@johnnynomad.bsky.social
You‘re paragraphing it wrong
December 23, 2025 at 4:45 AM
You‘re paragraphing it wrong
A 20% chance of developing AGI within 5 years still sounds delusional. No one has any actual clue how to get there yet. And even if they somehow had all the breakthroughs needed, they would still have to prove that the whole thing actually scales.
December 21, 2025 at 4:34 PM
A 20% chance of developing AGI within 5 years still sounds delusional. No one has any actual clue how to get there yet. And even if they somehow had all the breakthroughs needed, they would still have to prove that the whole thing actually scales.
Railroads and AI, that's got some poetry. Now we only need to figure out how to get tulips into this game
December 21, 2025 at 11:42 AM
Railroads and AI, that's got some poetry. Now we only need to figure out how to get tulips into this game
All you need is one right Alibaba bet in 2000 apparently...
December 21, 2025 at 11:38 AM
All you need is one right Alibaba bet in 2000 apparently...
Yep, this is one of the factors that will the crash the AI bubble sooner or later.
Running these open source models still needs powerful hardware, so there is some datacenter revenue. And currently the closed-source models are a bit more powerful. But i believe this gap will close.
Running these open source models still needs powerful hardware, so there is some datacenter revenue. And currently the closed-source models are a bit more powerful. But i believe this gap will close.
December 20, 2025 at 1:41 PM
Yep, this is one of the factors that will the crash the AI bubble sooner or later.
Running these open source models still needs powerful hardware, so there is some datacenter revenue. And currently the closed-source models are a bit more powerful. But i believe this gap will close.
Running these open source models still needs powerful hardware, so there is some datacenter revenue. And currently the closed-source models are a bit more powerful. But i believe this gap will close.
How do you interpret this? Does this mean the $22.5B from Softbank likely won't go through by the end of the year, and now OpenAI is scrambling to get more cash?
December 18, 2025 at 5:56 AM
How do you interpret this? Does this mean the $22.5B from Softbank likely won't go through by the end of the year, and now OpenAI is scrambling to get more cash?
Then the AI boosters will say dark compute means data centers in space
December 12, 2025 at 3:37 PM
Then the AI boosters will say dark compute means data centers in space
Maybe they mean 220 million people as in 220 million US taxpayers as in bailout lol
November 26, 2025 at 3:19 AM
Maybe they mean 220 million people as in 220 million US taxpayers as in bailout lol
They should sell NFTs for those GPUs. Finally a good use for that amazing invention
November 24, 2025 at 4:33 AM
They should sell NFTs for those GPUs. Finally a good use for that amazing invention
He meant he meant once they go public he can watch people get destroyed on their short positions. Sadly right now there is no way to short OpenAI
November 20, 2025 at 5:00 PM
He meant he meant once they go public he can watch people get destroyed on their short positions. Sadly right now there is no way to short OpenAI
They obviously couldn‘t just patch it on top of LLMs, this would require a whole new architecture
November 20, 2025 at 7:33 AM
They obviously couldn‘t just patch it on top of LLMs, this would require a whole new architecture
Dogs trying to become physicists is what LLMs are. The whole idea of world models is to create a model that allows reasoning and true understanding. And how to do that is big open question.
November 20, 2025 at 7:32 AM
Dogs trying to become physicists is what LLMs are. The whole idea of world models is to create a model that allows reasoning and true understanding. And how to do that is big open question.
Much bigger complexity… like how many variables does the real world have, and more importantly how many rules (cause and effect) does the world have. So you will need a very smart architecture. If it‘s even possble to replicate what the human brain does
November 20, 2025 at 5:41 AM
Much bigger complexity… like how many variables does the real world have, and more importantly how many rules (cause and effect) does the world have. So you will need a very smart architecture. If it‘s even possble to replicate what the human brain does
There seem to be several definitions. The narrow „world model“ would be modeling spatial relationships and physics. But they real deal would be a model that knows cause and effect. A model that describes how the whole world actually works and that allows planning and reasoning, like us humans do.
November 20, 2025 at 5:34 AM
There seem to be several definitions. The narrow „world model“ would be modeling spatial relationships and physics. But they real deal would be a model that knows cause and effect. A model that describes how the whole world actually works and that allows planning and reasoning, like us humans do.
If they succede (i hope they won‘t) it would be HUGE. This would create AI with actual understanding and solve all the problems LLMs have, like hallucinations and the inability to come up with any novel thoughts. Which means then they could actually build their agents™️ and make everyone unemployed
November 20, 2025 at 5:25 AM
If they succede (i hope they won‘t) it would be HUGE. This would create AI with actual understanding and solve all the problems LLMs have, like hallucinations and the inability to come up with any novel thoughts. Which means then they could actually build their agents™️ and make everyone unemployed
I don‘t know much about LeCun, but i have the feeling he‘s an idealist and actually wants to create AGI, or superinteligence or whatever the word of the month is
November 20, 2025 at 5:21 AM
I don‘t know much about LeCun, but i have the feeling he‘s an idealist and actually wants to create AGI, or superinteligence or whatever the word of the month is
The huge difference is that a world model would mean actual understanding. LLMs don‘t understand anything, they‘re just matching patterns from their training data
November 20, 2025 at 5:17 AM
The huge difference is that a world model would mean actual understanding. LLMs don‘t understand anything, they‘re just matching patterns from their training data
That‘s the narrow definition of world model. What they are probably trying to build is an AI that actually understands the world (not only physics) - as opposed to LLMs
November 20, 2025 at 5:14 AM
That‘s the narrow definition of world model. What they are probably trying to build is an AI that actually understands the world (not only physics) - as opposed to LLMs
What they‘re doing is probably fundamental research. From what I know, no one has any clue how to actually build world models yet. From LeCun‘s perspective it makes sense, since at Meta it‘s all in for LLMs and they‘re obviously a dead end.
November 20, 2025 at 5:11 AM
What they‘re doing is probably fundamental research. From what I know, no one has any clue how to actually build world models yet. From LeCun‘s perspective it makes sense, since at Meta it‘s all in for LLMs and they‘re obviously a dead end.
An image of the auditor just leaked
November 20, 2025 at 4:46 AM
An image of the auditor just leaked
Google definitely is using their in-house chips. They trained and ran all of their Gemini generations completely on TPUs
November 20, 2025 at 2:21 AM
Google definitely is using their in-house chips. They trained and ran all of their Gemini generations completely on TPUs
And Gemini 1
November 19, 2025 at 5:42 PM
And Gemini 1
This is not big news, Google already did the same for Gemini 2
November 19, 2025 at 5:37 PM
This is not big news, Google already did the same for Gemini 2
Why are you posting all of this under a comment for Ed Zitron‘s blog? Did you read any of his articles? Of course there is an AI bubble. And yes it will be like 2000. It‘s not even a debate anymore, the only big, million-dollar question is WHEN will it pop
November 18, 2025 at 4:41 AM
Why are you posting all of this under a comment for Ed Zitron‘s blog? Did you read any of his articles? Of course there is an AI bubble. And yes it will be like 2000. It‘s not even a debate anymore, the only big, million-dollar question is WHEN will it pop
Softbank actually *had* to buy more OpenAI stock, because they had promised the money ($22.5B) in an investment round earlier. They also had to sell off a bunch of other stuff to free up that amount.
November 18, 2025 at 3:27 AM
Softbank actually *had* to buy more OpenAI stock, because they had promised the money ($22.5B) in an investment round earlier. They also had to sell off a bunch of other stuff to free up that amount.