Pinned Tweet
Dario Amodei's son asked to borrow his Claude account. His answer: "absolutely not, I need my tokens."
That's the lightest moment of a 65-minute interview. The rest: why the first companies to test Mythos called it "a super weapon" and asked Anthropic not to release it, what is happening to engineers inside Anthropic, and what it feels like to run the company at the center of AI right now.
Watch now, save for later.
+10% software engineer jobs. +20% radiology jobs. AI automated the core task of both.
That's Jensen Huang's evidence, from the NVIDIA CEO's talk to students:
"The narrative about AI destroying jobs is exactly backwards. AI automates tasks away, but it doesn't necessarily eliminate jobs."
His example: "They said Harvey is going to eliminate all of the paralegal jobs... Turns out paralegals are growing like crazy."
The figures are his, quoted from the talk.
ReSourse retweeted
Dario Amodei's son asked to borrow his Claude account. His answer: "absolutely not, I need my tokens."
That's the lightest moment of a 65-minute interview. The rest: why the first companies to test Mythos called it "a super weapon" and asked Anthropic not to release it, what is happening to engineers inside Anthropic, and what it feels like to run the company at the center of AI right now.
Watch now, save for later.
ReSourse retweeted
NVIDIA doesn't pick one AI coding agent for its engineers. Jensen Huang lets them use all of them.
Claude Code runs autonomously in sandboxes all over the company. Some teams use Codex, some Cursor, some Cognition. NVIDIA learns from every one.
And his bar for agents is lower than you'd think:
"We don't need the agents to be 100% accurate... It could literally be 80%, and then we help it the rest of the way."
NVIDIA doesn't pick one AI coding agent for its engineers. Jensen Huang lets them use all of them.
Claude Code runs autonomously in sandboxes all over the company. Some teams use Codex, some Cursor, some Cognition. NVIDIA learns from every one.
And his bar for agents is lower than you'd think:
"We don't need the agents to be 100% accurate... It could literally be 80%, and then we help it the rest of the way."
DEMIS HASSABIS: Silicon Valley's "move fast and break things" is the wrong way to build AI.
"I think it's not appropriate, in my opinion, for this type of transformative technology. I think instead, you know, we should be trying to use the scientific method and approach it with a kind of humility and respect that this kind of technology deserves."
Via his lecture at Cambridge
ReSourse retweeted
This is how a SpaceX AI engineer ships a feature now:
She writes down the task at the end of the day.
She fires off a cloud agent, closes her laptop and goes for a walk.
The agent builds the whole page, figures out how to test it on its own, and sends back a video proving it works.
"implementing a feature like this probably would have taken me weeks. And Cursor was able to do it in an hour."
Amrita Venkatraman, SpaceX AI
This is how a SpaceX AI engineer ships a feature now:
She writes down the task at the end of the day.
She fires off a cloud agent, closes her laptop and goes for a walk.
The agent builds the whole page, figures out how to test it on its own, and sends back a video proving it works.
"implementing a feature like this probably would have taken me weeks. And Cursor was able to do it in an hour."
Amrita Venkatraman, SpaceX AI
ReSourse retweeted
KARPATHY: LOGGED OUT, CHATGPT DOESN'T EVEN TELL YOU WHICH MODEL YOU'RE TALKING TO. IT MIGHT BE "A SMALLER KIND OF DUMBER MODEL"
myth: the $200 thinking model always wins
reality: he gave one bug to o1 pro. it thought for a minute and fixed it. Claude 3.5 Sonnet fixed the same bug with no thinking at all
myth: pick one model and stick with it
reality: he pays for several and asks them all the same question. he calls it his "LLM council"
his rule: try without reasoning first, switch only if it fails
00:45 what the free tier actually gives you
02:17 o1 pro vs Claude on the same bug
03:24 when to reach for a reasoning model
for a bot desk: pin the exact model in the config. never let a vendor swap it under you
Bookmark and watch it today
KARPATHY: LOGGED OUT, CHATGPT DOESN'T EVEN TELL YOU WHICH MODEL YOU'RE TALKING TO. IT MIGHT BE "A SMALLER KIND OF DUMBER MODEL"
myth: the $200 thinking model always wins
reality: he gave one bug to o1 pro. it thought for a minute and fixed it. Claude 3.5 Sonnet fixed the same bug with no thinking at all
myth: pick one model and stick with it
reality: he pays for several and asks them all the same question. he calls it his "LLM council"
his rule: try without reasoning first, switch only if it fails
00:45 what the free tier actually gives you
02:17 o1 pro vs Claude on the same bug
03:24 when to reach for a reasoning model
for a bot desk: pin the exact model in the config. never let a vendor swap it under you
Bookmark and watch it today
THE MAN WHO BUILT CLAUDE CODE WAS ASKED FOR THE HARDEST PART. HE DIDN'T SAY THE MODEL. HE SAID THIS:
"Bash is inherently pretty dangerous"
3 rules from his 3 minutes I'd tape above any bot desk:
1. read-only commands don't wait for a click
2. one config line and a URL "can never be fetched". no employee can override it
3. the bot lives in a pipe: prompt in, JSON out. they run it in CI and incident response
01:09 the policy file
02:03 claude -p
02:29 how they made bash safe without approving every command
the bot decides. the limits in code say no
8
383
HOLY SH*T. SAM ALTMAN AT THE FEDERAL RESERVE: SOME BANKS STILL LET YOU MOVE "A LOT OF MONEY" WITH JUST YOUR VOICE. "AI HAS FULLY DEFEATED THAT"
"a significant impending fraud crisis"
voice prints → selfies and waves → cloned calls from your kid → video calls "indistinguishable from reality"
00:00 - the banks that still take a voice print
00:13 - "AI has fully defeated most of the ways that people authenticate"
00:50 - the fake urgent call in your kid's voice
01:08 - next: a video FaceTime you can't tell from real
steal it for a bot desk: no voice, no selfie. keys sit in a vault the bot can't read, and every order passes a check in code
while you trust a voice on the phone, someone is cloning it
Bookmark and watch it today
Then read how I wired my Grok bot desk in the guide below ↓
1
10
333
ReSourse retweeted
THIS IS F*CKING WILD. ELON MUSK: ONE SLIGHTLY BETTER DECISION AT TESLA CAN BE WORTH $100 MILLION IN AN HOUR
"if I have zero recreational time, then I make worse decisions"
$100B a year → $2B a week → a slightly better call swings $1B → so he judges it in percentages, not dollars
00:05 - Tesla's $2 billion week
00:22 - the $100 million hour
00:46 - why he thinks in percentages, or he'd "never get any sleep"
01:37 - depressed = worse decisions
steal it for a bot desk: size every trade as a % of the account, never in dollars
while you size trades on feel, he prices his hours at $100M
Bookmark and watch it today
Then read how I wired my Grok bot desk in the guide below ↓
THIS IS F*CKING WILD. ELON MUSK: ONE SLIGHTLY BETTER DECISION AT TESLA CAN BE WORTH $100 MILLION IN AN HOUR
"if I have zero recreational time, then I make worse decisions"
$100B a year → $2B a week → a slightly better call swings $1B → so he judges it in percentages, not dollars
00:05 - Tesla's $2 billion week
00:22 - the $100 million hour
00:46 - why he thinks in percentages, or he'd "never get any sleep"
01:37 - depressed = worse decisions
steal it for a bot desk: size every trade as a % of the account, never in dollars
while you size trades on feel, he prices his hours at $100M
Bookmark and watch it today
Then read how I wired my Grok bot desk in the guide below ↓
HOLY SH*T. ANTHROPIC'S AGENTS NOW "DREAM". THEY READ EACH OTHER'S MISTAKES, AND THE NEXT DAY'S AGENTS SHOW UP SMARTER
"the next day's agents automatically get better"
agents write memory → dreaming reads every session → merges duplicates, drops stale notes, verifies the rest → next run starts smarter
01:16 - one team cut first-pass mistakes by 90%
01:47 - the runbook is read-only, the working notes are read-write
02:47 - Harvey: 6x task completion in one legal scenario
05:22 - the 60-second pattern no single agent noticed
steal the setup for a bot desk: rules read-only, trade notes read-write, every memory change logged
while your bot forgets every session, theirs learns from the whole team
Bookmark and watch it today
Then read how I wired my Grok bot desk in the guide below ↓
8
292
THIS IS F*CKING WILD. NOBEL WINNER DEMIS HASSABIS: HIS AI WENT FROM PLAYING RANDOMLY IN THE MORNING TO STRONGER THAN A WORLD CHAMPION IN 24 HOURS OR LESS
"you want to be like five years ahead of your time, not 50 years ahead"
v1 plays itself 100,000 games → v2 learns from them → v2 must win 55% to replace v1 → repeat 17-18 times
00:24 - why hand-coded systems break on the unexpected
01:37 - self-play and the 55% rule
03:56 - his 3 tests: huge search space, one clear score, data or a simulator
05:11 - the first couple of years "nothing worked"
steal the 55% rule: a new bot version doesn't go live until it beats the old one head to head
while you tweak your bot on gut feel, his ran 17 versions in a day
Bookmark and watch it today
Then read how I wired my Grok bot desk in the guide below ↓
ReSourse retweeted
HOLY SH*T. AN ANTHROPIC ENGINEER SHOWED WHY WRITING "CRITICAL: DO THE MATH RIGHT" IN YOUR BOT'S PROMPT DOES NOTHING
"instructions don't add capability"
kill old patches → give it a tool for the math → state both sides of every trade-off
01:25 - an old patch made the bot hide numbers it had
02:51 - "critical" doesn't fix mental math. a tool does
03:10 - the $8 rule that stopped the bot escalating
opus with adaptive thinking passed, but tripled the tokens and the latency. three small prompts in a loop passed too
my desk works the same way: the limits live in code, not in the prompt.
while you shout CRITICAL in caps, they hand the bot a calculator
Bookmark and watch it today
HOLY SH*T. AN ANTHROPIC ENGINEER SHOWED WHY WRITING "CRITICAL: DO THE MATH RIGHT" IN YOUR BOT'S PROMPT DOES NOTHING
"instructions don't add capability"
kill old patches → give it a tool for the math → state both sides of every trade-off
01:25 - an old patch made the bot hide numbers it had
02:51 - "critical" doesn't fix mental math. a tool does
03:10 - the $8 rule that stopped the bot escalating
opus with adaptive thinking passed, but tripled the tokens and the latency. three small prompts in a loop passed too
my desk works the same way: the limits live in code, not in the prompt.
while you shout CRITICAL in caps, they hand the bot a calculator
Bookmark and watch it today
THIS IS F*CKING WILD. A SPACEX AI ENGINEER DOESN'T RUN ONE GROK BOT. HE RUNS A STAFF: GROWTH ENGINEER, FOUNDING ENGINEER, CREATIVE DIRECTOR, CHIEF OF STAFF
"I think about bots as, like, different employees, really."
one bot per role → a new bot for every new task → too much scope? narrow it or clear the context
00:19 - matt's bot roster, one employee per role
00:59 - how to tell a bot has too much scope
01:55 - the idea: a bot that lights up every $1,000 in sales
02:39 - dr. eggbot, a bot taught to make other bots
my desk runs the same way: CRAWLNET reads the coin page, Grokbot decides, FOMO fills. one job each, nothing else
while you cram everything into one chat, they hire bots like staff
Bookmark and watch it today
Then read how I wired my Grok bot desk in the guide below ↓
2
8
224
HOLY SH*T. GROK BOT NOW HAS A VAULT. IT FILLS IN YOUR PASSWORDS WITHOUT EVER SEEING THEM
"the bot doesn't have access to the information. it's encrypted"
you type it once → it's encrypted in the vault → the bot fills it in on the page → the bot never reads it
00:12 - pages: the bot turns your data into a shareable visual
01:05 - the vault: passwords and sensitive info the bot can't see
02:01 - a deeper 1password integration
05:46 - fewer approval cards, same review on sensitive actions
for a bot desk this is the whole game: the bot needs your logins to work, it should never be able to read them
while you paste keys into chats, other people's bots fill them in blind
Bookmark and watch it today
Then read how I wired my Grok bot desk in the guide below ↓
6
230







