Armchair Data Scientist and nerd. PhD in automated theorem proving just before AI became a subfield of Linear Algebra. Machine Learning, statistics, data visualisation, biology, recreational maths, computation, old pocket calculators, and video games
Benchmarking the MOE Gemma 4 vs Claude for writing Z3 specification: Local model is not quite there yet, Claude basically one shots everything correctly, local model has a hard time with a less familiar spec language.
Released this to help customize it - it packages the normal not behavioral stuff and lets you include your own. I included some default options but fully customizable per project, or global config. It just runs Claude code with a custom dynamically creates prompt
A foundation model trained on 36M cells across 8 species creates a universal embedding space for cell typesβno fine-tuning needed for new data. It recovers developmental lineages and cell functions it was never explicitly trained on.
One common critique of AI is that it's imprecise and nondeterministic. We programmers have held too long to the notion that precisionβin algorithms, in the programming languages we use, in the data we collectβis essential. 1/13
I went looking for one Japanese chart book and ended up with 19 prewar volumes, a new collecting method, and a stronger conviction that graphic history is far from fully known. www.chartography.net/p/ascending-into-the-realm-of-japanese
Someone on Reddit already solved the Claude Code bug that ate everyone's token limits:
This might not be 100% of it (it's a lot of code) but it's 100% at least part of it. And it explains why Anthropic have trouble catching it internally.