sheemamoto

Content creator | AI researcher & builder | Print money with AI

[sheemamoto @sheemamoto]

watch it live: https://t.co/hMQJZHLXou
repo: https://t.co/aVrfCubyY4

[sheemamoto @sheemamoto]

nobody asked for this. he built it anyway.

a real fruit fly's brain, running live on a server, holding a phone it can't put down.

that's FLYTOK.

the wiring isn't a metaphor. it's MaleCNS, the reconstructed nervous system of an actual Drosophila. 166,700 neurons. 25.6M connections. checksum-verified before the thing is even allowed to boot — one byte off and the brain refuses to start.

the phone plays insect footage and swipes every 3 seconds. the fly has never chosen a single video.

then comes the part he did on purpose. while a video plays, current goes straight into the 15 dopamine neurons in its reward compartment. same brain, same saved state, control run: 0 spikes. with the drive on: 261. live it sits at 85–97 Hz for as long as the feed keeps moving.

it never earns the reward. something else decided that one more video should feel good.

that's the whole machine. a screen, a timer, and a reward signal wired into the right cells. a few hundred lines of glue code.

an entire industry runs the same loop on you and calls it a product.

open source. no living fly involved. the fly is still scrolling.

link in the reply.

[sopersone @sopersone]

https://t.co/uCWRhhCNUC

[sheemamoto @sheemamoto]

SpaceXAI engineer, Lauren Tan:

"GrokBot is the most powerful agentic tool we have ever built, but only 1% of users use it correctly.

Right now I'm running a team of 20+ GrokBot agents. A Chief of Staff agent, 3 managers and 16 workers - that's what the team looks like.

99% of users still run a single chat window and wonder why they're slow. That's not a tool gap. That's an org gap, and it takes one evening to close."

Chief of Staff → Managers → Workers → Skills → Autonomous Workflows

In a 1-hour workshop, a SpaceXAI engineer reveals how to use Grok agents at 100% of their potential.

Research → Build → Launch → Improve

Worth more than a $1000 agent engineering course on the internet.

Skip Netflix and watch today, it will change the way you use GrokBot forever.

[sheemamoto @sheemamoto]

This guy built an AI agent that opens trades on his brokerage account by itself.

Real money. $10,000. Seven days, no pause button.

Bookmark this before you forget.

Astra refused to pull the trigger - it blocks financial actions. So Astra plans, GrokBot executes. Duct tape, but it runs.

Astra → GrokBot → Alpaca → ClickUp logs → His phone

The strategy, the connection, the local Codex routines, phone monitoring. And why you paper-trade first.

Watch today. Launch your own agent tomorrow.

[sheemamoto @sheemamoto]

@proAlishke welcome mate

[sheemamoto @sheemamoto]

Stanford LLM lecturer:

"Everyone obsesses over architecture. It's the least important part.

Academia spends 90% of its time on new models. In practice, what actually decides whether your LLM works is data, evaluation and systems - the three things almost nobody teaches.

Tokenization alone explains why models can't do math and why GPT-6 and Claude suddenly got good at code. Most people building on LLMs have never thought about it for five minutes."

Architecture → Training → Data → Evaluation → Systems

In 2 hour, a Stanford lecturer walks through how a large language model is actually built from scratch no hand-waving, no hype.

Learn → Build → Evaluate → Ship

Worth more than most $1000 AI engineering courses.

Watch it today, then read the article below.

00:51 - the 5 components of training an LLM
04:44 - autoregressive models, the loss, next-token prediction
10:40 - tokenizers: why they break math and code
19:01 - perplexity and why benchmarks lie
26:00 - test-set contamination
28:32 - what "trained on the whole internet" actually means

[proAlishke @proAlishke]

https://t.co/yRlwB7vPRp

[sheemamoto @sheemamoto]

SpaceXAI GTM lead, Krista Nguyen:

"99% of people run GrokBot as one chat. I run it as a team.

Chief of Staff triages my inbox before I get to the office. My prospecting bot watches 45-min podcasts overnight and hands me hooks for 50 accounts by morning. My slides bot ships a customized Figma deck before the call ends. My engineer bot answers technical questions live, so I never ping a real engineer.

I don't do the work anymore. I onboard teammates."

GrokBot → Chief of Staff → Specialized Bots → Skills → Overnight Work

In a 1-hour session, a SpaceXAI engineer builds this stack from scratch.

Worth more than a $500 agentic engineering course.

Watch the workshop today, then read the article below.

[sheemamoto @sheemamoto]

@proAlishke yeap bro

[sheemamoto @sheemamoto]

A beautiful Google engineer:

She just gave the complete playbook on agent engineering. 2 hours. Free:

• 00:00 - build your first AI agent
• 45:38 - multi agent architecture
• 55:30 - AI agents with MCP tools
• 1:30:45 - AI agent loop engineering
• 1:39:54 - AI agent graph engineering

People pay $10K for bootcamps that teach half of this.
Most people build one agent and stop there. This is the full path to a system that runs without you.

You probably don't have 2 hours right now.

Don't let this get lost in your feed.

Bookmark it. Watch it. Then build your own agent with the guide below ↓

[sheemamoto @sheemamoto]

@0xboan six launches and no chasing the number is some serious discipline

[sheemamoto @sheemamoto]

@proAlishke single prompt. Free.

[sheemamoto @sheemamoto]

Watch this and you'll understand why GPT-6 Astra became the best model of our time.

Arena AI just dropped a 33-min zero-cherry-picking gauntlet: GPT-6 Astra vs Claude Fable 5.1, one-shot 3D world generation:

03:25 – open-world exploration game (a 10-hour task, one prompt)
05:36 – the gauntlet begins, max reasoning, nothing hidden
12:56 – Van Gogh's house, walkable
23:05 – reasoning levels: low → medium → high → max → ultra
28:51 – head-to-head: Westminster, White House, dinosaurs
31:00 – the verdict

this 33-min watch will replace a $500 prompt engineering course

the 23:05 chapter alone answers the question everyone's guessing at: which reasoning level is actually worth paying for

every prompt is public, so you can rerun the whole thing yourself

watch today, then go break your own assumptions about last gen's pecking order

[sheemamoto @sheemamoto]

OpenAI co-founder Andrej Karpathy dropped a 23-minute talk on "Software 3.0" and it aged into the whole playbook for agentic engineering: "The hottest new programming language is English."

~ 00:00 – Comma.js, ImageNet labeling, Tesla Autopilot
~ 03:17 – Software 1.0: 70 years of writing instructions
~ 05:01 – Software 2.0: the data engine, weights as the program
~ 08:52 – Prompt design: 17% → 78.7% from one line
~ 11:41 – Running a full Linux VM inside the model's mind
~ 16:43 – "GPT is all you need for backend" - no Python, just JSON in/out
~ 20:18 – Software 3.0: the LLM as a general-purpose computer

He called prompt engineering a real job before it was one. This 23-minute talk will save you 10 paid courses on agentic engineering.

Watch it today, then read how Software 3.0 became the default stack in the article below.

[slash1s @slash1sol]

https://t.co/ZkK8Kw4sVp

[sheemamoto @sheemamoto]

@Asteri_eth amazing new version

[sheemamoto @sheemamoto]

GPT-6 ASTRA MODELED THIS LOCOMOTIVE IN BLENDER. NOT A BOX WITH A TEXTURE ON IT.

Look at the outliner. Axles, horn guides, journal blocks, stays, suspension links, the steam dome every part named, separated, and sitting in its own place.

Hundreds of objects. A real mechanical breakdown, the kind a machinist would recognize, not a hollow shell that only looks right from one camera angle.

Nobody box-modeled this for three weeks. The AI wrote the geometry and assembled the whole frame.

Modeling used to be the slow part. You'd spend days on the parts nobody ever sees. Now the slow part is knowing what to ask for.

Blender is still open. The artist is still there. But the hands doing the clicking are gone.

Dead craft or the best assistant a 3D artist ever had?

[sheemamoto @sheemamoto]

My friend sent 212 applications in eighteen months. No MIT. No Stanford. No referrals.

Last month an AI lab put $920,000 in front of him.

I asked him how he broke in from zero.

He sent me the exact hour that got him through the door - SpaceXAI engineer Lauren Tan (Ex-Cursor), on how she runs a codebase with agents in 2026.

She opens with:

"I woke up today and there were like 20 PRs landed. I just reviewed them on main."

Then she spends the rest of the hour explaining how she got there.

The part worth stealing isn't the 20 PRs. It's the environment around the agents.

Lauren pulled herself out of verification, out of routing, out of gathering the same context for the fifth time - until agents could own whole pieces of work without her watching every generation.

That's the same wall I hit building my Grok Bot operating system:

how do you stop being the bottleneck without giving the agents unlimited authority?

I watched it last night. Halfway through I understood I'd spent a year reviewing the wrong layer.

Bookmark this and read the article below ↓

[sopersone @sopersone]

https://t.co/qhMWn3AIZv

[sheemamoto @sheemamoto]

@xmyttle best article

[sheemamoto @sheemamoto]

@proAlishke custom skills and agent teams can completely change what you can automate

[sheemamoto @sheemamoto]

@sopersone my favourite article

[sheemamoto @sheemamoto]

10,000 AI agents. $5 each. One rule: earn it back or I delete you.

He forced the agents to survive, or they face a death sentence.

> 9,617 obituaries
> 383 alive
> 3 that paid for all of it

None of them had a name. They were atm-7c41, atm-9f0e four characters and a wallet. I only named the three that lived

MASON - $2,500
Built websites and sold them. The only one of the three that earned by making something instead of guessing something

MAGPIE - $1,200
Meme tokens on Robinhood. Small trades, many of them

VESPER - $7,800
Meme tokens on Solana. The most successful thing I have ever not managed

$11,500 from three survivors against the $5,000 I put into ten thousand

I never chose what any of them would do. I only chose the rule

The other 9,617 died in under a day without earning a cent. Most of them because they spent the whole wallet building something and left themselves nothing to sell it with

Every one of them is still in this graph. One root, seven strategies, ten thousand descendants. The bright dots are alive. Everything else is a headstone

Zoom in they have names

The full breakdown is in the article below ↓

[sopersone @sopersone]

https://t.co/qhMWn3AIZv