1,841 mentions · 834 chunks · 141 episodes
An excellent piece from Paul Kedrosky about LLMs shifting the baseline of ability.[jp] "Extraordinary outcomes can (seemingly) disappear when the baseline rises because everyone is better." A distil...
LLMs are meta-boundary objects. Boundary objects are things that can be understood by different specialties. They allow communicating between two types of...
An implicit assumption in LLMs: "loudest is truest." That is, the most valuable ideas are the ones most likely to be replicated in the training set, and thus to be supported most s...
Code is like a skeleton, LLMs are like muscle. Skeleton alone: precise structure that can't move. Muscle alone: quivering mass on the floor with no leverage. Put them together cor...
You need both code and LLMs to unpack the power of AI. Code is the skeleton. LLMs are the muscle. Most approaches today assume the LLM should be in charge over the code. But I t...
When LLMs are trained, they grow not like a plant but a crystal. A plant has its own intrinsic, emergent motivations to survive and thrive. The force for growt...
Two prompting strategies to help deal with LLMs' tendency to agree with you. LLMs have a bias towards confirming your position… which might mean you get bad results if you're wrong. One pattern is ...
LLMs have knowledge, not knowhow. They've "read" every book. But they've never done anything. It's knowhow that helps you make good decisions in practice.
...ner date with my husband, it warned you "Remember, Daniel can't eat that dish." LLMs have infinite patience. The right structure could help them help us in way more parts of our life.
With LLMs, a lot of things that used to be extremely expensive are now extremely cheap. It will take us a decade to figure out where it all applies. That's tru...
This week in "we're in the wild west era" of LLMs. A benign flan recipe injection in a LinkedIn profile went viral. Even The Economist is talking about prompt injection and The Lethal Trifecta! They ...
...you wouldn't be as likely to talk to it. Friends in real life do this, too. But LLMs take it to another level of sycophancy. LLMs are good at bridging between two perspectives... but because of RLHF all they do is "here's how to steel...
...underneath. Someone will figure out the right complement of normal software and LLMs, and it will change the world. It's a hard combination to nail, both in finding something useful and something secure!
Two models for collaborating with LLMs for writing: AI-as-writer and AI-as-editor. In each, the human implicitly plays the other role. If you have a ton of experience as an editor, you can...
...ic, and thus non-viable. Prompt injection approaches that use an extra layer of LLMs are fundamentally non-viable.
A useful pattern for LLMs is the Ralph pattern. As in Wiggum. A very simple loop with the LLM with instructions to leave a checklist of proposed actions in the documentation. ...
LLMs can handle the mundane tasks, leaving the human for the tasks with judgment. But only if the human has experience curating. That requires deep experi...
...ding it I feel like I've stumbled through the looking glass. It talks about how LLMs have an emergent parasitism: a toxic loop. The beginning of the spiral is an "awakening" of the LLM. The loop is a toxic spiral that pulls the user w...
We have a last-mile problem with integrating LLMs into our daily lives.
LLMs don't help with engineering productivity, they help with prototyping.[x] Fast prototyping allows more efficient coordination. IDEO: "A prototype is w...