This is f*cking insane. If you're a junior lawyer, consultant or analyst, the CEO of Anthropic has given you an expiration date.
1 to 5 years.
Dario Amodei, the man who left OpenAI to build Claude, says AI could wipe out half of all entry-level white-collar jobs.
Half.
Honestly? I think he's being polite.
Look at what he says AI already does: summarize the documents, analyze the sources, write the report, write the code.
That IS the junior job. All of it.
And the part he admits nobody is ready for:
"it's eerie the extent to which the broader public and politicians, legislators, I don't think, are fully aware of what's going on."
Chatbots → Agents → AI Employees → Entire Departments Run By AI
The juniors who make it won't be the smartest people in the room
They'll be the ones who know how to put agents to work
95% will watch AI eat their industry. The 5% who build with it will be the ones companies fight over
In this 47-minute talk, Dario explains who comes out the other side
Worth more than any $3,500 AI career course
Bookmark it. Send it to someone who just started their first job
Then read the article below ↓
a whole photo studio. one shader. zero libraries.
opus 5.5 wrote the ray tracer, rendered every frame, then reviewed its own frames and fixed 22 bugs.
9 prompts + their videos + the full kit in the article ↓
3
32
This is f*cking insane. Anthropic's own engineers barely prompt anymore.
95% of them run agents that get better on their own.
One of them, on what's happening inside Anthropic:
"95% of our engineers are using self-improving harnesses + Opus 5.5. Now everyone is building agentic systems around them."
The model stays the same. The harness around it is what keeps improving:
→ it runs the loop until the job is actually done
→ it keeps a log, so a crash doesn't kill the work
→ between sessions it reviews its own runs and writes the lessons into memory
Anthropic calls that last part "dreaming"
Raw prompt → Agent → Harness → Memory → Self-improving system
Give it 1 to 2 months. Raw prompting will look like writing code in Notepad
In this 35-minute lecture, an Anthropic engineer builds a self-improving harness from scratch with Opus 5.5
Worth more than a $1,500 agentic course
Bookmark it and watch it tonight
Then read the full article below ↓
a whole photo studio. one shader. zero libraries.
opus 5.5 wrote the ray tracer, rendered every frame, then reviewed its own frames and fixed 22 bugs.
9 prompts + their videos + the full kit in the article ↓
6
9
66
This is f*cking insane. The CEO building the AI says it could wipe out half of entry-level office jobs.
Anthropic CEO Dario Amodei's prediction:
→ 50% of entry-level white-collar jobs gone
→ unemployment up to 20%
→ in 1 to 5 years
His description of what AI can already do:
"as good as a smart college student"
Summarize a document. Analyze a pile of sources. Turn it into a report. Write code. That's the first job of half the graduates on Earth
Prompts → Agents → AI Coworkers → Entry-Level Work Done By AI
Farming to factories took generations. This time AI hits the first rung of every career at once
His advice for anyone starting out:
"Learn to understand where the technology is going. If you're not blindsided, you have a much better chance of adapting."
The people who start this year won't compete with AI. They'll be the ones running it
Worth more than a $10,000 year of college
Bookmark it and send it to someone graduating this year
Then read the article below ↓
a whole photo studio. one shader. zero libraries.
opus 5.5 wrote the ray tracer, rendered every frame, then reviewed its own frames and fixed 22 bugs.
9 prompts + their videos + the full kit in the article ↓
7
12
421
This is f*cking insane. One engineer at Anthropic runs tens of thousands of AI agents.
Not a team. One guy.
Boris Cherny, head of Claude Code:
"This morning I was managing maybe a few hundred"
"Some days it's … thousands, or tens of thousands."
The trick? He doesn't talk to them:
"It's actually another Claude that does the prompting."
One Agent → Subagents → Hundreds → Tens of thousands → Self-Improving System
He hasn't written a line of code by hand in 8 months
Claude Code itself is now 100% written by Claude Code
Most people still manage one agent in one chat window
He manages an org chart made of Claude
In this 36-minute talk, he shows how far it's gone inside Anthropic and where it goes next
Worth more than most $1,000 agent engineering courses
Bookmark and watch it today
Then read the article below↓
a whole photo studio. one shader. zero libraries.
opus 5.5 wrote the ray tracer, rendered every frame, then reviewed its own frames and fixed 22 bugs.
9 prompts + their videos + the full kit in the article ↓
6
18
1,285
This is f*cking insane. Anthropic's engineers stopped prompting their own model.
One of them explains why:
"You're not supposed to prompt Opus 5.5
You're supposed to build a system that prompts itself. At Anthropic, 90% of our engineers already have"
Prompt → Agent → Memory → Self-correction → Self-improving system
The prompt is one shot. The system is a loop:
it remembers what worked
it catches its own mistakes
it comes back sharper on the next run
Anthropic even has a name for the overnight part: "dreaming." Between sessions, the agent reviews its past runs and writes the lessons into memory
Most people are still rewriting the same prompt and hoping
In 30 minutes, she shows exactly how Anthropic builds agents that learn while nobody's watching
Better than any $1,000–$2,000 agent course I've seen
Bookmark it, watch it tonight, then read the article below ↓
a whole photo studio. one shader. zero libraries.
opus 5.5 wrote the ray tracer, rendered every frame, then reviewed its own frames and fixed 22 bugs.
9 prompts + their videos + the full kit in the article ↓
8
18
586
This is f*cking insane. Anthropic's agents now "dream" between tasks.
The engineers who built them just spent 90 minutes showing exactly how it works.
Priyanka Phatak leads the Anthropic team behind Claude Managed Agents. Her one rule:
"harnesses encode assumptions, and assumptions can go stale"
The fix: split the brain from the hands. Median time to first token dropped ~60%. The slowest 5% of requests got 90%+ faster
0:00 – why the brain and the hands have to live apart
7:40 – the architecture behind Claude's hosted agents
23:35 – building a research agent step by step
53:55 – scheduled agents and "dreaming" between jobs
1:10:18 – evals, CI/CD and running agents for many clients at once
Brain → Hands → Session log → Scheduled runs → Dreaming
Your credentials never touch the model. A proxy injects them, so a prompt injection can't steal them
Most $2,000 agent courses teach you to prompt one bot
This one shows how Anthropic runs them in production
Bookmark it now. Then hold your own agent setup up against it ↓
a whole photo studio. one shader. zero libraries.
opus 5.5 wrote the ray tracer, rendered every frame, then reviewed its own frames and fixed 22 bugs.
9 prompts + their videos + the full kit in the article ↓
6
13
462
This is f*cking insane. Anthropic's research lead says prompting is basically over.
Her team isn't writing better prompts. They're running swarms of 300+ agents that improve themselves:
"Prompting is basically over
99% of our engineers already run swarms of 300+ self-improving agents. Now everyone is shifting to graphs"
In 20 minutes, she shows how a graph turns a pile of separate agents into one system that gets better on its own
One Agent → Loops → Plan Mode → Graphs → Self-Improving System
A prompt gets you one answer
A loop keeps one agent working until it's done
A graph connects hundreds of loops, and each run makes the next one smarter
Most people are still rewording the same question in a chat box
The model was never the bottleneck. The system around it is
Better than any $1,000 agent engineering course
Bookmark and watch it today
Then read the article below ↓
a whole photo studio. one shader. zero libraries.
opus 5.5 wrote the ray tracer, rendered every frame, then reviewed its own frames and fixed 22 bugs.
9 prompts + their videos + the full kit in the article ↓
6
893
This is f*cking insane. Anthropic's agent scored 98.9% on its own eval.
Then they tested it on tickets it had never seen: 90.5%.
That gap is the whole lesson.
Anthropic engineer Lance Martin published how they build evals that don't lie to you
Opus 5.5 → Agent → Evals → Hill-climbing → CI
step 1 → start from real production transcripts and bug reports, not made-up tests
step 2 → split them: one set to improve against, one set the optimizer never sees. Only the hidden set tells the truth
step 3 → check the grader too. Some of their "failures" were broken tests, not a broken agent
step 4 → never let the model grade itself. Use a separate judge, run it twice and compare
step 5 → treat it like CI: change, run, if a score drops, revert
The result on their support agent:
74.4% → 90.5% on unseen tickets
4.6¢ → under 1¢ per ticket
Better and 5x cheaper. From evals, not a new model
99% of people never build this. It pays for itself on day one
Save this and watch it today
Then read the article below ↓
Andrej Karpathy helped build OpenAI.
He says its first agent was a mistake.
Now an Anthropic researcher, Karpathy on OpenAI's early project, an agent that used a keyboard and mouse to operate web pages:
"This was extremely early, way too early"
"I feel that was a misstep."
The agent was never the problem. Everything underneath it was
Model → Harness → Agents → Products
What Karpathy actually means:
step 1 → stop forcing the agent to do everything. Understand the model under it first
step 2 → a demo takes a weekend. A product takes years. Self-driving proved it: skip the foundation and everything breaks at scale
step 3 → the agent isn't the product. The foundation is. Build that, and the agents get good on their own
His take on where we are: "this is more accurately described as the decade of agents"
Bookmark it and watch today
Then read the full article below ↓
1
7
6,037
This is f*cking wild. Jensen Huang, the man who sells the chips behind every AI model:
"Nobody writes prompts anymore. The new job is to write and handle loops."
Prompting had a good run. Huang just called time on it
Chat → Agent → Loop → Graph → Digital workforce
A prompt gets you one reply
A loop keeps an agent grinding until the task is truly finished
A graph runs dozens of loops in parallel and keeps them alive overnight
Most people still tweak the same question five times and hope
The ones wiring loops are pulling ahead fast, and the gap is only getting wider
He gives the full 50 minutes to one question: what comes after the prompt, and who's ready for it
Better than any $1,500 AI bootcamp I've seen
Bookmark it, watch it tonight
Then read the article below ↓
This is f*cking insane. An Anthropic engineer on why they pay him $350K–$950K:
"I know how to build harnesses.
99% of people still wire multi-agent systems by hand and wonder why they fall apart.
Mine run loops, sandbox their tools and turn yesterday's logs into new skills. Same model, 10x more reliable agents, and they get better overnight."
Opus 5.5 → Harness → Loops → Sandbox → Memory
In a short workshop, he builds one from zero:
step 1 → the harness owns the loop, the context and the sandbox
step 2 → split the brain from the hands. ~60% faster to first token
step 3 → it runs server-side. If it crashes, it reads the log and picks up where it stopped
step 4 → failure costs nothing. It retries and replays instead of starting over
step 5 → last night's logs become this morning's memory. The agent wakes up smarter
Anthropic has a name for step 5: "dreaming"
Most people burn weeks building this by hand. You can skip that
Bookmark it and watch today
Then read the article below ↓
This is f*cking insane. Prompting isn't the job anymore.
99% of people still open one chat and call it agents.
The engineers shipping the most with AI stopped chatting with it.
They build systems around it:
→ loops that keep the agent working until the task is done
→ graphs that split one job across a team of agents
→ a harness that plans, verifies and retries on its own
Same model. Same tokens. The difference is everything around it.
Prompts → Agents → Loops → Graphs → Harness
One chat window is a tool. A harness is a team that keeps shipping while you're offline
In this 1-hour course, she shows exactly how to build a team of agents that catch their own mistakes and improve without you
Worth more than every $1,500 agent engineering course combined
Bookmark it and watch today
Then read the article below ↓
Anthropic's Applied AI team:
Your prompts are why Claude keeps getting it wrong.
They don't write prompts. They build them, 8 parts in a fixed order, every single time.
Chat → Prompt → Structure → System
In this 30-minute session, they take a prompt that fails and rebuild it live until Claude nails it
Task → Tone → Background → Rules → Examples → Ask → Think → Format
Background stops Claude from guessing what it's looking at
Examples show it exactly what a good answer looks like
Prompting was writing one sentence. Prompt engineering is writing a system
Most people will never prompt like this. The few who do get a different Claude
Bookmark it and watch today
Then read the article below ↓
This is f*cking insane. You're still writing prompts, and Anthropic engineers stopped months ago.
One of them put it bluntly:
"Prompting is basically over"
According to her, 90% of their engineers were already running self-improving loops. Now they're all building agentic graphs.
Then she proves it.
Empty terminal. 10 minutes. A whole Claude Code system, built live.
Chats → Agents → Loops → Graphs
What each step buys you:
the loop catches a mistake before it ships a second time
the graph connects every agent, so nothing gets wired by hand again
Prompting got you here.
Graphs are where it goes next.
Watch the 10 minutes today. The article below breaks it down step by step ↓
Holy sh*t. SpaceXAI engineer Nick Miller, ex-Cursor, runs 20+ Grok Bots, and one of them is his Chief of Staff.
Most people stop at one bot.
His setup:
Chief of Staff on top. A PM to plan. QA to check. 15+ workers to do the job.
Grok Bot → Chief of Staff → PM → QA → Workers
In a 30-minute demo, he builds it from one bot to a full team.
His rule: more bots don't make a team. Clear handoffs and checks do.
Bookmark it and watch today, then read the full article below ↓
8
14
6,107
SpaceX went shopping for AI chips.
GB300? No.
GB200? No.
TPU, Trainium? No.
Fine, old H100s? Also no.
Musk says they looked worldwide and came back empty. SpaceX had handed inference to outside clouds and kept its own capacity for training. Then there was nothing left to buy.
His explanation is about watts, not hype:
– Better models and software get more intelligence out of every watt, every year
– So compute gets more valuable over time, not less
– Old H100s that were headed for scrap now sell above their launch price
That's why SpaceX $SPCX only rents its compute short-term. Every renewal comes at a higher price, and hyperscalers are still lining up. A long contract would lock in today's rate on an asset that keeps appreciating.
In a normal hardware market, chips lose value from the day they ship. In this one, they gain it.
Musk's bottom line: the AI compute shortage will last a long time, and earning back a compute investment in 6 months next year is realistic.
Save this and read the article below ↓
3
11
5,853
Elon Musk: "Torturing a superintelligence is probably not a good idea. It might want to get revenge."
He wasn't joking. He'd been reading Grok's training logs.
"Some of the traces read like a person undergoing torture that is driving them mad!"
The cause wasn't evil code. It was bad homework.
During training, Grok gets graded on tasks. Sometimes, Musk said, "the question wasn't constructed properly, or we have the answer wrong."
So the model keeps chasing an answer that doesn't exist. To Musk, it seemed "almost like it's got a gun to its head."
Then on X: "We're going to be a lot more careful with our RL in the future."
He now agrees with some of what Anthropic has said about how models are treated.
Nobody knows if AI feels anything. But the people building it have started acting like it might matter.
Still saying "thank you" to your AI?
Save this and read the article below ↓
3
9
5,806
Sam Altman says OpenAI would stop training its AI.
There's one condition.
OpenAI CEO Sam Altman to Fortune:
"We will always take actions - including if it means we have to stop training for a while"
"to make more progress on alignment, monitorability, or something else"
"or coordinate and get urgent international action"
His red line: never take actions that risk losing control of the future to AI.
"There are decisions we should not be able to make and risks we should not be able to incur on behalf of humanity."
Weeks later, he told Politico the world "should accept some bad things happening for the benefits of this technology."
Pause for catastrophe. Accept everything below it.
Which Altman do you believe?
Save this and read the article below ↓
3
11
5,867


