Weekend Reading List — 18/04/2026
Have a great weekend.
Links
“Long live the harness (wrapper)” - Jamin Ball. Interesting read on new research from Stanford that proves you can significantly improve model performance (i.e. on benchmarks) by improving the harness. This is intuitive (and in some ways obvious) for sure, but is a strong signal that there is still great value in building great applications and harnesses around models, and that we may be a ways off generalised models themselves eating the application layer.
“How to make sense of AI” - Cedric Chin. Not the first time I’ve featured one of Cedric’s articles on here. I find he has a very compelling writing style, and this is no exception. A good reminder that, the best way to make sense of any new technology is not to just read endless “takes” about it, but to use it, and talk to others using it. Tinker away.
“The Cathedral and the Stones: What “Next Word Prediction” Actually Means” - Jinx. At this point, the “stochastic parrot” and “it’s just next token prediction” argument feels kinda lame. This is a good piece on why those arguments don’t make sense (anymore). There is some good healthy debate and push back in the comments, and this topic is certainly above my pay grade intellectually, but I honestly think it’s just semantics at this point. Whether the models are actually reasoning or just pretending to reason, you seem to get better results when they do whatever it is they are doing. So, does it matter?