- LLMs build internal Knowledge Graphs in their network layers.
- LLMs shine in the kinds of situations where “good enough is good enough”.
- LLMs are compressed information stores.
- Classic ML systems where humans are designing how the information is organized (feature engineering, linking, graph building) scale poorly (the bitter lesson). LLMs are able to learn how to organize the information from the data itself.
- LLMs may not yet have human-level depth, but they already have vastly superhuman breadth.
- Learning to prompt is similar to learning to search in a search engine (you have to develop a sense of how and what to search for). A good example from OpenAI.
- LLMs have encyclopedic knowledge but suffer from hallucinations, jagged intelligence, and “amnesia” (no persistent memory).
- AI tools amplify existing expertise. The more skills and experience you have on a topic, the faster and better the results you can get from working with LLMs on that topic.
- LLMs are useful when exploiting the asymmetry between coming up with an answer and verifying the answer (similar to how a sudoku is difficult to solve, but it’s easy to verify that a solution is correct).
- LLMs are good at the things that computers are bad at, and bad at the things that computers are good at. Also good at things that don’t have wrong answers.
- The use of LLMs and agents can be successful on tasks that are; verifiable, brute forceable, automation-like, PoC friendly.
- Context is king. Managing the context window effectively is crucial for getting good results.
- LLMs amplify existing expertise rather than replacing it.
- AI speeds up generation, not necessarily throughput. Bottleneck are often upstream: unclear scope, missing context, legal/product/domain ambiguity.
- Typing faster was never the main software bottleneck. Software is mostly decisions, tradeoffs, and understanding.
- Use AI to reduce friction around the bottleneck, but first identify the real bottleneck
- “AIs” can be dangerous in under-specified environments (e.g: pausing games to last longer in the level) but those are the places where we will use them most. If something is well specified, there might be better solutions/optimizations (maths, code, …).
- When AI is “right”, people become less skilled. When AI is “wrong”, you need skilled people to catch and fix the errors.
- When the main purpose of writing is to demonstrate your thinking (building trust, applying for a job), don’t use LLM output. Use LLMs when you need to communicate info, or do admin stuff, where the person really just wants info and doesn’t need to be convinced “how you think”. LLMs are good at writing but bad at thinking.
- LLMs are helpful when you want the output/result and don’t need to do the work yourself (e.g: going to the gym doesn’t work if the weights are automatically lifted).
- Personal communication and writing as trust and self-expression. Rewriting with LLMs changes meaning, blurs authorship, erodes voice.
- What you send should take more effort to produce than it takes the recipient to read. Raw, unread, unrequested AI output (sloppypasta) pushes the cost of reading, verifying, and distilling onto everyone else.
- If you do share AI output, disclose it and ideally link the chat. The conversation is more reusable than the transcript, and others can fork your context instead of re-reading your output.
- If I can have a robot say something, so can you. It reads as inconsiderate to post un-digested AI output as though it’s your own writing.
- If you are requesting human attention, demonstrate human effort.
- For now, AI generated content should be consumed by you, not shared with others. This applies to code, bug reports, emails, and so on.
- Don’t outsource thinking. That means tasks that:
- Build complex tacit knowledge you’ll need for navigating the world in the future.
- Are an expression of care and presence for someone else.
- Are a valuable experience on its own.
- Are deceptive to fake.
- LLMs as “stateless functions”. Fixed weights, no updating. LLMs are in-context learners.
- LLMs are a lot better at fill in the blanks (the micro) than grand strategy (the macro). Generalists might increasingly outperform specialists.
- The skill you want to build is the ability to understand problems and have some concept of how to solve them.
- Knowing how the style is named makes prompting easier.
- Understand benchmarking and what it really represents.