x0elllse9

[x0elllse9 @elllse9]

https://t.co/8qHvb3Y9pZ

[x0elllse9 @elllse9]

@elonmusk Monday is going to be fire🔥🔥🔥

[x0elllse9 @elllse9]

Small addition since it was asked for:

One thing that didn't fit in the main post — what happens when the classifier is wrong.

Because it is wrong. not often, but it happens.

I keep a separate log of every escalate_to_human and kill_angle decision from the last 30 days. once a week I go through the kill_angle branch — not to revive them, but to understand why the classifier called them dead.

a couple times the novelty threshold was set too high and it was killing normal, "quiet" angles — not viral, but factually solid and worth publishing.

I fixed the threshold itself, not the classifier's prompt — cheaper and more predictable than rewriting the instructions and hoping the model "gets it right" next time.

second thing — hold_for_fact_check shouldn't be a dead end. if the fact checks out, the branch goes back into the graph from where it stopped, not from the start. the draft that was already written before the check doesn't get rewritten — only its status changes.

that's the actual difference between "rerun everything" and "rerun what broke."

the router isn't perfect on day one. it gets sharper because I look at its decisions, not because I bolt on more instructions.

next step — same five actions in the decision loop, but with weights that adjust week to week instead of being hardcoded.

[x0elllse9 @elllse9]

https://t.co/DwOnddlKEP

[x0elllse9 @elllse9]

https://t.co/DwOnddlKEP

[x0elllse9 @elllse9]

@sylithon FB

[x0elllse9 @elllse9]

@Airhunt109r Hi

[x0elllse9 @elllse9]

GPT-6 ASTRA MAKES ONE AGENT MISTAKE

IMPOSSIBLE TO IGNORE a 1M+ context window sounds massive, but long context alone does not make an agent reliable. what matters is how planning, action, checking, and memory are separated.

plan → tool use → verification → memory → output 5 stages keep one bad decision from silently reaching production, while useful results can be stored for future runs. long-running work starts looking completely different.

instead of one giant prompt carrying everything, each layer gets a narrow job and a clear failure point. better agents are starting to look less like smarter chatbots and more like small operating systems built around a model.

[x0elllse9 @elllse9]

@CryptoAvenue0 ACTIVE

[x0elllse9 @elllse9]

@sylithon Done

[x0elllse9 @elllse9]

The progress in AI over the past four years is staggering

It's already hard to tell AI from reality
What will it be like in 1–2 years?

What do you think this will lead to?

[x0elllse9 @elllse9]

@CryptoAvenue0 YES FB

[x0elllse9 @elllse9]

@iamelvisfx IFB

[x0elllse9 @elllse9]

@SFZOGEX69 Lets go!!

[x0elllse9 @elllse9]

@sylithon Yes

[x0elllse9 @elllse9]

@mrtbrgl Ifb permanentny

[x0elllse9 @elllse9]

@dtitser Fb

[x0elllse9 @elllse9]

@sylithon Yes

[x0elllse9 @elllse9]

@X_supporttop1 Yes

[x0elllse9 @elllse9]

@Chisom95109689 Grab

[x0elllse9 @elllse9]

@crypto_mooncity Grab