US apparently has more resource and funds; what is concerning that a number of Chinese companies (with very well trained engineers) figured that they can train frontier model with a small fraction of GPUs and money, needed for OpenAI, meta, google, etc. In a way, it is super encouraging that Chinese companies showed that little guys can compete with giants. Also, it could mean that we did not understand AI that well, we may be able to train AI models on our mobile phones in the future. Finally, maybe "talent > money/GPUs" to make be successful in AI race? If that is the case, here in US, we probably need to brew or import talents like crazy right now ...
@ from strawberry to o1 was about 16 months o1 to r1 was about 3 months. By a team that mainly works on investing and they spent less money training it than the salary of an OpenAI executive
@ o1 is still more powerful than r1, and o3 is days to weeks away. That’s likely a lie about this being a side project for those guys. They are also doing open source so they are able to build directly on top of the research of other teams. Based on the number of folks on the paper that is a 10m dollar team in annual cost, the number of gpus used was something like $500m in gpus and $50m in power cost. Now if their government is subsidizing all of that and it only cost the team $5m, that isn’t the true cost of what it cost to build and train that model.
@@cyberft It's not going to be as much of a leap. o1 was different because it moved towards inference-time training. o3 will most likely do inference-time better by generating higher quality and quantity of synthetic data.
It really makes me relax from seeing just the huge amount of indians (Elite Human Capitals 😄) involved in the american AI industry. The chinese century is certainly upon us 🇨🇳🙏
Currently, AI has not yet reached the level of Agents, so profit margins are very limited. However, once AI evolves into AGI-level Agents, the situation will change dramatically. For example, using O1 to write code can boost my work efficiency by over 100 times. I can create complete projects based solely on conceptual frameworks, although I still need to iterate to fix bugs. This is something previous models like GPT-4 couldn't achieve; they could only assist with coding. Agents, however, are on a completely different level. You no longer need to manually copy and paste; all corrections are done automatically. The key is that Agents can work 24/7 without getting tired. As long as the quality is slightly higher than the current standard, human programmers will definitely be at risk of losing their jobs. From my work experience, many small experiments or verification programs are difficult to implement because they're time-consuming and only boost efficiency by about 10%. But with AI, these concerns will no longer be an issue. Note that this phenomenon won't just affect programmers-it will impact all industries.
This fella has no clue how most things in AI engineering work and seems to work with inherent bias against Chinese engineering. Strange given he publishes good stuff in semics.
This is a racist rant if ever I seen one. Dylan Patel saying oh look Chinese companies claiming they are "Singaporean" now just because they moved to Singapore. They are Singaporean because they were incorporated in Singapore, with International funding. With that reasoning, people can say that Dylan Patel is "claiming" to be American just because he was assembled in the US, but with Indian-origin genetic materials, but hey, he should still be considered as Indian.
Prolly wanna redo this again tmr haha
great show - presenter is excellent at letting the guest speak and asking thoughtful questions where needed.
thank you!
Please consider sharing with others that might get value from the show - we're always looking to increase our reach :)
@@RedpointAIThis age is like milk.😂. In just 3 months China caught up to o1. R1 Equal Ability and more Cheaper
thank you based deepseek
i had to check my playback speed setting
Sniffle sniffle
Dylan is the best 👑👑👑😄
agreed
I saw this dude on another podcast with Asiannometry, great stuff
He was Omega cracked
Great Host, Wonderful guest.. informed trailblazer..Buy Nvidia..Buy Bloom Energy
This aged like milk 😂
R1 wiped the floor with these predictions
He already mentioned R1 in this video, in what part of the video did he say something false about R1?
@@default2826thank you. Nothing wrong said here.
they may raise north of 100 billion by the end of 2027. ... They just got given 500 billion lol
Great conversation. Please convert this into a podcast playlist on RUclips. Makes it possible to listen on RUclips Music. Thanks a ton!!
US apparently has more resource and funds; what is concerning that a number of Chinese companies (with very well trained engineers) figured that they can train frontier model with a small fraction of GPUs and money, needed for OpenAI, meta, google, etc. In a way, it is super encouraging that Chinese companies showed that little guys can compete with giants. Also, it could mean that we did not understand AI that well, we may be able to train AI models on our mobile phones in the future. Finally, maybe "talent > money/GPUs" to make be successful in AI race? If that is the case, here in US, we probably need to brew or import talents like crazy right now ...
what does he mean when he says RAG is overhyped but Retrieval is underhyped? what papers is he talking about?
Wow! The Great Oracle of ai!
46:00 R1 already caught up to o1 and o3 isnt even out yet tho
Ok, but o3 is supposedly exponentially better than o1, and o4 is already in training.
@ from strawberry to o1 was about 16 months o1 to r1 was about 3 months. By a team that mainly works on investing and they spent less money training it than the salary of an OpenAI executive
@ o1 is still more powerful than r1, and o3 is days to weeks away. That’s likely a lie about this being a side project for those guys. They are also doing open source so they are able to build directly on top of the research of other teams. Based on the number of folks on the paper that is a 10m dollar team in annual cost, the number of gpus used was something like $500m in gpus and $50m in power cost. Now if their government is subsidizing all of that and it only cost the team $5m, that isn’t the true cost of what it cost to build and train that model.
@@cyberft It's not going to be as much of a leap. o1 was different because it moved towards inference-time training. o3 will most likely do inference-time better by generating higher quality and quantity of synthetic data.
R1 didn't catch up neither with o1-pro nor with o3. But nice try.
wondering about the cool retrieval papers that came out lately he mentions, any idea what those might be?
lol R1, new world
The USSR was first to get a man into space... The AI race has only just started
@@thingexplainer China isn't USSR
E.g. USSR collapsed while China ×50
Bruh USA's top ai researchers are from China. What are on about.
@@thingexplainerAnd also bad analogy.
USA v. USSR space race =
USA v. PRC hyper Sonic jet arms race.
'cause china beat USA to a hypersonic jet
@@thingexplainer USA'S top AI researchers are of Chinese origin.
How is it any different than Huawei or someone else releasing an iPhone knockoff? Chinese just innovate on cost, not technology
It really makes me relax from seeing just the huge amount of indians (Elite Human Capitals 😄) involved in the american AI industry. The chinese century is certainly upon us 🇨🇳🙏
Currently, AI has not yet reached the level of Agents, so profit margins are very limited. However, once AI evolves into AGI-level Agents, the situation will change dramatically. For example, using O1 to write code can boost my work efficiency by over 100 times. I can create complete projects based solely on conceptual frameworks, although I still need to iterate to fix bugs. This is something previous models like GPT-4 couldn't achieve; they could only assist with coding.
Agents, however, are on a completely different level. You no longer need to manually copy and paste; all corrections are done automatically. The key is that Agents can work 24/7 without getting tired. As long as the quality is slightly higher than the current standard, human programmers will definitely be at risk of losing their jobs.
From my work experience, many small experiments or verification programs are difficult to implement because they're time-consuming and only boost efficiency by about 10%. But with AI, these concerns will no longer be an issue. Note that this phenomenon won't just affect programmers-it will impact all industries.
Wow Dylan Patel is on point, following that many things and having such deep insights is remarkable, interviewer is also good.
total "like" word count: 1,316
It's insane
Dylan was correct! OpenAI raised $500 billion in early 2025!
🌍forever live 🌍forever live
40:00 aged like milk😂😂, they got 500B
this is already outdated lol
This age is like milk.😂. In just 3 months China caught up to o1. R1 Equal Ability and more Cheaper
It's close but it's not equal, check out how poorly it did on simplebench.
Dylan Patel is the Vivek Ramaswamy of tech.
DeepSeek, sorry I’m late to the party
matrix intro ,,, sneaky ;)
Like, like, like... like how come this pal is considering himself a whatever EXPERT? 😮
This fella has no clue how most things in AI engineering work and seems to work with inherent bias against Chinese engineering. Strange given he publishes good stuff in semics.
This is a racist rant if ever I seen one. Dylan Patel saying oh look Chinese companies claiming they are "Singaporean" now just because they moved to Singapore. They are Singaporean because they were incorporated in Singapore, with International funding. With that reasoning, people can say that Dylan Patel is "claiming" to be American just because he was assembled in the US, but with Indian-origin genetic materials, but hey, he should still be considered as Indian.
dylan being wrong on pretty much a lot :D
He’s busy shilling nvda all day and all
Night. I am waiting for his mi300x part2 for inference. When ? Next year ?
Wtf is this a tanned Kyle Kulinski?
this aged like rotten milk
India tech bro will say anything to defend their US motherland and their super idol Elon Musk 😂😂
Dylan is a born and raised American
;-) experts and toilet papers