Radar ·
Radar: data-constrained scaling laws for undertrained LLMs
Radar briefing on a post arguing that LLMs are heavily undertrained and proposing data-constrained scaling laws, questioning Chinchilla-style assumptions about optimal scaling.
The post, dated November 2, 2026 in Radar, argues that LLMs are heavily undertrained compared with the number of synapses in the brain and with token exposure. The author presents data-constrained scaling laws and learning algorithms for very large transformers. It also questions optimal-scaling assumptions in Chinchilla-type compute and proposes scaling parameters and data differently. The source gives no numerical results, experiments or link; its stated scope is theoretical and propositional.
To consult and verify the original, search for the exact title in preprint repositories or on the author's blog, since the bookmark records no address. Compare the claims about synapses and tokens with the sources cited in the original before reusing them. Because the topic is model research and involves no personal data, the relevant protection is in use: do not paste fragments containing customer or employee data into prompts while studying the material.
Original source: Leis de escala com restrição de dados para LLMs subparametrizados