1,841 mentions · 834 chunks · 141 episodes
People who think LLMs can't be useful either haven't tried them or aren't curious enough. Curious people (especially ones with a need for cognition[ajv][ajw]) will keep pu...
...ge evaporates. Prompt-fu is the equivalent for getting much better results from LLMs. That implies that people with prompt-fu will get less and less differentiated results as the systems we all use get more mature. But presumably ther...
...y do a surprisingly great job at predicting what a given sequence will fold to. LLMs are tapping into a hidden structure of the universe that reveals itself only if you are patient enough to sift through it.[akb][akc] LLMs are patient...
There are a lot of new techniques to "frack" LLMs to squeeze more out of them. If you get just a single english-language append-only log designed for a human, you make the LLM take its rich understan...
You can view LLMs as a lemon to squeeze. The quality of an LLM's juice is tied to who is doing the squeezing. An implication of all outputs from LLMs being co-created ...
...ference time has wildly different characteristics. But a lot of discourse about LLMs doesn't differentiate the two. I see even technical people muddle this all the time. There's a difference between an LLM in training absorbing a holo...
...viders to electricity providers. Capital intensive, but commodity. But actually LLMs are a worse business than electricity providers[ako]. Electricity providers typically have a geographic monopoly; end consumers only have one option ...
Someone got an LLM to create a pretty good stand-up routine for other LLMs. My first reaction was: wow, the LLM is pretty funny! But I realize that that's partially an illusion. The LLM, working in tandem with the human, is ...
The append-only conversation log to interact with LLMs leaves me wanting more. In the past few months I've had a number of revelatory, multi-hour conversations with LLMs where I explore and pull on thread...
Interacting with LLMs today is like doing computing before the GUI.[ami] Clear raw potential that it takes specialist skills to unlock and apply. The command line was a re...
A nice little use for LLMs: discovering best practices. For example, download three different popular templates for Project Management from Notion, and provide them to the LLM,...
LLMs are unlike normal programs in that they allow open-ended output. Normal programs are close-ended[amm]: given similar inputs they will behave similarl...
An insightful quote from Ilya Sutskever via an old Less Wrong post about why LLMs are so powerful: "to learn to predict text, is to learn to predict the causal processes of which the text is a shadow,"[amo]
LLMs are the input to the next big thing.[amq] LLMs are like electricity[ams]: a general technology that reconfigures the costs of key inputs, making wild...
My friend Anthea pointed out that my assertion that LLMs capture "all of society" is wrong. LLMs give a slice of the content represented on the internet, which has a strong western, English bias. She imagin...
...gy it might mean a new disruptive technology, like electricity, jet engines, or LLMs. The new disruptive technology reconfigures the "fitness landscape" of viable ideas, by radically changing a key dimension of cost. With all of the n...
Last week I framed LLMs as dowsing rods. The more I think about it, the more I like that frame. A dowsing rod is a fuzzy kind of imprecise 'magic' that you should hold light...
When extracting information from LLMs, we're like cavemen poking them in the dark. LLMs encode vastly more information than we know how to retrieve. We're in the very early stages of figu...
An interesting use case for LLMs: on-demand cozy schlock novels. For example, fan fiction or formulaic romance novels. These novels already aren't great literature, they're formulaic...
...uman in the loop, but humans are expensive and get bored.[anu][anv] Now we have LLMs to do some of the squishy, high-context things that can float around the problem domain. But that means that if you've iterated to find something you...