Qwen3.8-Max arrived with strong multimodal results, so I wanted to try something more concrete than visual Q&A: give it an image, ask for structured bounding boxes, and draw the answer back onto the pixels. This is a common workflow also supported in previous models.
Satya Nadella recently published an excellent article on what a future firm looks like in an AI-driven economy. He also introduces the concept of “token capital” which now exists alongside human capital (and financial capital).
… or how to process DataFrames with RLMs and DSPy
Before we begin: Of course this doesn’t replace a proper (human) audit or security testing process, nor should it. But this was a fun afternoon experiment.
Originally written September 2025 - a lot has changed between now and December when I’m posting this!
There’s been lots of discussion recently on DSPy and the GEPA optimizer. And for good reason: the results are compelling. As evidenced by the below experiment using automatic prompt optimization, we’re seeing 20+ percentage point improvements in exact match accuracy over vanilla LLM structured output calls with little engineering effort required. This simple example demonstrates how much…
DSPy is becoming increasingly popular (at least in my bubble on X) - and, in my opinion, for good reason! It provides a sense of control and composability that becomes addictive once you get a few reps and understand how it all fits together. It allows you to inject LLMs into your program’s control flow, providing useful leverage.
Link to Paper
Tools Used: bat ffmpeg llm* mlx_whisper pdftotext shot-scraper* uvx
Large Language Models (LLMs) have burst into the conversation and have already proven incredibly powerful in accelerating all sorts of knowledge work. Initially explored by the wonderful open source library instructor (and others), the concept of generating structured outputs from unstructured text is an extremely powerful yet simple concept. Think PDF in, Excel out; for example, using a vendor…