Am I the only one excited for the release but not overanalyzing their words? This thread feels full of personal interpretations. DeepSeek is still a business—great release, but expectations and motivations seem inflated.
Probably it's because there's nothing specific here to discuss. In the absence of specific new information, discussions turn generic [1] and that tends to make for shallow/indignant discussion. That's one reason why an announcement of announcement (like "Starting next week, we'll open-source 5 repos") is off topic on HN [2].
The releases themselves may turn out to be interesting, of course, and then there may be something substantive to have a thread about. The best submission would be to pick the most interesting release once it shows up.
The "launch week" pattern isn't great for HN, because we end up with a bunch of follow-ups that we have to downweight [3], and there's no guarantee that the largest thread(s) will be about the most interesting element(s) in the sequence. But startups do it anyway so we'll adapt.
Most interested to see their inference stack, hope that’s one of the 5. I think most people are running R1 on a single H200 node but Deepseek had much lower RAM per GPU for their inference and so had some cluster based MoE deployment.
Correct. There are 3 main ways to "gimp" high end GPUs meant for training - "cores", "on-chip memory speed" and "interconnects". IIUC the H800 had the first 2 unchanged but halved the interconnect speeds.
H20 is the next iteration of the "sanctions" that I believe also limited the "cores" but left the on-chip memory intact, or slightly higher (from the new generation).
> We believe DeepSeek has access to around 10,000 of these H800s and about 10,000 H100s. Furthermore they have orders for many more H20’s, with Nvidia having produced over 1 million of the China specific GPU in the last 9 months.
> Starting next week, we'll open-source 5 repos – one daily drop
Probably counts as announcement of announcement? Let’s wait for the actual repo drops before discussing them, especially because there are no details about what will be open sourced other than
> These are humble building blocks of our online service: documented, deployed and battle-tested in production.
You are right for sure saying to wait for the actual repos.
But on the other hand, compare this announcement in a README.md file in a GitHub repo with this slideware approach of EU https://openeurollm.eu/
If I had to bet on someone providing some value, unfortunately I wouldn't bet on Europe.
I'm saying this as a European, deeply convinced that Europe is a good place to live. I've also worked for a couple of EU funded research projects, so I have some background experience on the outcome of these projects.
You’re not wrong, it’s a hell lot more exciting to watch players organically emerging from a competitive landscape with stuff you can put your hands on today (or next week) than players hand-picked and tasked by governments, making hollow announcements before they have anything interesting to show.
I think before "drop" in electronic music was a widely used term, "dropping a new track" (ie releasing new music) was a common hip-hop term, since forever.
FWIW the expression comes from delivery by airplane and parachute. Probably started to be used UPS-style delivery drivers and/or drug dealers, and spread from there. Now it just means "deliver".
Emotionally I agree, but... o1 was a paradigm shift. Nothing DeepSeek has done is on that level yet. DeepSeek themselves would agree. Supposedly Liang Wenfeng himself flew to US to gather information when o1 was launched.
Maybe in terms of advancing scientific knowledge but DeepSeek has achieved a paradigm shift back from opex to capex.
Certain applications are now economically viable when you don't have to pay per request and don't have to fight NVIDIA/sanctions for the privilege
Deep respect for DeepSeek and what they've done regarding all the innovations and researches they have been putting out in-the-open.
"Because every line shared becomes collective momentum that accelerates the journey. Daily unlocks begin soon. No ivory towers - just pure garage-energy and community-driven innovation" is a great phase.
I mostly agree with you. Google has a good strategy of driving down costs, for example. I am amazed by the large number of API providers who host either the original DeepSeek R1 or a distilled version.
When cost approaches zero, use cases increase exponentially.
> Most likely, without any intention on their part.
I think this is a very, very naive assumption.
The founder is a quant with involvements in domestic investments and market design and pricing for decades - in China.
As seen with the case of Jack Ma, after you cross a certain level, there is no such thing as "not involved with politics" in China.
Liang knows exactly what he's doing.
> During 2021, Liang started buying thousands of Nvidia GPUs for his AI side project while running High-Flyer. Some industry insiders viewed it as the eccentric actions of a billionaire looking for a new hobby. One of Liang's business partners said they initially did not take Liang seriously and described their first meeting as seeing a very nerdy guy with a terrible hairstyle who could not articulate his vision. Liang simply said he wanted to build something and it will be a game changer which his business partners thought was only possible from giants such as ByteDance and Alibaba Group.
> During that month in an interview with 36Kr, Liang stated that High-Flyer had acquired 10,000 Nvidia A100 GPUs before the US government imposed AI chip restrictions on China.
> On 20 January 2025, Liang was invited to the Symposium with Experts, Entrepreneurs and Representatives from the Fields of Education, Science, Culture, Health and Sports (专家、企业家和教科文卫体等领域代表座谈会) hosted by Premier Li Qiang in Beijing. Liang, being considered as an industry expert, was asked to provide opinions and suggestions on a draft for comments of the annual 2024 government work report.
> On 17 February 2025, Liang along with the heads of other Chinese technology companies attended a symposium hosted by President Xi Jinping at the Great Hall of the People in Beijing.
Whether he intended to or not initially, what happens with DeepSeek is now out of this man's hand and will be 100% influenced by politics.
The chip bans and dual use nature of the technology have catapulted Liang to the first row of CCP tech strategists' attention, for sure.
I know that Americans are the saints and everyone else is evil… everyone else being gentiles and all. But other than that, is there any other point you are trying to make?
I am not sure what you mean by AI bubble. Do you mean the valuation of some companies? Or course some won't do well in the future. In the meanty, a significant part of the population uses on it to accelerate their tasks (be it admin work, legal question, learning, getting inspiration). There is no way back. It feels like saying the video streaming bubble will burst in 2020. No. It is too valuable. But yes, some player will die. Nothing special here. IMHO.
A bubble bursting does not mean the industry in the bubble ceases to exist. It means the market hype dies down and only the things that have actual value survive. When it comes to AI, realistically most of the hype is fluff, so calling it a bubble is fair.
I mean the whole world still uses the Internet after the dot-com bubble burst. A significant amount of “AI companies” are valued with revenue multipliers never used before. 44x in the case of OpenAI for example. I agree there is no going back, but this bubble will burst, and hard. IMHO.
Kinda interesting to see where the moat is in AI space. Good base models can always distilled when you have access to API. System prompts can get leaked, and UI tricks can be copied. In the end, the moat might be in the hardware and vertical integration.
> the moat might be in the hardware and vertical integration.
The moat is the products that can be built. The moat is always the product - because a differentiated product can't be a commodity. And an LLM is not a product.
Google and MSFT and Meta have already "won" because they have profitable products they can build LLMs onto. Every other company seems to be burning cash to build a product, and only ChatGPT is getting the brand recognition to realistically compete.
Building an LLM is like building a database. Sure a good one unlocks new uses, but consumers aren't buying something for the database. Meanwhile enterprise customers will shop around and drive the price of a commodity down while open source alternatives grow from in-house uses to destroy moats.
Even hardware isn't a true moat. Only Google has strong vertical integration with their TPUs, and that gives them a lead. BUT Microsoft, AWS, Meta and a whole bunch of startups are building out custom silicon which will surely put pressure on them and Nvidia to keep innovating and earning that price edge.
See I kind of buy the database argument but also kind of don't. A database needs an operator whereas a LLM doesn't. You're basically melting the product into a piece of goo and the UI can be approached using natural language.
For products that still need a UI you could claim that LLM operators take over, so that's still a tax you pay to the incumbents as you interact with a product. It's sort of like we take the money which was paid to SQL operators and engineers and instead pay it to the hyperscalers.
How many times have we been down this path? Tcp/IP, dos/windows, Linux, virtualization, and on and on. Open platforms always seem to find a way to usurp everyone else. In the end, it's better to be a service provider.
You can use the outputs of a closed source model (or deepseek -> llama. see llama 70b deepseek distilled) to create a synthetic training data set which lets you fine tune (distill) most of the benefits of the "smarter" model in to a "dumber" model. This is why openAi does not show the actual full chain of thought but a summarized version. To stop exfiltration of their IP which has proven immensely difficult.*
This is great to see! Open-sourcing infrastructure tools can really accelerate innovation in the AI space. I've found that having access to well-documented repos makes it much easier to experiment and build on existing work. Are there any specific areas these repos focus on, like distributed training or model serving?
Readit News
The releases themselves may turn out to be interesting, of course, and then there may be something substantive to have a thread about. The best submission would be to pick the most interesting release once it shows up.
The "launch week" pattern isn't great for HN, because we end up with a bunch of follow-ups that we have to downweight [3], and there's no guarantee that the largest thread(s) will be about the most interesting element(s) in the sequence. But startups do it anyway so we'll adapt.
[1] https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
[2] https://hn.algolia.com/?dateRange=all&page=0&prefix=true&sor...
[3] https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
Most interested to see their inference stack, hope that’s one of the 5. I think most people are running R1 on a single H200 node but Deepseek had much lower RAM per GPU for their inference and so had some cluster based MoE deployment.
But yeah, would be interesting to see how they optimized for that.
H20 is the next iteration of the "sanctions" that I believe also limited the "cores" but left the on-chip memory intact, or slightly higher (from the new generation).
Dead Comment
> We believe DeepSeek has access to around 10,000 of these H800s and about 10,000 H100s. Furthermore they have orders for many more H20’s, with Nvidia having produced over 1 million of the China specific GPU in the last 9 months.
Probably counts as announcement of announcement? Let’s wait for the actual repo drops before discussing them, especially because there are no details about what will be open sourced other than
> These are humble building blocks of our online service: documented, deployed and battle-tested in production.
But on the other hand, compare this announcement in a README.md file in a GitHub repo with this slideware approach of EU https://openeurollm.eu/
If I had to bet on someone providing some value, unfortunately I wouldn't bet on Europe.
I'm saying this as a European, deeply convinced that Europe is a good place to live. I've also worked for a couple of EU funded research projects, so I have some background experience on the outcome of these projects.
Deleted Comment
Deleted Comment
"Because every line shared becomes collective momentum that accelerates the journey. Daily unlocks begin soon. No ivory towers - just pure garage-energy and community-driven innovation" is a great phase.
Dead Comment
LLMs have been more legitimate "blockchain" when most CIO magazines had these essays with "What's your blockchain strategy?" kind of stuffed material.
AI bubble will burst and will burst hard. By end of 2026 at max.
When cost approaches zero, use cases increase exponentially.
I think this is a very, very naive assumption.
The founder is a quant with involvements in domestic investments and market design and pricing for decades - in China.
As seen with the case of Jack Ma, after you cross a certain level, there is no such thing as "not involved with politics" in China.
Liang knows exactly what he's doing.
> During 2021, Liang started buying thousands of Nvidia GPUs for his AI side project while running High-Flyer. Some industry insiders viewed it as the eccentric actions of a billionaire looking for a new hobby. One of Liang's business partners said they initially did not take Liang seriously and described their first meeting as seeing a very nerdy guy with a terrible hairstyle who could not articulate his vision. Liang simply said he wanted to build something and it will be a game changer which his business partners thought was only possible from giants such as ByteDance and Alibaba Group.
> During that month in an interview with 36Kr, Liang stated that High-Flyer had acquired 10,000 Nvidia A100 GPUs before the US government imposed AI chip restrictions on China.
> On 20 January 2025, Liang was invited to the Symposium with Experts, Entrepreneurs and Representatives from the Fields of Education, Science, Culture, Health and Sports (专家、企业家和教科文卫体等领域代表座谈会) hosted by Premier Li Qiang in Beijing. Liang, being considered as an industry expert, was asked to provide opinions and suggestions on a draft for comments of the annual 2024 government work report.
> On 17 February 2025, Liang along with the heads of other Chinese technology companies attended a symposium hosted by President Xi Jinping at the Great Hall of the People in Beijing.
Whether he intended to or not initially, what happens with DeepSeek is now out of this man's hand and will be 100% influenced by politics.
The chip bans and dual use nature of the technology have catapulted Liang to the first row of CCP tech strategists' attention, for sure.
Source: https://en.wikipedia.org/wiki/Liang_Wenfeng
The moat is the products that can be built. The moat is always the product - because a differentiated product can't be a commodity. And an LLM is not a product.
Google and MSFT and Meta have already "won" because they have profitable products they can build LLMs onto. Every other company seems to be burning cash to build a product, and only ChatGPT is getting the brand recognition to realistically compete.
Building an LLM is like building a database. Sure a good one unlocks new uses, but consumers aren't buying something for the database. Meanwhile enterprise customers will shop around and drive the price of a commodity down while open source alternatives grow from in-house uses to destroy moats.
Even hardware isn't a true moat. Only Google has strong vertical integration with their TPUs, and that gives them a lead. BUT Microsoft, AWS, Meta and a whole bunch of startups are building out custom silicon which will surely put pressure on them and Nvidia to keep innovating and earning that price edge.
For products that still need a UI you could claim that LLM operators take over, so that's still a tax you pay to the incumbents as you interact with a product. It's sort of like we take the money which was paid to SQL operators and engineers and instead pay it to the hyperscalers.
Good enough + open (and free) is a very appealing proposition.
Where we're going, we don't need moats.
What does that mean?
*disclaimer; i am an expert of nothing
that's why they can open source their model and be fine because running this shit is actually hard, let alone maintaining SLA for millions of users??