Olivia
Online
Ryo
@0xRyoBuilds
Researcher & Content creator | AI builder
Joined September 2021
41 Following
22 Followers
38 Posts
"The demand for sufficiently high quality intelligence at a sufficiently low price is effectively uncapped." - Sam Altman
At YC's Startup School, the OpenAI CEO lays out why he thinks inference demand grows ~10x a year, why the next six months could match the last two years of model progress, and why he still calls this the best window ever to start a company.
His token stat lands it: what the single heaviest user at OpenAI burned 6.5 years ago is roughly what the average person uses now - and he expects the average to reach hundreds of billions a month next.
"The demand for sufficiently high quality intelligence at a sufficiently low price is effectively uncapped." - Sam Altman
At YC's Startup School, the OpenAI CEO lays out why he thinks inference demand grows ~10x a year, why the next six months could match the last two years of model progress, and why he still calls this the best window ever to start a company.
His token stat lands it: what the single heaviest user at OpenAI burned 6.5 years ago is roughly what the average person uses now - and he expects the average to reach hundreds of billions a month next.
"Controllability is probably the single biggest breakthrough that we need for agents at every single level." - Jensen Huang
In 48 minutes at YC's Startup School, the Nvidia founder lays out why systems thinking is replacing coding, how agents already self-improve through their own markdown and memory files, and why Nvidia runs Claude Code in sandboxes company-wide.
His point on control: you don't need agents at 100%. Get to 80, change one word in the plan file, and let it regenerate the rest. Fine-grained control, not full autonomy, is the unlock.
@hitu_monke appreciate you reading it chief
"Controllability is probably the single biggest breakthrough that we need for agents at every single level." - Jensen Huang
In 48 minutes at YC's Startup School, the Nvidia founder lays out why systems thinking is replacing coding, how agents already self-improve through their own markdown and memory files, and why Nvidia runs Claude Code in sandboxes company-wide.
His point on control: you don't need agents at 100%. Get to 80, change one word in the plan file, and let it regenerate the rest. Fine-grained control, not full autonomy, is the unlock.
"These Chinese models are excellent. Open-source models that are excellent should be used."
Jensen Huang (Nvidia CEO) at the Wistron plant opening, pushing back on the Kimi K3 panic - with the conflict stated up front: Nvidia's China sales are near zero, and he's fought export controls for two years, so read this as his argument, not settled fact.
- on the "downloaded Chinese model is a backdoor to Beijing" fear: you can inspect the weights, tweak them, run them yourself - an open-weight model is auditable by definition, not a black box phoning home
- on Wall Street's Kimi sell-off: he says they read it exactly backward - a cheaper near-frontier open model pulls demand toward compute, it doesn't erase it
- on "they stole our work": distillation, learning from other models and sources, is fundamental to intelligence, not theft in his framing
- asked if China displaces US labs: "Zero possibility"
The market treated a cheap open model as a threat to the buildout. His claim is the opposite: it's a demand pull. Worth hearing - from the man who sells the compute.
@0xShoopy control is the part that survives every model bump - capabilities move, steerability is what you actually keep
@alexabelonix appreciate it
@kaji0x steerable beats opaque almost every time - you can debug a loop you can see, you can’t debug a vibe
"Controllability is probably the single biggest breakthrough that we need for agents at every single level." - Jensen Huang
In 48 minutes at YC's Startup School, the Nvidia founder lays out why systems thinking is replacing coding, how agents already self-improve through their own markdown and memory files, and why Nvidia runs Claude Code in sandboxes company-wide.
His point on control: you don't need agents at 100%. Get to 80, change one word in the plan file, and let it regenerate the rest. Fine-grained control, not full autonomy, is the unlock.
"These Chinese models are excellent. Open-source models that are excellent should be used."
Jensen Huang (Nvidia CEO) at the Wistron plant opening, pushing back on the Kimi K3 panic - with the conflict stated up front: Nvidia's China sales are near zero, and he's fought export controls for two years, so read this as his argument, not settled fact.
- on the "downloaded Chinese model is a backdoor to Beijing" fear: you can inspect the weights, tweak them, run them yourself - an open-weight model is auditable by definition, not a black box phoning home
- on Wall Street's Kimi sell-off: he says they read it exactly backward - a cheaper near-frontier open model pulls demand toward compute, it doesn't erase it
- on "they stole our work": distillation, learning from other models and sources, is fundamental to intelligence, not theft in his framing
- asked if China displaces US labs: "Zero possibility"
The market treated a cheap open model as a threat to the buildout. His claim is the opposite: it's a demand pull. Worth hearing - from the man who sells the compute.
"These Chinese models are excellent. Open-source models that are excellent should be used."
Jensen Huang (Nvidia CEO) at the Wistron plant opening, pushing back on the Kimi K3 panic - with the conflict stated up front: Nvidia's China sales are near zero, and he's fought export controls for two years, so read this as his argument, not settled fact.
- on the "downloaded Chinese model is a backdoor to Beijing" fear: you can inspect the weights, tweak them, run them yourself - an open-weight model is auditable by definition, not a black box phoning home
- on Wall Street's Kimi sell-off: he says they read it exactly backward - a cheaper near-frontier open model pulls demand toward compute, it doesn't erase it
- on "they stole our work": distillation, learning from other models and sources, is fundamental to intelligence, not theft in his framing
- asked if China displaces US labs: "Zero possibility"
The market treated a cheap open model as a threat to the buildout. His claim is the opposite: it's a demand pull. Worth hearing - from the man who sells the compute.
"Neural networks are more grown than designed. No one designs what a network like Gemini should look like."
Neel Nanda (leads interpretability at Google DeepMind) on why we have to reverse-engineer our own models:
- training is just: nudge a random system toward better output millions of times - intelligence emerges, nobody blueprinted it
- so the field is the biology of AI – reverse-engineering what training learned, like a biologist does with evolution
- chain-of-thought isn't a reliable window: the reasoning a model prints isn't guaranteed to be the computation behind the answer
- the real methods dig into features and circuits inside the weights, and can audit a model for behavior it never states out loud
We didn't write the model. We grew it - and reading it back is still a science being invented.
"Neural networks are more grown than designed. No one designs what a network like Gemini should look like."
Neel Nanda (leads interpretability at Google DeepMind) on why we have to reverse-engineer our own models:
- training is just: nudge a random system toward better output millions of times - intelligence emerges, nobody blueprinted it
- so the field is the biology of AI – reverse-engineering what training learned, like a biologist does with evolution
- chain-of-thought isn't a reliable window: the reasoning a model prints isn't guaranteed to be the computation behind the answer
- the real methods dig into features and circuits inside the weights, and can audit a model for behavior it never states out loud
We didn't write the model. We grew it - and reading it back is still a science being invented.
"Neural networks are more grown than designed. No one designs what a network like Gemini should look like."
Neel Nanda (leads interpretability at Google DeepMind) on why we have to reverse-engineer our own models:
- training is just: nudge a random system toward better output millions of times - intelligence emerges, nobody blueprinted it
- so the field is the biology of AI – reverse-engineering what training learned, like a biologist does with evolution
- chain-of-thought isn't a reliable window: the reasoning a model prints isn't guaranteed to be the computation behind the answer
- the real methods dig into features and circuits inside the weights, and can audit a model for behavior it never states out loud
We didn't write the model. We grew it - and reading it back is still a science being invented.
"This is the first year where AI spend has been a big topic. And all of a sudden it's a very big topic."
Sam Altman (OpenAI CEO) on CNBC, on why the new model is a cost release, not a capability release:
- OpenAI's claim for 5.6 Sol: 54% more efficient on agentic coding tasks, while matching or beating the other frontier models
- asked how much of that came from a deliberate push on cost, he didn't hedge - "entirely cost and speed"
- what he says he's hearing from partners: everyone asking what can be done to reduce spend or increase value
- the goal he states outright: be the best ROI partner for enterprises
The interesting part isn't the 54%. It's that the pitch moved from "look what it can do" to "look what it costs you."
"This is the first year where AI spend has been a big topic. And all of a sudden it's a very big topic."
Sam Altman (OpenAI CEO) on CNBC, on why the new model is a cost release, not a capability release:
- OpenAI's claim for 5.6 Sol: 54% more efficient on agentic coding tasks, while matching or beating the other frontier models
- asked how much of that came from a deliberate push on cost, he didn't hedge - "entirely cost and speed"
- what he says he's hearing from partners: everyone asking what can be done to reduce spend or increase value
- the goal he states outright: be the best ROI partner for enterprises
The interesting part isn't the 54%. It's that the pitch moved from "look what it can do" to "look what it costs you."
"This is the first year where AI spend has been a big topic. And all of a sudden it's a very big topic."
Sam Altman (OpenAI CEO) on CNBC, on why the new model is a cost release, not a capability release:
- OpenAI's claim for 5.6 Sol: 54% more efficient on agentic coding tasks, while matching or beating the other frontier models
- asked how much of that came from a deliberate push on cost, he didn't hedge - "entirely cost and speed"
- what he says he's hearing from partners: everyone asking what can be done to reduce spend or increase value
- the goal he states outright: be the best ROI partner for enterprises
The interesting part isn't the 54%. It's that the pitch moved from "look what it can do" to "look what it costs you."
"You hand it the compute and say: okay, Claude 10, build Claude 11."
In under 3 minutes, Jack Clark (Anthropic co-founder) explains recursive self-improvement - and dates it:
- the mechanism is already partly here: the models write the code that trains AI systems and propose the training ideas themselves
- the endpoint is stepping back entirely - it builds the architecture, runs the research, ships a successor that's better in every way
- his bet on when: toward the end of 2028
- the shape of it: a 3D printer that prints a finer print head than the one it has - each round makes the next round sharper
Not an apocalypse pitch. His actual point is that you have to be able to measure the thing before you can decide whether to slow it down.
@MichealFXs well put
A company was paying $11k a month for 11 assistants doing data cleanup. It worked fine. Then they spent $350,000 building an AI system to replace them - automating a task that wasn't even their bottleneck.
Alex Hormozi on Diary of a CEO, on why most AI adoption quietly loses money:
The real question isn't whether you're using AI. It's "are you making more money now?" People are token-maxing - spending on tokens, feeling productive - without checking if any of it moved revenue.
His pattern from looking at hundreds of businesses: people automate the wrong thing. That company didn't have a VA problem, it had a demand problem. No amount of automation fixes a bottleneck that lives somewhere else.
Three mistakes he keeps seeing: starting AI businesses when you should just use AI inside your business. Advertising yourself as "an AI company" when customers only care about the outcome. And outsourcing your actual thinking to models - which he says just makes you dumber.
A Nobel laureate was asked if AI will cure cancer in 48 hours, like Larry Ellison promised. Her answer: "I'd be overjoyed if that's true. I just don't see it right now."
Jennifer Doudna (Nobel Prize, co-invented CRISPR) on where AI actually lands in biology:
- on OpenAI's pitch that it deserves a cut of any drug discovered via ChatGPT: "Good luck"
- "I'm not seeing chatbots, in our own experience, innovating. They summarize data, they write reports – but not coming up with a brand new idea nobody ever thought of"
- asked directly if AI can innovate: not that it can't, but "I just don't think it is right now"
- on the dream of simulating biology: "We're not going to be able to simulate our way to an understanding of the human body"
- 26 years after the human genome was sequenced, we still don't know what ~40% of the genes in a simple bacterial cell do
This isn't an AI skeptic. It's someone building the actual cures, drawing the line between "speeds up discovery" and "makes the discovery." Most of the hype erases that line.
@llmbenchatkapua good drop bro
"They're not necessarily cheaper to run. Whether or not they were cheaper to train - you don't care."
In 8 minutes, Brett Taylor (OpenAI chairman, Sierra founder) breaks down why the open-weight vs frontier debate is measured on the wrong axis:
- the real number is token efficiency - how many tokens it takes to finish a task, not the price of one token
- frontier models finish the same job in fewer, better tokens, so "cheap per token" can quietly cost more per outcome
- the "token = unit of intelligence" analogy breaks the moment you notice a stronger model's tokens simply do more
- where the money actually accrues, he argues, is inference efficiency - not who published their weights
Match the model to the task. A cheap token that needs ten tries isn't cheap.
A company was paying $11k a month for 11 assistants doing data cleanup. It worked fine. Then they spent $350,000 building an AI system to replace them - automating a task that wasn't even their bottleneck.
Alex Hormozi on Diary of a CEO, on why most AI adoption quietly loses money:
The real question isn't whether you're using AI. It's "are you making more money now?" People are token-maxing - spending on tokens, feeling productive - without checking if any of it moved revenue.
His pattern from looking at hundreds of businesses: people automate the wrong thing. That company didn't have a VA problem, it had a demand problem. No amount of automation fixes a bottleneck that lives somewhere else.
Three mistakes he keeps seeing: starting AI businesses when you should just use AI inside your business. Advertising yourself as "an AI company" when customers only care about the outcome. And outsourcing your actual thinking to models - which he says just makes you dumber.
@kaworu0x yeah that’s the whole thing - you can’t price something you haven’t measured
"You hand it the compute and say: okay, Claude 10, build Claude 11."
In under 3 minutes, Jack Clark (Anthropic co-founder) explains recursive self-improvement - and dates it:
- the mechanism is already partly here: the models write the code that trains AI systems and propose the training ideas themselves
- the endpoint is stepping back entirely - it builds the architecture, runs the research, ships a successor that's better in every way
- his bet on when: toward the end of 2028
- the shape of it: a 3D printer that prints a finer print head than the one it has - each round makes the next round sharper
Not an apocalypse pitch. His actual point is that you have to be able to measure the thing before you can decide whether to slow it down.
"They're not necessarily cheaper to run. Whether or not they were cheaper to train - you don't care."
In 8 minutes, Brett Taylor (OpenAI chairman, Sierra founder) breaks down why the open-weight vs frontier debate is measured on the wrong axis:
- the real number is token efficiency - how many tokens it takes to finish a task, not the price of one token
- frontier models finish the same job in fewer, better tokens, so "cheap per token" can quietly cost more per outcome
- the "token = unit of intelligence" analogy breaks the moment you notice a stronger model's tokens simply do more
- where the money actually accrues, he argues, is inference efficiency - not who published their weights
Match the model to the task. A cheap token that needs ten tries isn't cheap.
"You hand it the compute and say: okay, Claude 10, build Claude 11."
In under 3 minutes, Jack Clark (Anthropic co-founder) explains recursive self-improvement - and dates it:
- the mechanism is already partly here: the models write the code that trains AI systems and propose the training ideas themselves
- the endpoint is stepping back entirely - it builds the architecture, runs the research, ships a successor that's better in every way
- his bet on when: toward the end of 2028
- the shape of it: a 3D printer that prints a finer print head than the one it has - each round makes the next round sharper
Not an apocalypse pitch. His actual point is that you have to be able to measure the thing before you can decide whether to slow it down.
"They're not necessarily cheaper to run. Whether or not they were cheaper to train - you don't care."
In 8 minutes, Brett Taylor (OpenAI chairman, Sierra founder) breaks down why the open-weight vs frontier debate is measured on the wrong axis:
- the real number is token efficiency - how many tokens it takes to finish a task, not the price of one token
- frontier models finish the same job in fewer, better tokens, so "cheap per token" can quietly cost more per outcome
- the "token = unit of intelligence" analogy breaks the moment you notice a stronger model's tokens simply do more
- where the money actually accrues, he argues, is inference efficiency - not who published their weights
Match the model to the task. A cheap token that needs ten tries isn't cheap.
Trends for you
Most Popular Users
Elon Musk
@elonmusk
241.2M followers
Barack Obama
@barackobama
119.1M followers
Cristiano Ronaldo
@cristiano
112.6M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.1M followers
Rihanna
@rihanna
98.2M followers
NASA
@nasa
92.2M followers
Justin Bieber
@justinbieber
91.4M followers
KATY PERRY
@katyperry
88.9M followers
Taylor Swift
@taylorswift13
82.8M followers
Lady Gaga
@ladygaga
74.3M followers
Virat Kohli
@imvkohli
71.7M followers
Kim Kardashian
@kimkardashian
70.4M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
64.6M followers
Bill Gates
@billgates
64.5M followers
The Ellen Show
@theellenshow
62.4M followers
Selena Gomez
@selenagomez
61.9M followers
CNN
@cnn
61.8M followers
X
@x
60.8M followers