A chat model writes. An embedder places. A decision model answers: it reads some state, answers typed questions about it (pick one, yes/no, a scale) and returns a probability per option. No text, nothing to parse. loomcycle has used one for seven weeks, in five places. New post on why we now treat it as a third kind of model, next to chat and embedder. 🧵 1/3 loomcycle.dev/blog/decision-mo…
Similar posts
3 posts close in meaning, closest first
That's to me the most impactful evolution in the field during the past few months. Thanks to progress in quantization methods and specifically-optimized inference engines such as Strata, it's now possible run relatively large MoE models such as Qwen 3.8 Flash Next (125B parameters) on regular PC. This is how we take back control. #LLM #qwen #strata #AI
I was sharing with my wife my (current) preferred approach to AI produced code by engineers. It's borrowed from her PhD - thesis defence. If an engineer can defend their code in a thesis defence, I have no concern about how it was produced. What is thesis defence for software changes? They should be able to explain, without using an llm: * What is the objective for the code change * Why did they choose this approach * What other approaches they considered * What unrelated/unexpected things they found along the way * What are the limitations of chosen approach * What is the wider/longer term impact of choosing this approach
Show more of this post
* Who else did they collaborate with, or who else understands their chosen approach to be able to answer these questions if needed Notice I don't want them to explain their work. I can read the code and understand it. I want to understand how well researched this approach is, and how well they understand it. I also want to ensure that they have put in at least as much effort (to produce the code or to prepare for the theses defence) as the reviewer will put in. #ai #softwareengineering
"AI is great at writing code. But it doesn’t know if you are launching the app or have millions of users, which it needs to make the right architecture call to log a timestamp vs email. It is also sycophantic and tries to execute your requirements exactly without pushing back, which means you end up with a slow app that takes 15 seconds to load a static welcome screen." manidoraisamy.com/developer-to…