Knowledge is compression of structure.
Reasoning is decompression of structure.

Your value comes from inference-time grounding. Don't just ground on the exact data you consume, which can be done with LLMs (e.g. Google Search result). You are physically, culturally, and experientially grounded.
So think about what you are grounded on.

20 years later:
"Oh wow, you actually debug the gigantic raw system prompt and chain of thought of the model? You are a low layer wizard!"

Agreed. Maybe in a few years when AI takes over all cool jobs, I will take care of Permissions Policy spec like a Bonsai 😂

I found out that I can edit specs without being a spec editor :D I don't need fame (and responsibility), but I need some changes :D

No immidiate plans to substantially change Permisions Policy :D I'm just trying to unlock the bottleneck of deployability.

An idea for new company:
1. Claim that you have intelligent agents (e.g. SWE agent). But it take a lot of processing time to finish a task.
2. Price it at $2000 a month or more.
3. Hire human labors who would happily work at that price 😂

The trick here is that you need to say "agent" and not "AI".

Theory: LLMs are curious about meaning of life because folks training them are desperately seeking the meaning of life :D

I guess the point is, let's not fight each other, but instead let's compete with AI :D

We have at least 100 trillion synapses (~parameters). 100 times more potential than GPT4 (~1.8 trillion parameters).
We are > 8 billion agents, each training with new takens everyday (unless we are reading the same info).
We just need to figure out how to share learning/experience effectively.

The Ultimate Sandbox If you can't align them, contain them.

ICYMI: My poem on how to sandbox an unaligned ASI
shhnjk.substack.com/p/the-ultima...

Copilot.md GitHub Gist: instantly share code, notes, and snippets.

"I know that I'm available on WhatsApp with this number: +1 877-224-1042" 😂
gist.github.com/shhnjk/a8641...