1,841 mentions · 834 chunks · 141 episodes
...o multiplied by infinity is infinity. "Prompt injection won't be a problem once LLMs get perfectly good at not being tricked" is absurd. Prompt injection comes from a coevolving adversary, not a static distribution of quality. That me...
...ed on the current task. I wonder what other innovations will come form treating LLMs as not append-only ping-pong chat.
I think the world is ready for the dishwasher and microwave oven of LLMs. Today we have the faux humanoid robot that talks and acts like a (weird) human). Chat is a new UI to put on top of all those tools but not be the an...
The fact that LLMs still require humans in the loop to get good recurrent results undermines the "AGI is imminent" perspective. Even if the LLM is right 95% of the time...
...nd a potentially dangerous one! The most natural and powerful interactions with LLMs will not be via a chatbot.
Will LLMs lead to compounding ambiguity in communication? There's value in shrouding a potentially controversial statement in load-bearing ambiguity to keep th...
...um. Also, it's becoming more clear that owning the context will matter, to make LLMs maximally useful for a given user. The cookie jar is an obvious, highly distilled source of context. But if you actually look inside your cookie jar ...
...e missing the GUI for LLM native software. Chatbots are the command line to get LLMs to do structured things. You need to know arcane knowledge for LLMs to do structured or complex things for you today. It just so happens that they're...
To be interesting it has to be opinionated. That's why LLMs by default aren't interesting. They give you the view from nowhere[bw]. They seek perfect "objectivity", a one-size-fits-none impossibility.
...n top. Weaving stories required human effort. Now with the infinite patience of LLMs it's less about the human doing it and more about the human giving permission for data to be built upon.
LLMs can do qualitative nuance at quantitative scale... but at 90% quality. So they can do a good enough job in a lot of cases, but if you feed their own ...
The best way to think about LLMs for writing code is like managing a team of interns. You could have done the work yourself, but they did it and you verified.
LLMs are infinitely patient, which means they can loop indefinitely. This can be a powerful force. An LLM looping and curating, with a stable goal to opti...
...they could fix a bug you do want. Instead of vendoring a dependency, he has the LLMs draw on their inherent knowledge from all of open source to distill a bespoke, fit for purpose library in place.
...tural schelling points requires cognitive labor that used to be scarce. But now LLMs might help sift through and find the obvious schelling points. Of course, this would change the internal politics meta-game…
...ently. It used to be a pain to think through it for each decision. But now with LLMs and their infinite patience, it's easier to have a fuzzy set of values and mission statement operationalized. Every decision can be run through an LL...
LLMs are losing the ability to simulate real people. LLMs are largely a warped mirror of all of the human input. It used to be possible to use LLMs as a k...
Apps are the wrong soil for the seeds of LLMs to blossom into their full potential.
Three pace layers of prototyping with LLMs: 1) The LLM does everything. Expensive, loosey-goosey, flexible 2) The LLM behavior is sublimated into a mechanistic harness that can be run inside o...
The printing press kicked off the enlightenment. LLMs have the potential to be the printing press for turing completeness.