
max_tokens=700 on a reasoning model returned empty replies — hidden thinking tokens was the whole budget
For three nights straight my job-digest pipeline produced nothing. The pipeline pulls raw job-posting...
Tag archive

For three nights straight my job-digest pipeline produced nothing. The pipeline pulls raw job-posting...
Explore how sharing AI progress in mathematics is reshaping technology. Analyze the shift from...
The data says reasoning models make developers 19% slower. Here
In an earlier post I wrote off gemma4:12b's empty replies as a packaging bug. They weren't: it's a reasoning model. So I ran 13 questions with thinking ON and OFF. Reasoning got one more answer right while spending 68× the output tokens and 19× the wall-clock. Here's when I now turn it on in an agent, measured.
This week's tooling news splits cleanly between performance and compliance: a Go Protobuf parser that...

DeepSeek R1 vs V3 in 2026: When to Use Each (and Why They're Merging) Here is the trap...
ZAYA1-8B packs 760M active parameters into an 8.4B MoE that beats DeepSeek-R1 on AIME 2025. Here is what developers need to know.
How DRA-GRPO uses Submodular Mutual Information to fix GRPO's diversity collapse problem—with a minimal Python PoC and benchmark results from arXiv 2505.09655.
Complete developer guide to the OpenAI o3-pro API. Pricing, benchmarks, Responses API setup, background mode, reasoning effort, and when to actually use it.