Tag
The article argues that while AI is paid for by tokens, tokens are not the true unit of inference because efficiency improves with model advancements, suggesting a new unit based on problem-solving chains.
Halo modifies deep forecasters to include scale estimation via heteroscedastic learning, significantly improving point forecast accuracy in benchmarks on electricity price markets.
An autonomous research program by Qiushi Engine conducted end-to-end research on BabyLM 2026 Strict-Small, improving data-efficient language models through principle-guided methods and achieving the highest score in the public snapshot.
Andrew Chen argues that open weight AI models are improving rapidly and will cover most consumer/prosumer use cases, leaving frontier models to compete for the remaining high-value 10% of applications like coding, science, and math.
The updated Grok model (0.5T) is less lazy, more autonomous, and more accurate; improvements are ongoing.