1,841 mentions · 834 chunks · 141 episodes
...eading this week. Amelia Wattenberger's LLM fish eyes. It does track to me that LLMs will be an ingredient that allows new kinds of UX to become possible that wasn't possible before. The ability to generate high-quality summaries of p...
LLMs are best used as a thought partner, not as an oracle. They will often be wrong; you should assume anything they tell you is likely to be approximatel...
I like magic as a frame for things that are powered by LLMs. Normal programming is mechanistic. It does exactly what it was told to do, even if that's not exactly what the creator meant. But LLM-powered experi...
LLMs are like chainsaws, allowing you to hack and slash through dense problems. Before those problems might have been dense, foreboding jungles, with no g...
LLMs are now good enough to be better than all but the experts in any given domain. Which produces a problem: how can you judge if its answer is good in a...
LLMs should be good at generating possible multi-disciplinary insights. LLMs are worse than domain experts, but better than most everyone else nowadays. B...
LLMs are a good dowsing rod: is your hunch good enough to be worth digging deeper on? There are lots of questions where even an approximate answer says "y...
...running an LLM. The ruined society doesn't (yet) know how to build a way to run LLMs themselves. That would require millions of specialists with knowhow to produce all the inputs necessary to build chips and servers and program them, ...
An agent is a bit of software that is animated by LLMs. Not normal software that does precisely what it was programmed to do, agentic software that has some squishiness. Expensive, dangerous, intrinsicall...
The most important characteristic of LLMs is their patience[apc]. They can do tasks that real humans would get distracted or bored by. For example, carefully reading many pages of material to...
Don't use an LLM to write for you, use it as a thinking partner. LLMs' ideas are never good. They're always mush, just frog DNA. I was chatting with an author who told me he refuses to use an LLM. He told me you "write ...
LLMs are never blunt.[ape][apf] When someone is blunt they're definitely a human. You might love or hate their taste but they're definitely human.
LLMs are great at giving you the superficial vibe of quality. But the closer you look, the more you realize it's fundamentally hollow. A gilded turd. "Wow...
LLMs compress nearly all of humanity's background context into a teensy weeny little hyperobject package. They have effectively infinite background contex...
...eek I observed that the spec seems more important than the code in the world of LLMs. Which layer is the most important? The layer where you spend most of your time iterating. This is especially true if there is a robust, automatic tr...
...object with near infinite background context? If you sprinkle the pixie dust of LLMs on those old ideas, how many could spring to life?
...opelling them through the problem. This same ability is important in a world of LLMs too, where the hard part now is to sequence the incremental extensions of work in the same way a human programmer would, but have an LLM turn the exe...
One of LLMs' primary superpowers: they have human-level ability, but they never get bored. For example, human engineers get bored very quickly writing tests for ...
Frog DNA is average, mush, generic. That's one of the reasons LLMs pull everything to the bland centroid; they fill any ambiguities with mush. What if you could make it so instead of using frog DNA, it used shark DNA...
English specs can now be "compiled" to runnable code by LLMs. When we make programs, we compile the source code down to an executable binary. But then we make sure to keep the source code, so when we tweak it i...