Daniel Grady

Bit of a disconnect? "So far Anthropic has chosen not to participate in the ‘look what our unreleased model can do’ math olympics, but also math tends to be a relative weak point for Anthropic." thezvi.substack.com/p/ai-162-vis.... Vs www-cs-faculty.stanford.edu/~knuth/paper....

This is the week of "internet hot takes Claude 4," so … These examples are all with Claude Sonnet 4 with web search on, extended thinking off. General theme is that LLMs aren't very good at specific questions about how to use LLMs.

"Anthropic recently launched a new "Max" plan. Does this plan allow me to make API calls?" Basically right, but confusing. Which to be fair is also true of the docs. claude.ai/share/33f152...

This is the week of "internet hot takes Claude 4," so … These examples are all with Claude Sonnet 4 with web search on, extended thinking off. General theme is that LLMs aren't very good at specific questions about how to use LLMs.

"I'm working with FlatBuffers in Python. It seems like there is not a clean, idiomatic way to introspect on the data fields of a payload. Is this right? Why?" Right. (I think.) claude.ai/share/ac8c7f...

Nice essay from Oliver Reichenstein at iA. Nice essay. People learn by doing and making: although this is an obvious and agreed statement, it should be taken much more literally and much farther than it normally is. ia.net/topics/maker...

Echoes of this even in very abstract problem spaces: in data science, the number one thing is to look at the data, literally look at Claudia-Perlich-style, load the data into a spreadsheet and stare at it and explain what you see.

Oh wow. This morning, I mailed a physical letter to the CA AG requesting action against OpenAI's proposed theft (notforprivategain.org), and today OpenAI announced changes to their plan (www.wired.com/story/openai...). Love seeing this level of efficiency from the USPS!