Things I Learned - 06 Apr 2025
This week, I learned: <select> will soon be very customizable via CSS. Including custom HTML inside options - even SVG. MDN. Edge/Chrome already support it. The Vitali Set is every real number none of whose difference is rational. A sparse collection of irrational sets. It’s like a line but doesn’t have a measurable “length”. The Lebesgue measure measures the length of broken lines. You add up the lengths of the smallest continuous intervals that cover the line. The Cantor set (take a line, drop every middle third, repeat) has a Lebesgue measure of 0 because the sum of the removed thirds = 1/3 + 2/9 + 4/27 + … = 1. You’ve removed every “length” though infinitely many points remain. The Vitali set built so that if you shift it by every rational from -1 to +1 and add them up, you definitely cover every real from 0-1, but never anything beyond -1 to +2. So the length must be between 1-3. Yet, there’s no number you can add infinitely many times to get something between 1-3. If you add up multiple unmeasurable sets like the Vitali set, you can get any total length you want. The Banach Tarski paradox splits a sphere into unmeasurable sets and adds them to get 2 spheres. Ctrl+Alt+F1/F2/… on Ubuntu switches the terminal. Typically Ctrl+Alt+F2 switches back to Gnome. But it’s a useful hack if Gnome freezes and you need to kill a process. Press Ctrl+Alt+F3, log in, and kill what you need. Notes from AI 2027. BTW, this is the most impactful piece I’ve read recently. It’s been on my mind continuously for 36 hours. A bit distubring, too. 2025: AI can act as autonomous agents, like Glean, Devin, Operator. turn bullet points into emails take instructions via Slack or Teams and make substantial code changes on their own spend half an hour scouring the Internet to answer your question 2026: automating AI R&D is the biggest enabler for AI Labs job market for junior software engineers is in turmoil people who know how to manage and quality-control teams of AIs are making a killing 2027: potential demand for ~20,000 FTEs solving long-horizon tasks to train AI every researcher/coder becomes the manager of an AI team hiring new programmers has nearly stopped, but there’s never been a better time to be a consultant on integrating AI into your business CSS Speech is a W3C spec that lets you control how screen readers should read pages. No browser support now, though. Clipboard2Markdown is a utility that lets you paste rich text and convert it to Markdown. ChatGPT can’t yet create good sketchnotes. Here’s the impact of US tariffs on India. ChatGPT #IMPOSSIBLE OHDSI has a vocabulary you can download from Athena that includes ICD codes and a lot of medical data standards. It also has a hostable WebAPI No open source LLM-based tool handles live transcription and allows you to query notes so far during the transcription. The closest seems to be Meetily Learnings on AI code editors via Deep Research from ChatGPT, Gemini, Grok, Perplexity: #ai-coding GitHub Copilot can identify the source of a code snippet as a repo. That helps with copyright issues. Cursor uses a shadow workspace - a temporary sandbox where it edits files before applying changes at one shot. Cursor auto-complete has context of other files, i.e. inserting an class in a .js file based on another HTML file’s contents. Windsurf seems to be best for large code bases and for large-scale refactoring. It can also run test results fix them. Windsurf includes a browser and lets you click on an element and prompt to change its behavior, etc. That’s good for front-end developers. Roo Code can run scripts as part of the workflow, letting you run linting, tests, starting web apps, query databases, etc. Roo Code lets you create persona, e.g. code reviewer, data storytelling and analysis, etc. with access to different tools and behaviors. Roo Code does not support auto-complete. There’s outrage around Cursor not taking responsibility for a rules file backdoor (via Grok Deep Research) and pricing. Zapier has an MCP server. That should make most integrations easier. Airflow AI SDK is a clever idea. Airflow is a workflow system. Agents are a workflow system (sort of). This SDK exposes LLMs as Airflow tasks. Hidden Factual Knowledge in LLMs finds that the hidden states in LLMs contain much more knowledge than they share. (Sort of like sub-consciously knowing the answer.) Even after asking 1,000 times, the answer is not expressed. ChatGPT Reasoning to Learn from Latent Thoughts finds that the internal reasoning process of LLMs is useful to train other models. Notes from AI Engineering Summit, NY, Day 1 When deploying in production, you need reliable output with fundamentally unreliable components. Sort of like how the ENIAC worked with 17,000 vacuum tubes that would fail every few hours. This is a reliability engineering subject matter and needs to be thought of that way. Google Follow up Deep Research queries are a natural way to extend knowledge beyond just a single report Deep research offloads less relevant parts of the context to a separate memory store for selective retrieval later. Anthropic Don’t use agents if workflows can do the task. The reliability of each individual step of an agent is critical. Code, file access, search. These are the top three tools to use. Making agents budget aware can help deploy reliably in production. Having multiple agents like sub agents can help protect the main agents context window. Self evolving tools are a useful next step in the evolution of agents. Software development lifecycle is about how we iteratively improve consistently without getting worse. Almost like the scientific principle. Morgan Stanley It’s easy to improve knowledge in a problem. It’s very hard to influence skin in a problem. Reinforcement learning from deepseek seems one of the most promising approaches that allow llms to learn skills I published an eBook on Amazon. It takes about an hour if you have the content ready. Set up a Kindle Direct Publishing account with your address, bank details, and tax information. (10 min.) Export my London 2000 blog archive and convert to Markdown. (15 min) Reformat the Markdown by writing a script in Cursor (10 min). Here’s the prompt: Write a Python script that reads *.md including the YAML frontmatter, adds the YAML title as H1, date (yyyy-mm-dd) like Sun, 01 Jan 2000 in a new para after the frontmatter and before the content. ...