Posts
All posts.
Agents against broken ML workloads
The BrokenML benchmark runs nine models against 46 broken jobs at three access levels: configuration files and code only; live system logs; and full root access.
Half-written answers
What happens to an RL rollout when the sampling policy receives an update mid-sentence?
Decode Decoded
Decode performance of vLLM, SGLang, and TensorRT-LLM on H100 SXM across Qwen 3 0.6B-32B models for 1-256 batch sizes analysed with Nsight Systems.
Specialisation & Co-design
How specialisation and co-design work together in modern ML systems through careful boundary selection, interface design, and joint optimisation.
A window
Achieving synthetic general intelligence is likely impossible without decoupling the models from human interfaces.
Keeping up
Policies often balance short-term and long-term perspectives.
Hidden values
Language and audiovisual interfaces inform the current landscape and evolution of widely adopted AI applications.
Emergent grammar
We are overly optimistic when finding local structure in the environment and consistently underestimate its emergent properties.
Wait calculation
As with the interstellar travel wait calculation, we might be nearing a time when investing in non-AI scientific projects could prove suboptimal.
Limited throughput
There are only so many interactions with synthetic agents humans can maintain daily.