will it blend - live

Makefile

@@ -1,6 +1,6 @@
 # Makefile for Firehose monorepo
 
-.PHONY: check precommit deps compile test format
+.PHONY: check precommit deps compile test test-elisp format
 
 # Common check target that runs all static analysis
 check:
@@ -23,6 +23,10 @@ test:
 	@make -C app test
 	@make -C blogex test
 
+# Run Emacs Lisp tests for roam-export-blog (requires emacs on PATH)
+test-elisp:
+	emacs --batch -Q -l org -l ox-md -l scripts/roam-export-blog.el -l scripts/roam-export-blog-test.el -f ert-run-tests-batch-and-exit
+
 # Format code
 format:
 	@make -C app format

app/priv/blog/engineering/2026/10-08-will-it-blend.md

@@ -0,0 +1,50 @@
+%{
+  title: "Will It Blend?",
+  author: "Willem van den Ende",
+  tags: ~w(AI pi.dev ComprehensionMaxxing),
+  description: "Will it Blend was a series of videos promoting a blender brand, in which various unlikely objects were put in a blender to see if they would blend.",
+  image: "/images/blog/2026/will-it-blend.svg",
+  published: true
+}
+---
+
+**Problem** You find something interesting, but don't completely understand it, or are not sure where to start.
+
+**Solution** Point your coding agent at the thing you read (a blog post, research paper, product announcement) and build something so that you can play with it.
+
+Will it Blend was [a series of infomercials](https://en.wikipedia.org/wiki/Will_It_Blend%3F) promoting a blender brand, in which various unlikely objects, such as tablets, were put in a blender to see if they would blend.
+
+![will-it-blend](/images/blog/2026/will-it-blend.svg)
+
+I like it as a _metaphor_ for trying to get something to work that you have just come across. I often use it when reading about the new hot AI thing of the week (say decision models a couple of weeks ago). I’ll go and look for an open source implementation and have a play with it, asking my coding agent for help installing it and coming up with a micro project I can do with it.
+
+Worst case I lost half an hour to an hour, and put the idea back on the backburner. But more often than not, I get something out of it that I can take forward: working software and better, more grounded, understanding of a topic.
+
+
+## Examples of when I used this
+
+I use this all the time now. And so do other people, it is not that original. I want to encourage you to see if things blend too, only takes half an hour or so.
+
+### Decision models
+For instance, a day after [Jev](https://en.wikipedia.org/wiki/Jev_(AI_model))came out, Laya popped up on Reddit. I cloned the repository, opened Pi and asked some questions about what you could do with it. Then I had a look at the example code on the Laya page, used my coding agent to run it, had a chat about how I could use it in a coding agent, and had the agent make some sample questions with sample data. I also wondered why it was taking so long to start, got the coding agent to reduce startup by 40 times (28 seconds to about 0.6 seconds or so), and I learnt in the process that this was a BERT model with some pre-training. I had seen BERT’s but not played with one in the wild before. This enabled me to write some comments on LinkedIn, and learn, indirectly, about my mistakes :-). Not bad for an hour or two on a sunday afternoon.
+
+### Getting started with a local model in an open coding agent
+Another example would be John Nolan , who my blog post on a local coding agent on Mac OS through his coding agent to get Pi + local llms set up on his machine.  TODO source needed (LinkedIn). I liked this, because this is also how I got started with making my own coding agent, and later installing Pi. With a coding agent in hand, and some vigilance or sandboxing, this is not much time well spent.
+
+## Background
+
+I learnt this in the spring of 2025 through  Julia Turc, probably from her video on  [Mixture of Experts: How LLMs Are Getting Smarter Without Getting Slower](https://www.youtube.com/watch?v=7yR5ScbK1qk). In this video, Julia takes a 1991 research paper on mixture of experts, pulls it through ChatGPT (this is several hundred AI summer years ago) and creates an executable Collab Notebook out of it.
+
+I wrote a comment at the time:
+
+> +1. I was wondering if llms would make it easier to implement papers (something I haven’t done in a while), and later that day, youtube surfaced this video with a live demo. I came for the mixture of experts explanation and stayed for the implementation and paper reading. It is insightful to see what you look at in a paper.
+
+I was primed for this by  Katherine Breslin’s talk at Agile Manchester 2025 (spring, this is important in AI summer years), in which she mentioned making a voice assistant in an afternoon on her own, something which had taken the Alexa team (on which she worked) 150+ people and multiple years.
+
+## Your turn
+
+Do you have research papers, blog posts or videos that leave you wondering? Wonder no more. These days I often start with putting my question in a (gemini) notebook, getting a summary, finding more sources, and asking questions. Then I take a summary document to my coding agent, refer to that in an initial prompt and then it usually works.  I create a new git repository for every project, making small commits as  I go. Often [with tests](/blog/engineering/synthetic-tdd): they make it easy to get started, express expectations, and keep going if the experiment is a success.
+
+If you want specifics on what I use now: I use [Pi](/blog/engineering/tag/pi.dev) as coding agent. Models that I can run on a 64GB mac (e.g. Qwen 3.8 27B or 3.6 35B) are usually plenty. When I want it faster, or with less fan noise, I use a hosted GLM 5.3 Flash.
+
+I hope you find this useful and will see if something you are curious about will Blend. Happy blending!

scripts/.gitignore

@@ -0,0 +1 @@
+*.elc