I think the open source insight that was very explicitly called out by DeepSeek's CEO is part of a bigger story about organisational cultural transformation, that's also relevant to *every* organisation.
Indeed, over a year ago I wrote how AI would trigger a new, unconventional talent strategy:
"'Giving away’ solutions to systemic problems is an amazing way to attract and retain exceptional talent...
AI is radically reducing the ‘cost' of innovation – empowering small, brilliant and hyper-productive teams to move faster and reach further. DeepMind has less than 1/100th of the number of employees that Google has, yet its leveraging AI to transform multiple scientific fields at an astonishing pace...
It won’t just be ‘AI companies’ that experience this exponential empowerment.
AI is a general purpose technology. What’s happening today within the AI industry will happen tomorrow in your industry."
In your interview with Kai fu Lee, I recall one of his optimizations was a technical/hardware one, rather than a software algorithm one. It had to do with saving processed GPU tokens into the CPU memory, which can be recalled when needed without the need of being reprocessed. As the GPU's that Deepseek has access to (H100 and H800), these don't come with a CPU. I suspected they have an optimized CPU with a large cache for this very purpose. For reasoning in which variations of similar questions are being iterated, this might prove a useful feature for ergonomic reasons. It would be up to Nvidia to cover this hole.
A tangential observation. As it happens, I'm in Beijing for a few weeks. Last Wednesday a friend suggested I download Doubao, an app and a platform from ByteDance. It's UI/UX is remarkable. I've been impressed with Baidu's translation for years. [In 2020 I checked into a Chinese hotel easier than into a London hotel because Baidu's translation of Chinese is easier for my imperfect ears to understand than a London accent.]
Doubao's app is a game changer. I'm aware that Huawei's 5G is a lot faster than western 5g. Still Doubao's translation speed is astounding and the usability is profoundly higher than Baidu's.Two exponential technologies intersecting.
Powerful inference becomes important when it drives applications new applications. Unless the US manages to outlaw this technology, hard since is is so open, I think we have a next generation app platform. It's not limited to China because: open source.
I'd love to see some European companies jump into the fray.
The point about using a collective intelligence architecture thanks to cheaper models is huge - epistemically but also for novelty etc. having different models “collide” their ideas is the equivalent of having brain groups interact - it generates dissonance and resolution thereof. this is another step toward building supermind structures. I would feel better when those models don’t just have China values embedded in there. I assume that will happen. We might want a variety of models in there - including European and other values.
I think the open source insight that was very explicitly called out by DeepSeek's CEO is part of a bigger story about organisational cultural transformation, that's also relevant to *every* organisation.
Indeed, over a year ago I wrote how AI would trigger a new, unconventional talent strategy:
"'Giving away’ solutions to systemic problems is an amazing way to attract and retain exceptional talent...
AI is radically reducing the ‘cost' of innovation – empowering small, brilliant and hyper-productive teams to move faster and reach further. DeepMind has less than 1/100th of the number of employees that Google has, yet its leveraging AI to transform multiple scientific fields at an astonishing pace...
It won’t just be ‘AI companies’ that experience this exponential empowerment.
AI is a general purpose technology. What’s happening today within the AI industry will happen tomorrow in your industry."
Read more: https://thefuturenormal.substack.com/p/designing-a-people-first-ai-strategy-488
Nice observation
Great post Azeem...Pmarca could have at least credited YOUR Sputnik analogy!!??
In the midst of geopolitical reorientation, perhaps an important question, as the USA goes all imperial, is "who you gonna trust?"
run it locally
then. it is open weights
In your interview with Kai fu Lee, I recall one of his optimizations was a technical/hardware one, rather than a software algorithm one. It had to do with saving processed GPU tokens into the CPU memory, which can be recalled when needed without the need of being reprocessed. As the GPU's that Deepseek has access to (H100 and H800), these don't come with a CPU. I suspected they have an optimized CPU with a large cache for this very purpose. For reasoning in which variations of similar questions are being iterated, this might prove a useful feature for ergonomic reasons. It would be up to Nvidia to cover this hole.
A tangential observation. As it happens, I'm in Beijing for a few weeks. Last Wednesday a friend suggested I download Doubao, an app and a platform from ByteDance. It's UI/UX is remarkable. I've been impressed with Baidu's translation for years. [In 2020 I checked into a Chinese hotel easier than into a London hotel because Baidu's translation of Chinese is easier for my imperfect ears to understand than a London accent.]
Doubao's app is a game changer. I'm aware that Huawei's 5G is a lot faster than western 5g. Still Doubao's translation speed is astounding and the usability is profoundly higher than Baidu's.Two exponential technologies intersecting.
Powerful inference becomes important when it drives applications new applications. Unless the US manages to outlaw this technology, hard since is is so open, I think we have a next generation app platform. It's not limited to China because: open source.
I'd love to see some European companies jump into the fray.
The point about using a collective intelligence architecture thanks to cheaper models is huge - epistemically but also for novelty etc. having different models “collide” their ideas is the equivalent of having brain groups interact - it generates dissonance and resolution thereof. this is another step toward building supermind structures. I would feel better when those models don’t just have China values embedded in there. I assume that will happen. We might want a variety of models in there - including European and other values.
Wow, breathtaking!! Let’s enjoy the (wild) ride!
That's some serious DeepSeek usage already! When do you decide to use it vs Claude?
for structure style questions.
o1 Pro is the Chidi Anagonye ofAI. sometimes i don't have time to wait