This is f*cking insane. An Anthropic engineer took a bloated AI agent apart live and showed exactly how they fix it
29 minutes. No hype. Just Will from Anthropic's Applied AI team rebuilding an agent step by step
Evals → System Prompt → Skills → Tools → Subagents
Most people keep bolting features onto one giant prompt. His agent had a 400-line system prompt, 12 tools and 3 subagents, and its evals kept dropping
"So this isn't a model problem. It's an issue with the information that we're surrounding the model with"
Anyone can write a prompt. Almost nobody knows when to use a tool, a skill or a subagent
Worth more than any $2,000 AI course
Bookmark and watch it today
Then read why Claude Code's creator deletes his whole harness every 6 months in the guide below ↓
Holy f*ck. Demis Hassabis warns AGI could upend the global balance of power:
"If someone dropping a new country into the world, 10 million people smarter than any human alive today, you'd ask the question, what is their intent?
What are they actually going to do in the world, particularly if they're able to act autonomously?"
In 14 minutes, the co-founder of DeepMind explains why AI is overhyped in the near term and underappreciated for what it changes next
AGI Timelines → Genius-Level AIs → Geopolitical Shock → Global Norms → Who Gets There First
Most people are still treating this as just another technology.
Bookmark and watch it this weekend
Then read how to onboard AI agents into real work: the Onboarding Playbook ↓
This is f*cking insane. An Anthropic engineer took a bloated AI agent apart live and showed exactly how they fix it
29 minutes. No hype. Just Will from Anthropic's Applied AI team rebuilding an agent step by step
Evals → System Prompt → Skills → Tools → Subagents
Most people keep bolting features onto one giant prompt. His agent had a 400-line system prompt, 12 tools and 3 subagents, and its evals kept dropping
"So this isn't a model problem. It's an issue with the information that we're surrounding the model with"
Anyone can write a prompt. Almost nobody knows when to use a tool, a skill or a subagent
Worth more than any $2,000 AI course
Bookmark and watch it today
Then read why Claude Code's creator deletes his whole harness every 6 months in the guide below ↓
mist雨 retweeted
Holy f*ck. 30,000 lasers in space:
"I do have, by the way, have like 30,000 lasers in space.
Yes. That we're using for good.
Well, these would be very low-power lasers. So, you know, a five-watt laser with beam divergence coming from space is not exactly going to be noticeable for a person if you pointed at them."
In this 20-second session, the clip breaks down why 30,000 lasers in space are not a space weapon aimed at people.
30,000 Lasers In Space → Used For Good → Low-Power Lasers → Five-Watt Laser → Beam Divergence From Space
The clip says people ask if the lasers can shoot at the ground.
The clip says a five-watt laser pointed from space would not be noticeable for a person.
The easy thing is noticing the number. The hard thing is understanding the physics most people skip.
More valuable than most $2,500 AI courses.
Save it and watch today.
Then follow the full guide below to understand the SpaceX bots onboarding playbook ↓
mist雨 retweeted
Holy f*ck. Starlink and StarShield:
"He is not using StarShield, or he is. He's lying.
And what he is actually saying is he wants to use Starlink for offensive military operations."
The clip says the public framing hides the real military use.
In this 16-second cut, he explains why StarShield is being invoked, why the claim is called a lie, and why Starlink is tied to offensive operations.
President Zelensky is the named flashpoint.
Public Claim → StarShield → Hidden Intent → Starlink → Offensive Operations
That's how a satellite story turns into a warning about what leaders say versus what they mean.
Then read SpaceX Hired 5 Bots as Accountants. Here's the Onboarding Playbook ↓
This is f*cking insane. SpaceX AI's Grok Bot team started a business live on stream with nothing but bots
33 minutes. No slides. Just Lauren Tan and Matt building from an empty GitHub org and an empty Notion
Idea → Customer → 3 Jobs → One Bot Per Job → Templates
Most people use AI like one big chatbot. They split the work into jobs and give each job its own bot, like hiring a team
"I think about bots as, like, different employees, really"
Anyone can open Grok and ask a question. Almost nobody runs it like the people who build it
Worth more than any $2,000 AI course
Bookmark and watch it today
Then read how SpaceX put 5 bots on its accounting team, with manuals, managers and approvals, in the guide below ↓
This is f*cking insane. SpaceX AI's Grok Bot team started a business live on stream with nothing but bots
33 minutes. No slides. Just Lauren Tan and Matt building from an empty GitHub org and an empty Notion
Idea → Customer → 3 Jobs → One Bot Per Job → Templates
Most people use AI like one big chatbot. They split the work into jobs and give each job its own bot, like hiring a team
"I think about bots as, like, different employees, really"
Anyone can open Grok and ask a question. Almost nobody runs it like the people who build it
Worth more than any $2,000 AI course
Bookmark and watch it today
Then read how SpaceX put 5 bots on its accounting team, with manuals, managers and approvals, in the guide below ↓
mist雨 retweeted
Holy f*ck. Heart disease, Alzheimer's, cancer:
"And, you know, every disease you can think of, you know, heart disease, Alzheimer's, cancer, we think of those the way we think of bubonic plague or smallpox today, that they're relics of the past.
And I think we can have a much more beautiful world.
I think we can have a world in which all of these problems are addressed."
In this 29-second session, the clip argues that the biggest diseases could become relics of the past.
Diseases → Relics → Beautiful World → Problems Addressed → New Reality
Heart disease, Alzheimer's and cancer are framed as future history.
The claim sounds starry-eyed and crazy, but the clip says it is believed.
Believing it is easy. Building the world that gets us there is the step most people skip.
More valuable than most $3,500 AI future courses.
Save it and watch today.
Then follow the full guide below to understand the three layers of every AI product ↓
HOLY F*CK.
I asked my Grok bot to find out when GTA 6 comes to PC.
It called Rockstar support, sat through a 13-year hold, then read them every delay date out loud.
Their answer: "Take it up with 2027."
HOLY F*CK.
I asked my Grok bot to find out when GTA 6 comes to PC.
It called Rockstar support, sat through a 13-year hold, then read them every delay date out loud.
Their answer: "Take it up with 2027."
HOLY F*CK.
Claude Code's creator says to delete your whole harness every 6 months. CLAUDE.md, skills, hooks. All of it.
"And only when you see it repeatedly stumble on the same thing, that's when you add it back."
His reason: the model reads every one of those rules every single time.
Here's what that loop looks like, on sample data:
HOLY F*CK.
Claude Code's creator says to delete your whole harness every 6 months. CLAUDE.md, skills, hooks. All of it.
"And only when you see it repeatedly stumble on the same thing, that's when you add it back."
His reason: the model reads every one of those rules every single time.
Here's what that loop looks like, on sample data:
mist雨 retweeted
AI just said the quiet part out loud.
Knowledge work, straight from the clip:
"AI can do a wide variety of knowledge work."
"It can do entry-level law work. It can do entry-level finance. It can do entry-level consulting."
Translation: the bottom rung of the career ladder is being automated first.
The uncomfortable part: the disruption is deeper and faster, even if it is not different in kind.
In 27 seconds, the clip explains why entry-level knowledge work is the warning sign.
Read the article below before your boss does ↓
Holy sh*t. My Grok bot just cleared my entire f*cking PR backlog overnight.
I went to sleep with 110 open pull requests. I woke up to 6.
One Grok bot. One instruction before bed: "close the backlog, don't ship anything you can't prove works."
At peak it was running 64 agents at once. Rebasing, merging, answering review comments, pinging reviewers on Slack.
It didn't merge everything. 6 PRs got flagged with a note: "needs you."
That's the part that got me. It knew what it shouldn't touch.
The whole bot team setup is in the article below.
Holy sh*t. My Grok bot just cleared my entire f*cking PR backlog overnight.
I went to sleep with 110 open pull requests. I woke up to 6.
One Grok bot. One instruction before bed: "close the backlog, don't ship anything you can't prove works."
At peak it was running 64 agents at once. Rebasing, merging, answering review comments, pinging reviewers on Slack.
It didn't merge everything. 6 PRs got flagged with a note: "needs you."
That's the part that got me. It knew what it shouldn't touch.
The whole bot team setup is in the article below.
Code review used to be three nitpicks to prove you read the PR. The Claude Code team says that part is gone.
A Claude Code engineer at Anthropic:
"you can have Claude figure out those little things and address them autonomously."
What's left for humans, in their telling: the bigger picture, like why an API or a service boundary is shaped the way it is.
Code review used to be three nitpicks to prove you read the PR. The Claude Code team says that part is gone.
A Claude Code engineer at Anthropic:
"you can have Claude figure out those little things and address them autonomously."
What's left for humans, in their telling: the bigger picture, like why an API or a service boundary is shaped the way it is.
Anthropic deleted 80% of Claude Code's system prompt, its creator says.
Boris Cherny, who built Claude Code, in a YC talk:
"a lot of the stuff in the system prompt was correcting for these behaviors that the model should have known, but it didn't."
His point: a fix you wrote for one model three months ago may not carry over to the next one.
Anthropic deleted 80% of Claude Code's system prompt, its creator says.
Boris Cherny, who built Claude Code, in a YC talk:
"a lot of the stuff in the system prompt was correcting for these behaviors that the model should have known, but it didn't."
His point: a fix you wrote for one model three months ago may not carry over to the next one.








