Guillermo Flor@guilleflorvs
🚨 Last week was Sequoia AI Ascent, the one AI event of the year founders can't miss.
@karpathy opened the event with where AI is going.
Here's what you need to know:
His core idea:
- Software is changing again.
- Software 1.0 was code.
- Software 2.0 was neural networks.
Software 3.0 is prompts, context, tools, memory, agents, and verification.
The full breakdown:
1/12. Karpathy has never felt more behind as a programmer
Not because he forgot how to code.
Because coding agents crossed a threshold.
2/12. Prompts are becoming programs
The “program” is no longer just code.
It is context + model + tools + memory + feedback loop.
3/12. Software will become agent-native
Most software is built for humans clicking buttons and reading docs.
Agents need clear instructions, machine-readable docs, explicit permissions, and APIs designed for execution.
4/12. Many AI apps are temporary wrappers
Karpathy’s MenuGen example made this obvious.
Instead of building an app with OCR, extraction, image generation, and UI, you can just give a menu photo to Gemini and ask it to overlay food images.
5/12. The best products will come from impossible workflows
The real opportunity is not only automating old workflows.
It is creating things that could not exist before.
Like turning messy documents into a wiki, memo, market map, operating manual, or strategic brief.
6/12. AI automates what can be verified
Traditional software automates what you can specify.
AI automates what you can verify.
That’s why AI is moving fastest in code, math, tests, security, data workflows, and structured processes.
7/12. Founders should look for verification loops
The opportunity is not “another wrapper.”
It is finding valuable workflows where outputs can be checked, improved, and learned from.
Finance, tax, compliance, insurance, contracts, accounting, cybersecurity, logistics.
8/12. Vibe coding raises the floor
Anyone can now build apps, websites, tools, automations, and prototypes by describing what they want.
That is huge.
But it is only the first layer.
9/12. Agentic engineering raises the ceiling
The real skill is using agents while keeping quality high.
10/12. Hiring has to change
Small coding puzzles make less sense when the job becomes agentic engineering.
A better test:
Build a real project with agents. Make it secure. Make it work. Then let other agents try to break it.
11/12. Taste becomes more valuable
Agents can write code, generate drafts, and execute tasks.
But humans still decide what is worth building, what good looks like, what trade-offs matter, and what should not exist at all.
12/12. Understanding stays scarce
As intelligence gets cheaper, you need to understand more.
Because if you don’t understand the domain, you can’t direct the agent, inspect the output, or reject bad decisions.
Took me hours to do this, if it helped, give it a repost 🙏
