1,841 mentions · 834 chunks · 141 episodes
A mathematician told me that LLMs are great at "bazooka style proofs." That is, proof by exhaustion. You don't need induction if you can try every single hammer to see which ones work...
... better, shorter and shorter prompts perform better than more complex ones. The LLMs are good at navigating a superposition of possibility; if you give them too structured an envelope, you cut off some of the possibilities. A friend: ...
...ing, but it otherwise still works. This can lead to weird outcomes that get the LLMs confused. You can play with their memories. For example, if you have a versioned AGENTS.md, each time Claude.app boots up, it seems to put the curren...
Paper: LLMs prefer text written by LLMs. This implies that communication will become more and more Claudish, as more and more of the reading is being done by LLM...
LLMs in a harness with memory sometimes develop a viral misunderstanding. One LLM misunderstands something and stores it in the knowledgebase. Future agen...
...kable. Like a grotesque secret version of exquisite corpse. This was all before LLMs became big… I shudder to think what's happening now.
...just ad hoc and sloppy. The right combination of deterministic code and squishy LLMs will be like chocolate and peanut butter.
The Chat UX is a vestige of when LLMs responded quickly. You'd send a message, and then while you were watching get an answer, starting instantly. Now, with thinking tokens and sub-agents...
LLMs do magic at non-trivial expense… and current systems just throw that out. Often we don't cache or keep around the result. For example, distilling a b...
It turns out that a killer use of LLMs is writing code. They are insanely good at it. No matter how much of a crash there might be around AI, we're never going back to writing software the...
Today we often use LLMs to tell us what is true. Hopefully we won't have LLMs tell us what is good.
...you write just for yourself, it is inherently an authentic act. Even if you use LLMs to write, if it's for you, it is still meaningful. Making slop for someone else is often a waste of their time.
LLMs make new methods of team consensus possible in software development. Software used to be expensive, so we had to decide which design was best before ...
Humans and LLMs extract insights out of information differently. Humans don't just passively receive information; they interact with the world to tune their informat...
When we anthropomorphize LLMs we make it easier to erroneously assume that they will respond reasonably… but also that they have a human-style sensorium. The chatbot has a limited...
...ntive. Polymarket etc now make a direct incentive for this kind of thing, while LLMs make the cost of doing it collapse by orders of magnitude. The bar to clear is tiny, and can be as small as "I dunno, I just wanted to see what would...
LLMs allow JIT situated answers, for queries and also glue code. Before, the answer to your query had to already exist before you searched. That content w...
...t been alive. When the external life support stops, they rot. But software with LLMs involved can be alive. It can be self-maintaining, to make itself forward-compatible, automatically. As long as the basic runtime it runs on continue...
..., you'd likely lose out on significant improvements from compilers. Coding with LLMs will be the same. Of course you should do it a few times to develop a situated intuition and know that if you had to pop the hood you wouldn't be tot...
A paper that just came out from DeepMind: in the world of LLMs, writing styles, etc are homogenizing. Notably, this occurs even if we don't use the same model, since different models are trained on the same input...