Claims Browser

Parameter count is only meaningful alongside data quantity, compute allocation, and deployment conditions

"Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions."

Future ML research effectiveness depends more on training generally capable models then applying them to AI R&D rather than narrow training on specific R&D tasks

"My guess is this actually is not all that effective, and you would do better by mostly training a generally capable model and then turning it to AI R&D. Bitter lesson."

Z.ai has particular strength in post-training compared to Kimi which excels more at pretraining

"To risk a broad oversimplification, Z.ai seems to have a strength in post-training when compared to Kimi, which is more of a pretraining masterpiece."

Scaling generated distillation data should be paired with trait-aware curation and evaluation, even when data appears off-task or benign

"Our results suggest that scaling generated distillation data should be paired with trait-aware curation and evaluation, even when the data appears off-task or benign."

Current AI techniques are 4-6 orders of magnitude away from optimality in terms of data efficiency and test-time compute efficiency

After o3's test-time compute demonstration in late 2024, the new models showed genuine fluid intelligence and the LLM line of research could achieve unbounded capability scaling with no wall

There is an underserved market for tiny MoE models

Test-time compute (TTC) and harnesses are critical for AI capabilities, and the TTC breakthrough was not obvious

LLMs can work as a base to build systems actually capable of fluid intelligence, following the o3 test-time compute breakthrough in December 2024

The death of tokenmaxxing is a likely contributor to Anthropic's declining metrics, in addition to competition from Sol and opensource

Sudden silence on previously shared metrics is a bad sign for a company's performance

Deep learning may be hitting a wall in its progress

Scaling laws for AI have fundamental limitations that were predictable and documented

DeepSeek is finally relevant again after over a year of comparative obscurity, with V4 Pro in April 2026 as an exception

DeepSeek's V4-Flash update timing is well-positioned after their $70B pre-IPO fundraise

Whether Aschenbrenner's bets were correct long-term remains to be seen

The pace of progress on models from multiple organizations at once is genuinely incredible

Building LLMs isn't driven by rare secrets, but consistent effort, mass capital, and effective organization design