True Positive Weekly #138
The most important artificial intelligence and machine learning news and articles
Hey, in this issue: AI Scientist’s history; Seven thoughts on “AI Scientists”; the mirage of autonomous AI scientists; people see text, but LLM not; a tutorial on finetuning LLMs using reinforcement learning; using generative AI for data visualisation; tutorial on how to inference cost by 97% while preserving 96% quality; and more.
AI Scientist’s history plus Seven thoughts on “AI Scientists”
People see text, but LLM not
[Tutorial] Blend hybrid retrieval with structured data using MindsDB Knowledge Bases
[Project] Generative AI for data visualisation
[Project] Distilling intelligence: How to inference cost by 97% while preserving 96% quality
[ChapterPal] Less is More: Recursive Reasoning with Tiny Networks
Enjoy the newsletter? Please, don’t forget to like or comment to help the newsletter grow.
Honored to be featured alongside the "Less is More: Recursive Reasoning with Tiny Networks" piece. Feels like two sides of the same coin.
TRM shows a 7M parameter model beating billion-parameter LLMs through recursive architecture. My distillation work showed a ~7B model beating a 30B model through better pretraining alignment (instruction-following vs code-completion).
Different mechanisms, same conclusion: how you train and structure a model matters more than raw parameter count.
Thanks for curating these together, Andriy.
Fantastic curation this week! The juxtaposition of "people see text, but LLM not" with the recursive reasoning paper is particularly thought-provoking. It highlights a fundamental tension: while we're pushing towards more efficient reasoning through smaller networks that iterate, we're simultaneously discovering how differently LLMs parse information compared to humans. The 97% inference cost reduction project you linked is a perfect example of where this matters practically. If we're distilling models down while preserving quality, understanding how they actually process textversus how we think they process it becomes crucial for maintaining that quality threshold.