1,749 mentions · 790 chunks · 131 episodes
LLMs have only been getting incrementally better but new things are also made discontinuously possible. That's due to threshold effects. Some use cases do...
...s: Permuted Congruential Generators, Curricula, and Interpretability. Turns out LLMs are very good at learning the patterns underlying any sequence, no matter how subtle.
...re the artifact is default-converging. Useful for human teams… and working with LLMs too!
...ed to Claude in a loop instead of python and bash scripts to do routine tasks." LLMs are insanely powerful and able to handle surprise in a way mechanistic software never could. But they are ludicrous overkill for most tasks. If you h...
LLMs are explosively valuable in the domain of code. Is it because code is the first domain to explode… or is there something about LLMs being a perfect f...
... a logarithmic curve to quality. What looks 80% done is 20% done. Building with LLMs allows getting superficial results extremely quickly. But it doesn't get that last 80% done unless you hound it to go into the details that you haven...
...re about to have a very hard time: those who are "below LLM replacement level." LLMs have already become much better engineers than the vast majority of us. The only way to stay above water is to learn new ways of working that take ad...
...but a bit passive aggressive. "Yes, of course, why are you asking, just do it." LLMs know the connotations of that word and respond appropriately.
The AI Peter Principle is even stronger than the original. LLMs allow you to execute beyond your ability. The result is you're almost certainly out over your skis. You could stop within your ability, but humans ar...
...rging requires a process that reduces incoherence over time, that roots it out. LLMs have infinite patience[c] but are also prone to introduce subtle mistakes.
LLMs have the advantage in any domain where patience is more valuable than taste.
LLMs are fundamentally naive. Prompt Injection is just the most extreme version of that. But LLMs do subtly dumb things all the time. The practical risk o...
LLMs are like oxidation is for life. Critical and yet not sufficient. The force needs to be harnessed in the right structure to unleash its power in coher...
...to-the-middle advice can only ever make you good, never great. McKinsey is like LLMs, in that they both give you advice that will help you race to the middle.
... world. Of course the best jobs are knowledge work, and always will be. But now LLMs can do cognitive labor at an unimaginable scale. …Maybe knowledge work won't be as common, or often will take a different form?
...itive labor, they just required humans to do a new kind of cognitive labor. But LLMs can now truly handle it. But only if they have the context of your life and you trust them to act as an extension of you.
Some terminology: A harness is mechanistic code that LLMs run within. An agent is a loop with an LLM and tool calls. An assistant is an agent with a memory.[b]
LLMs are really good at rebasing.[c] Rebasing normally is tedious and error-prone. A form of cognitive labor that dominates the development of software. B...
LLMs are easy to trick by putting words in their mouths. If you give an LLM a task like "Write a poem about Strawberries" and then prefill its answer with...
...ade any data. A much larger surface area! All data can do things now, thanks to LLMs. Powerful, but catastrophically dangerous in our current physics of trust.