projects and what I'm reading about AI, security and the money behind it. Practical work alongside research, with sources for further reading.
12 notesEngineering & research
- B300s are starting to look like cloud instances, not mythical hardwareB300 access is spreading across GPU clouds, with public spot and preemptible prices. That is a real shift, even if it doesn't mean unlimited capacity on demand.3 min read
- Benchmaxxing is a real incentive, but a high score isn't proof of cheatingThere is solid evidence that training aimed at a benchmark task can move scores and rankings. That is different from proving a particular lab secretly trained on the test set.3 min read
- October was a rough month to be responsible for securityPHP, LiteLLM, certificate infrastructure and Binance all had a bad October. The incidents were different, but the same old weak links kept showing up.3 min read
- What if an LLM could read a neural network's weights?A set of experiments asks an LLM to inspect, edit and even write neural-network weights directly. The results are wild, and still very early.3 min read
- Connecting a team's autonomous stack to a simulatorMy work connecting Auria to CarMaker and EUFS Sim 2, with NATS at the integration boundary and Rerun to inspect a run.2 min read
- Read hyperscaler capex like a model metricBig Tech will spend well over $700 billion on infrastructure this year. The headline number depends on accounting choices, so I read it the way I'd read a sensitivity figure, operating point first.3 min read
- Three numbers for the AI market, and the fine print on eachVenture money, lab revenue and the price of intelligence all hit records this year. Each one has an accounting footnote that changes how you should read it.3 min read
- When a prompt becomes a shellTwo Semantic Kernel bugs this year turned prompt injection into code execution, and web pages are now hiding instructions for AI agents. The lesson is old: the model is not a security boundary.3 min read
- AI found twelve OpenSSL bugs the same month curl gave up on AI bug reportsIn January 2026 an AI system found every vulnerability in an OpenSSL security release, and curl closed its bug bounty because of AI-generated noise. Both stories are true, and they are the same story.2 min read
- Tokens keep getting cheaper and AI bills keep going upA curiosity of AI economics. The price of a fixed level of intelligence falls about 13× a year, yet teams spend more every quarter. Both are true.2 min read
- Looking inside the model is a breakthrough, with an asteriskMIT Technology Review put mechanistic interpretability on its 2026 list of breakthrough technologies. What it can do today, and why researchers still argue about how far it goes.2 min read
- Coding agents you can run on your own machineA short field guide to open-source coding agents that work with local models, what each is good at, and the one limitation they all share.2 min read