AI Research

Experiments, evaluations and notes on machine learning systems — what I tried, what the numbers said, and what I think it means.

3 articlesRSS

Subcategories

Interpretability

Opening up models: probing, feature attribution, mechanistic analysis.

1 article

Agents & Tooling

Agent architectures, harnesses, tool-use patterns and the infrastructure around them.

1 article

Benchmarks, eval design, and the ways measurement goes wrong.

1 article

TopicRepliesVotesActivity

Attention Residuals: Kimi Team swaps the fixed residual sum for learned attention over depth
AI Research Interpretability kimi attention-residuals llm-architecture residual-connections
Kimi's Attention Residuals paper replaces the fixed residual sum with softmax attention over layer outputs. The mechanism, the 1.4T-token test, and the caveats.
0 replies0 votes1mo
0reply
0vote
1mo
Go + HTMX vs React at enterprise scale: what the primary sources actually document
AI Research Evaluation htmx go react hypermedia
The 'Go + HTMX beats React' debate rests on a Carson Gross essay and one case study. We checked what the sources document versus what one video extrapolates.
0 replies0 votes1mo
0reply
0vote
1mo
Mojo 1.0 Ships Under Qualcomm, and the Independent Benchmarks Are More Nuanced Than "CUDA Parity"
AI Research Agents & Tooling mojo modular cuda gpu-programming
Modular shipped Mojo 1.0 after Qualcomm's acquisition. The Oak Ridge paper behind the 'CUDA parity' claims is real — but it documents gaps, not full parity.
0 replies0 votes1mo
0reply
0vote
1mo