@runes_leo: At Sequoia Ascent on 4/30, Karpathy compressed this year’s most valuable explanation of AI into three core arguments. You’ll see AI differently after reading this. 1. AI Isn’t Just “Faster,” It’s a New Paradigm For the past two years, the narrative has been that AI speeds things up. Karpathy says this is a misunderstanding...

On April 30 at Sequoia Ascent, Karpathy compressed this year’s most useful explanation of AI into three core arguments. You'll see AI in a completely different light after reading this. 1. AI Isn’t Just “Faster,” It’s a New Paradigm For the past two years, everyone’s been saying AI makes things faster. Karpathy says this is a misread. Here are three examples of AI redefining tasks: - menugen: Image in, image out, zero traditional code—the entire app is consumed by the LLM - .md skills: Instead of writing `.sh` scripts to install software, you write a short prompt in Chinese or English and let the LLM parse your environment to handle the installation - LLM knowledge bases: Something traditional code can't do—turning arbitrary unstructured text into computable knowledge The first category is "code reduction." The second is "using natural language as code." The third is "capabilities traditional code simply didn't have." 2. The Jagged Edge — Why AI Is Both Omniscient and Foolish The core argument. How can the same AI refactor 100k lines of code but also suggest you walk to get your car washed? It’s not the model glitching out. In Karpathy’s own words: "You're either on the rails of the RL circuits and flying, or off-roading in the jungle with a machete." Either you're soaring on well-trained RL tracks, or you're hacking through the jungle with a machete. Two factors determine which tasks make it into the training distribution: verifiability (can results be objectively checked?) + economics (is there enough market incentive for frontier labs to invest heavily in RL). Math competitions / coding / theorem proving: High verifiability + High TAM → Makes the cut → You're flying when you use it Everyday advice / niche literature & linguistics / long-tail tasks: Low TAM → Missed RL → You're swinging that machete in the jungle It’s not a linear narrative of "AI just keeps getting stronger." It’s a jagged boundary, and you must know which side of it you're standing on. 3. Agent-Native Economy Final argument: Future software breaks down into sensor (input) + actuator (execution) + logic (reasoning). The logic layer runs entirely on LLMs, while sensors and actuators use traditional code as coprocessors. Implication: Structuring information for maximum LLM readability becomes the core constraint for future software design. --- These three arguments form a cohesive framework: The new paradigm shows you what AI can do that it previously couldn't. The jagged edge helps you identify where its limits lie. The agent-native approach tells you how to wrap the remaining tasks for AI. It’s not "AI just keeps getting better." It’s "knowing which tasks fall inside the track and which fall in the jungle."