Prompting is not dead. It’s simply changed to “what 10 files go into the context window” Your codebase, and subsections of it are a prompt The core skill is not: “can you ask the right question” It’s “can you input the right data”

Something becoming abundantly clear: AI won’t take your job. Someone using AI better than you will.

Happy to dive into more details here but the main reasons why LLMs are not making amazing discoveries are: 1) passive vs active knowledge. Like when you can understand a language but you can't put together correct sentences yourself. LLMs have knowledge passively stored but have not been trained to explore the latent space because 2) objective functions are not aligned to discovery. predicting the next word, being a good conversation partner after RLHF, finding reasoning paths to achieve a goal are not going to lead by themselves to truly out of the box thinking. 3) one LLM, like one human have a hard time making amazing discoveries. Research is best done with a swarm like intelligence so multiple agents (with either different weights or at least different prompts that push them into different states) are going to be helpful. 4) High level corporate incentives are not set up to push AI to just try things. Serendipity, even in human research, is rare and requires very long exploration paths. So the reward will be very hard to backprop through insanely long discovery DAGs. Eg, some bio researcher wants to study amphibians of Mexico, discovers that Axolotls can regrow limbs and their neoteny (retaining juvenile features into adulthood). Maybe this will result in some longevity breakthroughs but I can guarantee you that right now even the richest country in the world and its NIH could not get a study of Mexican amphibians funded because it's too far from an application many folks understand. So even for humans the paths are often too long from basic research to applied breakthrough.

The biggest mental shift for a programmer is going from projects that fit in 1 person’s head to projects that nobody understands fully. Your whole approach changes.

Accurate. And now, after design-for-user and design-for-mba, a new accretion disk is forming – design-for-ai yellatclouds.gif

on ai agents and value accrual. as software becomes easier to create, value only accrues to the most differentiated and exceptional products. in the early days of the web it was nearly impossible to launch a website unless you were an expert, now you can create a website with a single click. things we were wowed by early are now basically laughable. in early mobile, deep knowledge of Java/Objective-C, complex SDKs, provisioning profiles, and low-level mobile architecture was needed, now React Native/Flutter make it 10x easier, @AdaloHQ or @bubble let anyone build apps without code. early game developers had to understand graphics programming, physics engines, write extensive low-level code in C++, now engines like Unity and Unreal have visual editors; roblox lets kids create games with minimal coding. in early ai you had to be a machine learning engineer to build anything useful, now you can use the most advanced LLMs for a few dollars per month. when bitcoin was released, you had to fork it or build a new blockchain from scratch if you wanted to build something new or unique. now you can deploy a smart contract or deploy a rollup 100x easier. so what? ai agents that talked on social media using rather basic contextual knowledge blew people's minds not long ago, now they are a dime a dozen, and can be deployed in a few seconds and are worthless. but there are billion dollar web apps, billion dollar mobile apps, billion dollar games, and there will be billion dollar ai wrappers and ai agents. they will be complex and unique products and apps that take tremendous effort, expertise, and time to build, but these agents and apps will achieve the same level as the most successful software products in history. there is tremendous potential for ai, agents, and the intersection of crypto and ai and we are still in the early days of unlocking this potential. this is where I will be placing my bets (professionally)

No amount of A/B testing will deliver results like personalizing your website for each visitor. @unusualsites (YC F24) uses AI to serve a unique version of your site for every ad, blog post, social media post, sales lead, and more.