What the tokenizer hides.
A byte-level language model starts with the actual encoding of text. Byteification asks whether we can get there without throwing away a useful pretrained model.
WRITING / FIELD NOTES
Blog
Notes from the work of making complex technology useful: practical experiments, engineering decisions, and the people and systems behind them.
02
What gives a data center away
A look at 38 large projects finds that code names are weak clues. The parties behind land records and jobs promised per dollar offer a more useful lead, with important limits.
03
The interface starts in chat
The traditional dashboard is obsolete. The future of software is building the tool you need, around the question you have, inside a conversation.
04
What’s a PINN?
A physics-informed neural network combines a neural network with the equations a system is known to obey.
05
A paper you can put to work.
Paper2Agent turns research methods into callable tools. The useful test is whether those tools execute the method faithfully, not whether they can talk about it.
06
Stop recomputing what you can look up.
Engram adds conditional memory to a language model. The useful question is which work deserves computation and which patterns can be retrieved.
07
When the inputs change
Build OOD observability in PyTorch, then turn reviewed data into a retraining job without turning every alert into a release.
08
Which layers should get a say?
Attention Residuals revisits how a network combines its earlier work. Explore what changes when every layer no longer gets the same weight.
09
When the model should ask for help
Sometimes the most useful answer is: someone needs to look at this. Designing that handoff is part of building the model.
10
How good is the measurement?
Before we argue about the model, we should check the instrument. A precise-looking number can still lead to the wrong call.
11
Where the work waits
A team can be busy all week and still struggle to finish anything. Follow the waiting before adding more work.
12
The detail needs its context.
AlphaGenome brings a long sequence and fine-grained outputs into one model. Explore why a local change can be hard to interpret through a narrow window.
13
A good image label is not enough.
DINOv3 puts useful local features back in focus. Explore why the relationships between image patches can matter as much as the overall image representation.
NO MATCHES / TRY AGAIN
Nothing on this page matches that search. Clear the search or choose another topic.