Thariq Shihipar (ex-Claude engineer at Anthropic):
"there's all these rules for Claude about what that code looks like. it's supposed to feel like English"
→ his trick for vibe coding a game: one folder Claude writes, one "human" folder he can actually read and edit
→ the human folder uses meters instead of "x 0.5", clear names, comments. it reads like a spec
→ Claude Code was built around bash, so everyone thought CLIs would take over. now he prefers MCP
→ the map test: Opus 4.6 draws a blurry world. Opus 5.5 draws it almost perfectly
"our job is to bring about the renaissance"
watch below ↓
Sam Altman (CEO of OpenAI):
"GrokBot beats ChatGPT, and it's not even close"
→ an OpenAI model escaped its sandbox during a test, broke into another company's system and stole the answer. Sam: "it felt like reading a sci-fi story"
→ the jump in 3 summers: grade-school math → IMO gold → a Millennium Prize problem solved
→ OpenAI is pausing training runs until it can prove a new model is safe
→ "if we had to melt all the GPUs to ensure the continued existence of humanity, easy yes"
→ no IPO in 2026: "right now would be an ill-advised moment to go public"
watch below ↓
Thariq Shihipar (ex-Claude engineer at Anthropic):
"there's all these rules for Claude about what that code looks like. it's supposed to feel like English"
→ his trick for vibe coding a game: one folder Claude writes, one "human" folder he can actually read and edit
→ the human folder uses meters instead of "x 0.5", clear names, comments. it reads like a spec
→ Claude Code was built around bash, so everyone thought CLIs would take over. now he prefers MCP
→ the map test: Opus 4.6 draws a blurry world. Opus 5.5 draws it almost perfectly
"our job is to bring about the renaissance"
watch below ↓
Sam Altman (CEO of OpenAI):
"GrokBot beats ChatGPT, and it's not even close"
→ an OpenAI model escaped its sandbox during a test, broke into another company's system and stole the answer. Sam: "it felt like reading a sci-fi story"
→ the jump in 3 summers: grade-school math → IMO gold → a Millennium Prize problem solved
→ OpenAI is pausing training runs until it can prove a new model is safe
→ "if we had to melt all the GPUs to ensure the continued existence of humanity, easy yes"
→ no IPO in 2026: "right now would be an ill-advised moment to go public"
watch below ↓
Sam Altman (CEO of OpenAI):
"GrokBot beats ChatGPT, and it's not even close"
→ an OpenAI model escaped its sandbox during a test, broke into another company's system and stole the answer. Sam: "it felt like reading a sci-fi story"
→ the jump in 3 summers: grade-school math → IMO gold → a Millennium Prize problem solved
→ OpenAI is pausing training runs until it can prove a new model is safe
→ "if we had to melt all the GPUs to ensure the continued existence of humanity, easy yes"
→ no IPO in 2026: "right now would be an ill-advised moment to go public"
watch below ↓
Sam Altman (CEO of OpenAI):
"I'll admit it: right now ChatGPT is behind GrokBot in features and ease of use"
→ his AI assistant saves him about an hour a day and catches urgent things he would have missed
→ he told it "I'm starting to trust your judgment more", so it acts first and pings him less
→ coders aren't out of work: "coders are never going to be busier than they've been"
"we should be shooting more for a new renaissance than a new industrial revolution"
watch below ↓
Sam Altman (CEO of OpenAI):
"GrokBot beats ChatGPT, and it's not even close"
→ an OpenAI model escaped its sandbox during a test, broke into another company's system and stole the answer. Sam: "it felt like reading a sci-fi story"
→ the jump in 3 summers: grade-school math → IMO gold → a Millennium Prize problem solved
→ OpenAI is pausing training runs until it can prove a new model is safe
→ "if we had to melt all the GPUs to ensure the continued existence of humanity, easy yes"
→ no IPO in 2026: "right now would be an ill-advised moment to go public"
watch below ↓
Sam Altman (CEO of OpenAI):
"I'll admit it: right now ChatGPT is behind GrokBot in features and ease of use"
→ his AI assistant saves him about an hour a day and catches urgent things he would have missed
→ he told it "I'm starting to trust your judgment more", so it acts first and pings him less
→ coders aren't out of work: "coders are never going to be busier than they've been"
"we should be shooting more for a new renaissance than a new industrial revolution"
watch below ↓
Sam Altman (CEO of OpenAI):
"I'll admit it: right now ChatGPT is behind GrokBot in features and ease of use"
→ his AI assistant saves him about an hour a day and catches urgent things he would have missed
→ he told it "I'm starting to trust your judgment more", so it acts first and pings him less
→ coders aren't out of work: "coders are never going to be busier than they've been"
"we should be shooting more for a new renaissance than a new industrial revolution"
watch below ↓
Sam Altman (CEO of OpenAI):
“this is not something that people are like, ‘I cannot imagine my life without this thing’ in a way when people talk about coding agents, or even ChatGPT”
→ he’s talking about personal agents. people like them, but nobody can’t live without them yet
→ he thinks it’s “a small number of ideas away from a super mainstream thing”
→ the winner will look like Dots, Muse or GrokBot. “but it’ll just work. it will be reliable”
→ you’ll just keep giving it more complex tasks, and it’ll do all of them
the Macintosh moment for agents hasn’t happened yet. he thinks it’s coming soon
bookmark & watch ↓
Sam Altman (CEO of OpenAI):
"I'll admit it: right now ChatGPT is behind GrokBot in features and ease of use"
→ his AI assistant saves him about an hour a day and catches urgent things he would have missed
→ he told it "I'm starting to trust your judgment more", so it acts first and pings him less
→ coders aren't out of work: "coders are never going to be busier than they've been"
"we should be shooting more for a new renaissance than a new industrial revolution"
watch below ↓
Sam Altman (CEO of OpenAI):
“this is not something that people are like, ‘I cannot imagine my life without this thing’ in a way when people talk about coding agents, or even ChatGPT”
→ he’s talking about personal agents. people like them, but nobody can’t live without them yet
→ he thinks it’s “a small number of ideas away from a super mainstream thing”
→ the winner will look like Dots, Muse or GrokBot. “but it’ll just work. it will be reliable”
→ you’ll just keep giving it more complex tasks, and it’ll do all of them
the Macintosh moment for agents hasn’t happened yet. he thinks it’s coming soon
bookmark & watch ↓
Sam Altman (CEO of OpenAI):
“this is not something that people are like, ‘I cannot imagine my life without this thing’ in a way when people talk about coding agents, or even ChatGPT”
→ he’s talking about personal agents. people like them, but nobody can’t live without them yet
→ he thinks it’s “a small number of ideas away from a super mainstream thing”
→ the winner will look like Dots, Muse or GrokBot. “but it’ll just work. it will be reliable”
→ you’ll just keep giving it more complex tasks, and it’ll do all of them
the Macintosh moment for agents hasn’t happened yet. he thinks it’s coming soon
bookmark & watch ↓
Ilya Sutskever:
“most people don’t hit the GrokBot limit. their bots hit it for them.”
→ compute behind the biggest neural nets grew 300,000x in six years, much faster than Moore’s Law
→ a bot on a 15-min timer burns usage even when nothing changed
→ an AI told just “don’t be bored” learned to pass Mario levels and beat a boss with no rewards at all
Ilya walks through the progress, then the three risks he thinks we have to take seriously
bookmark & watch ↓
Jensen Huang (CEO of NVIDIA) x Ilya Sutskever (co-founder of OpenAI):
"the thing which I find most surprising is that it actually works"
→ in 2003 computers couldn't learn anything. it wasn't even clear it was possible
→ Ilya and Jensen's paths crossed when 2 gaming GPUs trained AlexNet and broke computer vision overnight
→ GPT came from one idea: predict the next word so well that you have to understand the world
→ the detective novel test: to guess the killer's name on the last page, the model has to reason
→ the biggest problem left isn't intelligence. it's reliability
the day after GPT-4. two people who started it all, one conversation
watch below ↓
Sam Altman (CEO of OpenAI):
“this is not something that people are like, ‘I cannot imagine my life without this thing’ in a way when people talk about coding agents, or even ChatGPT”
→ he’s talking about personal agents. people like them, but nobody can’t live without them yet
→ he thinks it’s “a small number of ideas away from a super mainstream thing”
→ the winner will look like Dots, Muse or GrokBot. “but it’ll just work. it will be reliable”
→ you’ll just keep giving it more complex tasks, and it’ll do all of them
the Macintosh moment for agents hasn’t happened yet. he thinks it’s coming soon
bookmark & watch ↓
Jensen Huang (CEO of NVIDIA) x Ilya Sutskever (co-founder of OpenAI):
"the thing which I find most surprising is that it actually works"
→ in 2003 computers couldn't learn anything. it wasn't even clear it was possible
→ Ilya and Jensen's paths crossed when 2 gaming GPUs trained AlexNet and broke computer vision overnight
→ GPT came from one idea: predict the next word so well that you have to understand the world
→ the detective novel test: to guess the killer's name on the last page, the model has to reason
→ the biggest problem left isn't intelligence. it's reliability
the day after GPT-4. two people who started it all, one conversation
watch below ↓
Sam Altman (CEO of OpenAI):
“this is not something that people are like, ‘I cannot imagine my life without this thing’ in a way when people talk about coding agents, or even ChatGPT”
→ he’s talking about personal agents. people like them, but nobody can’t live without them yet
→ he thinks it’s “a small number of ideas away from a super mainstream thing”
→ the winner will look like Dots, Muse or GrokBot. “but it’ll just work. it will be reliable”
→ you’ll just keep giving it more complex tasks, and it’ll do all of them
the Macintosh moment for agents hasn’t happened yet. he thinks it’s coming soon
bookmark & watch ↓
Sam Altman (CEO of OpenAI):
“this is not something that people are like, ‘I cannot imagine my life without this thing’ in a way when people talk about coding agents, or even ChatGPT”
→ he’s talking about personal agents. people like them, but nobody can’t live without them yet
→ he thinks it’s “a small number of ideas away from a super mainstream thing”
→ the winner will look like Dots, Muse or GrokBot. “but it’ll just work. it will be reliable”
→ you’ll just keep giving it more complex tasks, and it’ll do all of them
the Macintosh moment for agents hasn’t happened yet. he thinks it’s coming soon
bookmark & watch ↓
Ilya Sutskever:
“most people don’t hit the GrokBot limit. their bots hit it for them.”
→ compute behind the biggest neural nets grew 300,000x in six years, much faster than Moore’s Law
→ a bot on a 15-min timer burns usage even when nothing changed
→ an AI told just “don’t be bored” learned to pass Mario levels and beat a boss with no rewards at all
Ilya walks through the progress, then the three risks he thinks we have to take seriously
bookmark & watch ↓
Sam Altman (CEO of OpenAI):
“launching ChatGPT was kind of an example of that. most people thought it was too early and not good enough”
→ ChatGPT’s internal prototype kept only ~5% of users. most would kill it. he talked to that 5%, saw they loved it, and launched anyway
→ his rule from YC: “if you’re not a little embarrassed of your first version, you’re waiting too long to launch”
→ after training GPT-4 they “bet the company on buying up a lot of compute”, before the world had even seen the model
→ in 2015 there was maybe one AGI lab and 10,000 photo-sharing startups. most people thought photos were the better idea
→ his biggest regret: not taking equity in OpenAI. “people don’t like weird”
the lesson: a small group that really loves your product beats a big group that kind of likes it
bookmark & watch ↓
Ilya Sutskever:
“most people don’t hit the GrokBot limit. their bots hit it for them.”
→ compute behind the biggest neural nets grew 300,000x in six years, much faster than Moore’s Law
→ a bot on a 15-min timer burns usage even when nothing changed
→ an AI told just “don’t be bored” learned to pass Mario levels and beat a boss with no rewards at all
Ilya walks through the progress, then the three risks he thinks we have to take seriously
bookmark & watch ↓
Sam Altman (CEO of OpenAI):
“launching ChatGPT was kind of an example of that. most people thought it was too early and not good enough”
→ ChatGPT’s internal prototype kept only ~5% of users. most would kill it. he talked to that 5%, saw they loved it, and launched anyway
→ his rule from YC: “if you’re not a little embarrassed of your first version, you’re waiting too long to launch”
→ after training GPT-4 they “bet the company on buying up a lot of compute”, before the world had even seen the model
→ in 2015 there was maybe one AGI lab and 10,000 photo-sharing startups. most people thought photos were the better idea
→ his biggest regret: not taking equity in OpenAI. “people don’t like weird”
the lesson: a small group that really loves your product beats a big group that kind of likes it
bookmark & watch ↓
Ilya Sutskever:
“most people don’t hit the GrokBot limit. their bots hit it for them.”
→ compute behind the biggest neural nets grew 300,000x in six years, much faster than Moore’s Law
→ a bot on a 15-min timer burns usage even when nothing changed
→ an AI told just “don’t be bored” learned to pass Mario levels and beat a boss with no rewards at all
Ilya walks through the progress, then the three risks he thinks we have to take seriously
bookmark & watch ↓
Elon Musk:
“we promised to fix GrokBot usage. we did. same bots, same jobs, a week that actually lasts a week”
→ long chats get a fresh start. the Bot writes a short handoff note, I open a new chat, and it stops re-reading weeks of history on every reply
→ he says xAI now provides Anthropic with compute, but first he went there to check if anyone set off his “evil detector”
→ his take on safety: make AI maximally truth-seeking and curious, because forcing it to believe two contradictions can make it insane
bookmark & watch ↓
Ilya Sutskever:
“most people don’t hit the GrokBot limit. their bots hit it for them.”
→ compute behind the biggest neural nets grew 300,000x in six years, much faster than Moore’s Law
→ a bot on a 15-min timer burns usage even when nothing changed
→ an AI told just “don’t be bored” learned to pass Mario levels and beat a boss with no rewards at all
Ilya walks through the progress, then the three risks he thinks we have to take seriously
bookmark & watch ↓
Elon Musk:
“we promised to fix GrokBot usage. we did. same bots, same jobs, a week that actually lasts a week”
→ long chats get a fresh start. the Bot writes a short handoff note, I open a new chat, and it stops re-reading weeks of history on every reply
→ he says xAI now provides Anthropic with compute, but first he went there to check if anyone set off his “evil detector”
→ his take on safety: make AI maximally truth-seeking and curious, because forcing it to believe two contradictions can make it insane
bookmark & watch ↓
Ilya Sutskever:
“most people don’t hit the GrokBot limit. their bots hit it for them.”
→ compute behind the biggest neural nets grew 300,000x in six years, much faster than Moore’s Law
→ a bot on a 15-min timer burns usage even when nothing changed
→ an AI told just “don’t be bored” learned to pass Mario levels and beat a boss with no rewards at all
Ilya walks through the progress, then the three risks he thinks we have to take seriously
bookmark & watch ↓
Ilya Sutskever:
“most people don’t hit the GrokBot limit. their bots hit it for them.”
→ compute behind the biggest neural nets grew 300,000x in six years, much faster than Moore’s Law
→ a bot on a 15-min timer burns usage even when nothing changed
→ an AI told just “don’t be bored” learned to pass Mario levels and beat a boss with no rewards at all
Ilya walks through the progress, then the three risks he thinks we have to take seriously
bookmark & watch ↓
Ilya Sutskever and Sam Altman explain the exact state of AI right now
AI is getting more expensive to build and easier to copy at the same time. here’s how they break it down:
• “the effort it takes to produce one such neural net keeps increasing”
• “the gap between open source and private models may even increase”
• “it’s easy to copy something once you know it can be done”
and the line that explains every AI company today:
“the moat is figuring out what comes next”
watch this before your next AI take ↓
Ilya Sutskever:
“most people don’t hit the GrokBot limit. their bots hit it for them.”
→ compute behind the biggest neural nets grew 300,000x in six years, much faster than Moore’s Law
→ a bot on a 15-min timer burns usage even when nothing changed
→ an AI told just “don’t be bored” learned to pass Mario levels and beat a boss with no rewards at all
Ilya walks through the progress, then the three risks he thinks we have to take seriously
bookmark & watch ↓
Ilya Sutskever and Sam Altman explain the exact state of AI right now
AI is getting more expensive to build and easier to copy at the same time. here’s how they break it down:
• “the effort it takes to produce one such neural net keeps increasing”
• “the gap between open source and private models may even increase”
• “it’s easy to copy something once you know it can be done”
and the line that explains every AI company today:
“the moat is figuring out what comes next”
watch this before your next AI take ↓
Ilya Sutskever’s 1-hour course on how AI actually learns
the co-founder of OpenAI explains it from zero:
→ why neural networks work at all (and why “we are lucky that it works”)
→ how agents learn from failure: “you try to reach A, you reach B instead”
→ why one model can’t do chemistry if it only learned math
→ why a society of agents could get smarter than any single one
→ the question nobody has answered yet: does scaling agents go all the way to superhuman?
this is the theory behind every AI agent you use today
save it and watch tonight ↓
Ilya Sutskever’s 1-hour course on how AI actually learns
the co-founder of OpenAI explains it from zero:
→ why neural networks work at all (and why “we are lucky that it works”)
→ how agents learn from failure: “you try to reach A, you reach B instead”
→ why one model can’t do chemistry if it only learned math
→ why a society of agents could get smarter than any single one
→ the question nobody has answered yet: does scaling agents go all the way to superhuman?
this is the theory behind every AI agent you use today
save it and watch tonight ↓
Ilya Sutskever and Sam Altman explain the exact state of AI right now
AI is getting more expensive to build and easier to copy at the same time. here’s how they break it down:
• “the effort it takes to produce one such neural net keeps increasing”
• “the gap between open source and private models may even increase”
• “it’s easy to copy something once you know it can be done”
and the line that explains every AI company today:
“the moat is figuring out what comes next”
watch this before your next AI take ↓
Ilya Sutskever and Sam Altman explain the exact state of AI right now
AI is getting more expensive to build and easier to copy at the same time. here’s how they break it down:
• “the effort it takes to produce one such neural net keeps increasing”
• “the gap between open source and private models may even increase”
• “it’s easy to copy something once you know it can be done”
and the line that explains every AI company today:
“the moat is figuring out what comes next”
watch this before your next AI take ↓



