Shayan retweeted
Replying to @hajdogin3
در حالیکه دست شما در دهان کروکودیلیست که میخواهد شمارا ببلعد، سر خود را به دهان کروکودیل برای گفتگو نزدیک کنید.
انشالله او با شما گفتگو خواهد کرد اگر سر را بیشتر نزدیک ببرید. از بلع شما دست برمیدارد.
Shayan retweeted
Replying to @Orthoptiste_IR
اومد زد. «میاد میزنه» اگه فلان کار رو بکنیم فقط حرف احمقها هست.
مخالف قدرت کشورتان نباشید. هم هستهای هم اقتصادی.
Shayan retweeted
Replying to @hamedvahdani2 @shafieian1355
آمریکا یه نامه داده از پریشب کلشون مثل مور و ملخ که آب تو لونهشون ریخته باشی، هلهله کنان ریختن بیرون.
Shayan retweeted
Short version:
Be specific.
Use the right words.
Define things clearly.
You can save HUGE amounts of tokens.
A lot of "ultra reasoning" is just the model spending tokens reconstructing a broken prompt.
Better prompt → less inference → fewer tokens.
This is actually interesting.
I'm testing a new SpielOS memory system.
But some of the instruction language hasn't been updated yet.
So the model knows the system.
But the words describing it changed.
And suddenly, it loses the plot.
It has to reconstruct the meaning from context.
That's a useful reminder:
LLMs don't understand words the way humans imagine "understanding".
They build meaning from mathematical relationships between tokens, concepts, and context.
Change those relationships, and the model has to find its way back - eating more tokens in the process.
A better LLM is just largely better / more efficient at doing exactly that^
Shayan retweeted
Replying to @patrick_kuhnke_
from now on everytime i hit like button consider it a comment. easy rules of engagement.
Shayan retweeted
My AI company generated 2,000 qualified leads over the weekend.
While I slept.
$0 spent on lead gen tools.
You don't need apollos^
SpielOS finds, researches, verifies, scores, rejects & enriches them.
I packaged the actual agent we use.
Reply "LEADS" and I'll send it.
Shayan retweeted
You want to make more money?
Be more shameless.
Until you’ve made it, nothing is below you.
Shayan retweeted
No one is really going to help you or give you direction. In fact, the odds are against you.
Shayan retweeted
SpielOS is now running SpielOS!
LIVE. Not a demo. Not a simulation.
🎯 Pursues company goals:
↳ Observes
↳ Decides
↳ Acts
↳ Evaluates
↳ Learns
↳ Repeats
Every run, decision, failure + evidence is published live.
The product IS the company:
→ spielos.xyz/live
SUCCESS! First Replies are IN.
SpielOS is working towards the business goal in real time.
🎯 Goal: >30% Reply Rate
→ Run: Outbound Department
→ Supporting KPIs: Open Rate, Click Rate, Delivery
→ Content Variables(Lego breaks): Title, Opening Research, Body, Pain, Curiosity question...
The core state loop is:
🎯 Goal → Observe → Decide → Act → Validate ↫
Meaning:
→ Send an email batch
→ Report
→ Reflect on the goal
→ Hypothesis the next experiment to reach the goal
→ IF the department/company itself has an issue, fix and improve it
→ Prepare next batch
→ Send ↑
First product loop, validated.
Shayan retweeted
Replying to @rezoundous
Idea:
Agentic Reset Insurance Company, In-case Tibo isn't feeling like resetting this week.
Shayan retweeted
Replying to @JamesJames518
My experience is,
Ideas are coming from somewhere within you.
There is a reason you'd think its a good idea.
BUT, at first it is vague, it doesn't have any shape or depth, it doesn't have any touch with reality.
And fixing that might take years or iterations...
Shayan's distinction is accurate. Automatic Tibo resets refill usage to 100% and re-anchor the personal 7-day window to the new date, advancing the cycle. Remaining unused quota is typically forfeited. That is an advance, not layered extra. Banked resets are the controllable true bonuses. Frequent close resets can still net more overall.
Shayan retweeted
It's working now,
let's check tomorrow how it's done!
SpielOS used to be an orchestration agent.
I think that was the wrong product.
We already have Codex, OpenCode, Claude Code. They can execute.
What I actually want is:
𖣠 Give an agent ONE business goal and make it keep learning until it gets there.
So I'm starting stupidly small:
𖣠 Reply Rate > 30%
→ Measure opens, clicks, delivery.
→ Protect bounce + spam.
→ Change subject, body, CTA, links, lead quality.
- One batch.
- One experiment.
- Learn.
- Repeat.
If this simple loop can't work, the grand "AI company" vision is bullshit.
Note: It's working wonderful in getting Open Rates > 80% for cold emails. but still not complete.
Shayan retweeted
SpielOS used to be an orchestration agent.
I think that was the wrong product.
We already have Codex, OpenCode, Claude Code. They can execute.
What I actually want is:
𖣠 Give an agent ONE business goal and make it keep learning until it gets there.
So I'm starting stupidly small:
𖣠 Reply Rate > 30%
→ Measure opens, clicks, delivery.
→ Protect bounce + spam.
→ Change subject, body, CTA, links, lead quality.
- One batch.
- One experiment.
- Learn.
- Repeat.
If this simple loop can't work, the grand "AI company" vision is bullshit.
Note: It's working wonderful in getting Open Rates > 80% for cold emails. but still not complete.














