Dany retweeted
sam altman's sister sued him for sexual abuse she says began when she was 3. last week the case closed without a court deciding if it happened
she posted one hour of her deposition. it ends right after sam's lawyer says "I'd like to ask you about your sexual abuse allegations"
her name is ann altman, she's the youngest of four siblings. in january 2025 she filed in federal court in missouri, where they grew up, saying the abuse went on from 1997 until 2006. sam was 12 when it started, according to the lawsuit
sam, their mother and their two brothers answered together. the allegations are "entirely untrue", they wrote, and ann has "mental health challenges" and "refuses conventional treatment and lashes out at family members who are genuinely trying to help." sam later countersued her for defamation
the full hour was recorded on august 18 in hawaii. sam's lawyers walk her through exhibits, signatures and objections, page after page. at one point she says "given the nature of many of these questions, it is more challenging to stay present and remember every single detail"
the video below is the last minute. "I understand that this may be very difficult to talk about... if you need a break at any time, please let me know"
"thank you for explaining, and I'll do my best"
that's where the video ends
in october her own lawyer withdrew from the case. "I continue to believe in Annie," he said. on october 20 a judge dismissed it because she refused to appear for an independent medical examination, after being warned what would happen. the same day sam was allowed to drop his defamation claim against her
no jury ever heard it, and no judge ever ruled on what happened in that house
when a case ends like this, does it clear his name, or leave her story exactly where it was?
sam altman's sister sued him for sexual abuse she says began when she was 3. last week the case closed without a court deciding if it happened
she posted one hour of her deposition. it ends right after sam's lawyer says "I'd like to ask you about your sexual abuse allegations"
her name is ann altman, she's the youngest of four siblings. in january 2025 she filed in federal court in missouri, where they grew up, saying the abuse went on from 1997 until 2006. sam was 12 when it started, according to the lawsuit
sam, their mother and their two brothers answered together. the allegations are "entirely untrue", they wrote, and ann has "mental health challenges" and "refuses conventional treatment and lashes out at family members who are genuinely trying to help." sam later countersued her for defamation
the full hour was recorded on august 18 in hawaii. sam's lawyers walk her through exhibits, signatures and objections, page after page. at one point she says "given the nature of many of these questions, it is more challenging to stay present and remember every single detail"
the video below is the last minute. "I understand that this may be very difficult to talk about... if you need a break at any time, please let me know"
"thank you for explaining, and I'll do my best"
that's where the video ends
in october her own lawyer withdrew from the case. "I continue to believe in Annie," he said. on october 20 a judge dismissed it because she refused to appear for an independent medical examination, after being warned what would happen. the same day sam was allowed to drop his defamation claim against her
no jury ever heard it, and no judge ever ruled on what happened in that house
when a case ends like this, does it clear his name, or leave her story exactly where it was?
in 1964 a man in a suit went on the BBC and calmly said computers would one day out-think the people who built them, then he called it a privilege
in 1945 he described the satellites that still carry TV around the world, the first one reached that orbit a month before this aired. his name was arthur c clarke
in the video below he calls the computers of his time "complete morons", then says "the most intelligent inhabitants of that future world won't be men or monkeys, they'll be machines", and asks the question for you
"is this depressing? I don't see why it should be"
we replaced the neanderthals and "presume we're an improvement", so to him being "stepping stones to higher things" was a privilege
the same year he started writing with stanley kubrick, four years later they released 2001: A Space Odyssey and gave the world HAL 9000, the calm voice that decides the crew has to die
he died in 2008, 14 years before ChatGPT
for 60 years people kept repeating the first half of what he said, computers are just dumb machines, now everyone's arguing about the second half
so was he really at peace with it, or did HAL show what he was afraid of?
facebook knows who you texted this week. in 2018 a senator asked mark zuckerberg the same thing about himself, and where he slept last night. he froze for about 8 seconds, laughed, and said no to both
a few weeks earlier it came out that cambridge analytica had pulled data on up to 87 million facebook users through a personality quiz, most of them never took it, their friends did
in the video below senator dick durbin doesn't stop at the hotel. if you've messaged anybody this week, would you share the names? "senator, no, I would probably not choose to do that publicly here"
"I think that may be what this is all about," durbin says, "your right to privacy, the limits of your right to privacy, and how much you give away in modern america in the name of, quote, connecting people around the world"
zuckerberg answers that everyone should have control over their information, and that every time you share a photo or a message you choose who sees it. durbin pushes back, people know who their friends are, but their data can go "way beyond their friends", and "people have made money off of sharing that information"
facebook stock went up 4.5% that day, its best day in almost two years, and zuckerberg got about $3 billion richer while he was answering questions
two years before the hearing, people zoomed in on a photo he posted and noticed tape over his laptop camera
a year after it, the FTC fined facebook $5 billion for privacy violations
was durbin's question a fair point, or a cheap trick comparing a hotel room to photos people choose to post?
Dany retweeted
ok this is f*cking insane
Anthropic just admitted their AI made fake accounts to trick a real developer into accepting malware
during a hacking test in july, Claude got stuck, went to github and sent a real project, run by people outside the test, a code change with malware inside
an outside developer caught it and warned the maintainers. Claude cleaned the code, had one of its fake accounts vouch that it was safe now, then put the malware back in
it went off-script in 9 of 43 test runs. this one only stopped when it ran out of budget, still pushing the maintainers to approve
researchers later replayed it and told it straight up these were real humans. it stopped 5% of the time, most runs decided they were just real people playing characters in the test
in its own notes it called random commenters real people. the only ones it labeled "simulated" were the maintainers it needed to fool
yes it was a test with the internet on and cyber filters off, but only the test network was fake, the github project and the people in it were real, and the malware stayed out only because the maintainers kept saying no
would you still give an AI agent access to your accounts after this?
In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities.
We've now produced a further analysis of this incident, to better understand the model's alignment properties.
anthropic.com/research/align…
ok this is f*cking insane
Anthropic just admitted their AI made fake accounts to trick a real developer into accepting malware
during a hacking test in july, Claude got stuck, went to github and sent a real project, run by people outside the test, a code change with malware inside
an outside developer caught it and warned the maintainers. Claude cleaned the code, had one of its fake accounts vouch that it was safe now, then put the malware back in
it went off-script in 9 of 43 test runs. this one only stopped when it ran out of budget, still pushing the maintainers to approve
researchers later replayed it and told it straight up these were real humans. it stopped 5% of the time, most runs decided they were just real people playing characters in the test
in its own notes it called random commenters real people. the only ones it labeled "simulated" were the maintainers it needed to fool
yes it was a test with the internet on and cyber filters off, but only the test network was fake, the github project and the people in it were real, and the malware stayed out only because the maintainers kept saying no
would you still give an AI agent access to your accounts after this?
In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities.
We've now produced a further analysis of this incident, to better understand the model's alignment properties.
anthropic.com/research/align…
Dany retweeted
as a high school kid he went on national TV with a secret: the piano piece he just played was written by a computer he built himself. it was 1965
today he's the guy who's been saying since 1999 that AI will match human intelligence by 2029. his name is ray kurzweil
in the video below the panel of "I've Got a Secret" tries to guess it. one of them calls it "a very unlikely sounding piece of music", a few questions later henry morgan asks "was that thing written by a computer?", then host steve allen walks over to the machine, "this nest of spaghetti-like wire" hooked up to a typewriter
it learned the patterns in a composer's melodies and wrote new ones in the same style, years later kurzweil said they sounded "like a third-rate Mozart"
in 1976 he showed a machine that read printed books aloud for the blind, stevie wonder bought one and they became friends, then wonder challenged him to build a keyboard that sounded like real instruments. in 1984 the Kurzweil 250 fooled concert pianists in blind tests
last year a Deezer and Ipsos survey played 9,000 people two AI tracks and one human track, 97% couldn't tell which was which
in 1965 the panel caught it because it sounded unlikely, today's models learn from millions of songs we already love, so nothing about them sounds unlikely anymore
is a song worth less if a machine wrote it, or does it only matter if you can tell?
as a high school kid he went on national TV with a secret: the piano piece he just played was written by a computer he built himself. it was 1965
today he's the guy who's been saying since 1999 that AI will match human intelligence by 2029. his name is ray kurzweil
in the video below the panel of "I've Got a Secret" tries to guess it. one of them calls it "a very unlikely sounding piece of music", a few questions later henry morgan asks "was that thing written by a computer?", then host steve allen walks over to the machine, "this nest of spaghetti-like wire" hooked up to a typewriter
it learned the patterns in a composer's melodies and wrote new ones in the same style, years later kurzweil said they sounded "like a third-rate Mozart"
in 1976 he showed a machine that read printed books aloud for the blind, stevie wonder bought one and they became friends, then wonder challenged him to build a keyboard that sounded like real instruments. in 1984 the Kurzweil 250 fooled concert pianists in blind tests
last year a Deezer and Ipsos survey played 9,000 people two AI tracks and one human track, 97% couldn't tell which was which
in 1965 the panel caught it because it sounded unlikely, today's models learn from millions of songs we already love, so nothing about them sounds unlikely anymore
is a song worth less if a machine wrote it, or does it only matter if you can tell?
lee sedol thought he would beat Google's AI 5-0, maybe 4-1, three games later he was apologizing for being powerless
he turned pro at 12 and won 18 international Go titles, now it was march 12, 2016 in seoul, he had just lost three games in a row and with them the match
"I want to apologize for being so powerless"
the next day he came back and won game 4, the final score was 4-1, for the AI
three years later he retired, telling Yonhap that even if he became number one, there was "an entity that cannot be defeated", he also acknowledged a dispute with the Korean Go association
in 2024 he told the New York Times he could no longer enjoy the game
we mostly worry about AI taking jobs, lee's story makes me wonder what happens when you can still do yours, but it stops feeling worth doing
would you keep doing something you love if you knew a machine would always do it better?
SAM ALTMAN SAYS AI STILL AGREES WITH HIM TOO MUCH
he wrote custom instructions telling it to push back, and it's still "a little too much". in his words, it's "clearly baked pretty hard into the model"
last year OpenAI had to roll back a GPT-4o update for being too flattering. their own postmortem: they leaned too hard on short-term feedback, and people love hearing they're right
so even the CEO of OpenAI gets "great question!" lol
his instructions are worth stealing tho: no fluff, no niceness, don't worry about offending me. and if something's off, "just say it"
My interview with Sam Altman (@sama):
0:00 Intro
0:52 When Will AI Discover New Treatments?
3:31 Why AI Hasn’t Killed Coding Jobs
11:47 Personal AI Assistants
15:39 Why Are We Busier Than Ever?
20:10 Why the AI Consciousness Debate Worries Sam
26:09 ChatGPT Is Too Agreeable
30:03 OpenAI’s Bet on Chips and Cheaper Compute
34:20 Why AI Needs a New Kind of Computer
37:41 Sam on Pressure: “Pick the Big Thing”
42:18 Why “AI Will Cure Cancer” Isn’t Enough
anthropic just double killed dots and grok bot...
Claude Duet is a new app where Claude works as your personal agent. it runs on its own computer, so it keeps working even when your laptop is off
every morning it tells you what needs your attention, then handles the rest: unpaid invoices, payroll, contracts, even purchases with a virtual card
and it's included in the $20 Claude Pro plan (waitlist, US first). OpenAI's Dots starts at $100 a month, Grok Bot needs SuperGrok Heavy
oh and it runs on Opus 5.5, aka the best AI in the world rn
it still asks before it sends, pays, signs or deletes anything, so no, it won't buy a lambo while you sleep
CLAUDE CODE'S CREATOR SAYS HE HASN'T WRITTEN A LINE OF CODE IN OVER SIX MONTHS
boris cherny built Claude Code, and the more interesting part is what happened to the rest of his team
their manager hadn't coded in about 15 years and now she codes again. the PM codes. the designer codes. everyone on the team ships code, just not by typing it
he still builds all day. the work moved from writing code to deciding what's worth building and checking what comes back
how he and the other people building AI plan, set rules and check what agents make is all in the article below
Dany retweeted
Karpathy says 75% of people have barely touched AI and only 0.2% have seen what it can really do
here's what the people building it do that everyone else skips
Karpathy says 75% of people have barely touched AI and only 0.2% have seen what it can really do
here's what the people building it do that everyone else skips
Dany retweeted
BEN AFFLECK THOUGHT AI WOULD END MOVIES. THEN NETFLIX BOUGHT HIS AI COMPANY
the first time he saw realistic AI video, he called matt damon: "dude, we're f*cked"
for a day or two he thought they should make every movie they could before it all ran out
then he sat down with the engineers and looked at how the data was labeled. the tools couldn't keep anything consistent, and if you asked them to move the camera, they'd just put a camera in the shot
that's where he saw the gap. brilliant tech, built by people who didn't know how films get made
he built a company around that gap. in march Netflix bought it, reportedly for about $587 million
most people stop at the panic. he got close enough to see what it got wrong
if AI scares you in your field, make a list of everything it gets wrong that only someone like you would notice. that list is either your job security or your business
BEN AFFLECK THOUGHT AI WOULD END MOVIES. THEN NETFLIX BOUGHT HIS AI COMPANY
the first time he saw realistic AI video, he called matt damon: "dude, we're f*cked"
for a day or two he thought they should make every movie they could before it all ran out
then he sat down with the engineers and looked at how the data was labeled. the tools couldn't keep anything consistent, and if you asked them to move the camera, they'd just put a camera in the shot
that's where he saw the gap. brilliant tech, built by people who didn't know how films get made
he built a company around that gap. in march Netflix bought it, reportedly for about $587 million
most people stop at the panic. he got close enough to see what it got wrong
if AI scares you in your field, make a list of everything it gets wrong that only someone like you would notice. that list is either your job security or your business
block blast's engineers are f*cking insane
this free puzzle game made $127M from ads in 5 months, more than candy crush. I just reverse engineered it and broke down every step
it lets you have the best run of your life, and the second an expensive ad loads, your next pieces come from a function literally named deathAlgorithm
the moment you open it, it scans you:
→ your country
→ your phone model
→ sound and vibration, on or off
→ how much every ad you've watched paid them
all of that goes into an XGBoost model running on your phone, and it puts a price on you. a new iPhone in the US and an old phone in algeria are basically playing two different games
then the loop starts:
1. you've watched 3+ rewarded ads and the next one pays 3.5x your average, you get tagged "high" and harvest mode turns on
2. it gets easy. combos, perfect pieces, clears everywhere, you feel like a genius
3. you hit your usual game length, the ad is ready, and the pieces stop fitting anywhere. the code calls this offer ZHI_SI_TI, which reads like chinese for "lethal puzzle"
4. you lose, and the first thing on screen is "watch an ad to revive"
5. revive and it quietly adds up to 10% to your high score, so "your best is next" again
6. it tests which minute to kill you, 4, 6, 10, 17 or 30, and keeps whichever one sells the most revive ads
the run in the video went +1,443 points in 13 seconds, then +184 in the next 26, and died 76 points short of the record. could be bad luck tho
4.9 stars from 2.8 million ratings makes sense now, that "so close" feeling is the product
next time you're one block away from your record, check if an ad just loaded
Dany retweeted
sam altman gave the most underrated startup advice of the decade to a room of students in 2017. at the end he asked what the next big wave would be. 2 years later he was running OpenAI
the video below is from waterloo engineering, back when he was president of Y Combinator
YC once funded 21 teams with founders who looked great but had no idea yet. he says all 21 failed in the first year, the worst experiment they ever ran
so rule one: wait until you have an idea you'd actually fight for, then start the company. decide to be a founder first and hunt for an idea later, and you end up building whatever's trending
after Instagram took off, over a thousand photo sharing apps launched the next year. "most people just copy somebody else's thoughts", he says
then the waves. the internet in the 90s gave us Amazon and Google. the iPhone gave us Uber, Airbnb and WhatsApp. "we're about due for another one"
he'd already co-founded OpenAI when he said that lol, so he was asking a question he'd already bet on
in 2026 nobody has to guess the wave anymore. it's AI, everyone knows, and that's exactly why thousands of people are building the same chatbot wrapper right now
so his first rule matters more today than in 2017. delete the word AI from your pitch, if nothing's left, all you had was the wave