AI & Compute
Where AI systems actually fail in production
Not in the model. In the data pipeline, the distribution shift, the feedback loop and the assumption that the world stays still.
AI & Compute
Models, chips, data centres and the claims that outrun them.
AI & Compute
What to actually worry about with AI
A great deal of the risk discussion is about scenarios, and a shorter list of problems is already causing measurable harm.
AI & Compute
Evaluating an AI product claim
A short checklist for reading announcements, which mostly consists of asking what was measured and against what.
AI & Compute
Small models, distillation and the case against always using the biggest
Capability per parameter has improved faster than parameter counts have grown, which changes the economics of what to deploy.
AI & Compute
Retrieval, context and grounding a model in your data
Supplying source material at inference time is the most reliable fix for fabrication, and it moves the difficulty into the retrieval step.
AI & Compute
Agents, tools and the gap between demo and deployment
Giving a model tools and a loop makes it far more useful and introduces reliability problems that compound with every step.
AI & Compute
Open weights, closed models, and what the distinction means
Releasing model weights is not the same as open source, and the practical consequences differ from the licensing argument.
AI & Compute
Training data, copyright and the unresolved question
Models are trained on text and images their creators did not license, several jurisdictions disagree about whether that matters, and the courts are still working.
AI & Compute
How models are actually deployed
Serving a model in production is a systems problem — batching, caching, routing and cost — and it looks nothing like running one on a laptop.
AI & Compute
What a GPU actually does, and why it is scarce
Matrix multiplication in parallel, memory bandwidth as the real bottleneck, and a supply chain with a single point of concentration.
AI & Compute
Benchmarks and what they fail to measure
A score on a public test set tells you about the test set. Contamination, saturation and construct validity all get in the way.
AI & Compute
Why AI needs so much electricity
Training is a large one-off cost. Inference is a small cost repeated an enormous number of times, and it is the one that scales with success.
AI & Compute
What a large language model is actually doing
Next-token prediction at scale, which explains both the capability and the failure modes better than any metaphor about thinking.