5 ms·

Deepseek's advancements are not going away, even if you ban them. Demand for NVDA chips will go down, and likely will take a long time before Jevon's Paradox k

Deepseek's advancements are not going away, even if you ban them. Demand for NVDA chips will go down, and likely will take a long time before Jevon's Paradox kicks in.

aurareturn 2y ago
>Demand for NVDA chips will go down, and likely will take a long time before Jevon's Paradox kicks in. reply I expect Jevons Paradox effect to start immediately. Small businesses and enterprises are likely wanting to order DGX systems for internal inference. OpenAI and larger AI labs will want to put a bigger distance between themselves and Chinese AI models by using compute muscle and advantage. All AI labs see the same efficiency benefit so they will have to continue to compete on compute capacity.
- zippyman55 2y ago
  Last week the moat was the large size or cost to generate a model. That has gone away. So there should be lots of competition now within the US from smaller companies.
  - aurareturn 2y ago
    New AI Lab: Trains a model using Deepseek's techniques on 2,000 GPUs xAI/OpenAI/Anthropic/Google: Trains a model using Deepseek's techniques on 100,000 GPUs I fail to see how this makes smaller companies competitive.
    - GianFabien 2y ago
      Depends on whether smaller companies can deliver sufficient results cheaper than the larger ones. There are some indications that suggest that there are diminishing returns on investing on ever more power. It's like you don't need a 1000HP supercar to get around town, a 55HP sedan is fine for most folks.
      - poobear22 2y ago
        Yep, that is my point. If the large scale LLMs are not sufficiently better than the new crop of startups, I suspect the large firms will need to acquire the startups (that would be their response due to a lack of a moat). Its hard to buy everything and to know where to place your bets.
        aurareturn 2y ago
        So you’re worried that LLMs have stopped scaling even though the biggest breakthrough from DeepSeek is scaling RF learning without humans?
        poobear22 2y ago
        As I see it: I'm waiting to see improvements in LLM performance. What I see is an improvement in computational efficiency (less hardware needed) If general LLMs do not show continued performance improvement, then there is a lot of excess HW that needs to be utilized somehow. If LLMs continue to show performance improvement, then the hardware can be used more efficiently.
        aurareturn 2y ago
        So how does Deepseek change anything for your view? What you wrote was true before Deepseeek and their non-human RF breakthrough.
        poobear22 2y ago
        Here are my thoughts, I am a little removed from some of these fields so you will not hurt my feelings if you want to be blunt. Ok, “What I wrote was true before Deepseek and their non-human reinforcement learning Breakthrough”. Right, I did make a general statement that could be applied pre- and post-Deepseek. I think I get your point. But, you are stating: that “So you’re worried that LLMs have stopped scaling even though the biggest breakthrough from DeepSeek is scaling RF learning without humans?” I’m not worried about it, but I am waiting to see continued LLM performance improvement due to HW scaling as opposed to algorithmic improvements. This seems to happen in industries, like in weather forecasting, model is developed, sucks up all resources, new model is developed, HPC company goes bankrupt, new super computer purchased, new model rolled out, new performance gains, sucks up all computer resources, rinse and repeat. But it takes a long time to release a new upgrade of a model. Now, regarding algorithmic improvements, the way I think about it is to break it down into two areas: a) throughput/efficiency improvement (faster/more_efficient) and b) performance improvement (better predictions). I’m not following this as close as I would like, but it seems Deepseek is more aligned with faster/more_efficient. My statement I wrote, I think what has changed is the realization that, as applied to LLM, investors have been neglecting the efficiency side of the problem. It seems many of the promises of performance around the corner are more investment/funding driven. The other option, of focusing on efficiency would leave more stakeholders with less wealth, so they have a nature bias to promote this technology. So, with the Deepseek approach/revelation, my thoughts are: Before Deepseek announcement: Established/large LLM producers: Resources: More hardware, data centers, power distribution, tax credits. Challenges: making a profit, establishing a true moat (but not publicly recognized) Moat: Significant cost challenges to new market entrants limited who could compete. New_entrant/small LLM producers: Resources: License large-scale LLMs, tune them for bespoke requests. Moat: Significant cost challenges to new market entrants limited who could compete. HW producers NVIDIA is dominant with CUDA and established HW products After Deepseek announcement: Established LLM producers: Called into question the excess resources: More hardware, data centers, power distribution, tax credits. Need to recognize and acquire new entrants that may pose a down-stream challenge. This will be tough as many companies will spring up. New_entrant/small LLM producers: Resources: New companies should sprout up as barrier to entry is reduced. Moat: Small vendors can specialize, or build business to be acquired by large LLM vendor. HW producers NVIDIA is dominant with CUDA and established HW products Increased GPU demand and high margins of established HW players will bring in new market entrants (competition). Numerous small LLM producers will increase HW sales. Margins are high with NVIDIA, so other HW suppliers will see this as an opportunity to enter the market. Also, since NVIDIA margins are so high, large-scale LLM vendors will welcome competition.
  - Jlagreen 2y ago
    That was never the moat. Nvidia's moat is controlling compute for 80-90% of workloads. Why do you think Steam has more RTX 4090 than 4080 in their survey? Ever since RTX 2080TI you could buy multi server consumer GPUs for ML. We sell PCs with RTX cards regularly to customer for small local AI applications. Project Digits is the next thing. It is not only a Nvidia GPU in some PC, it's the real AI PC. The real deal and we're already considering switchting to that as a system instead of a PC since it's size is perfect for our Vision application. Do you think Nvidia cares if you buy 1 Blackwell DC GPU, 13x Digits or 20x RTX 5090? In the end it's all the same turnover for Nvidia. Nvidia's goal is to spread and dominate workloads worldwide and that no matter if DC, enterprise or consumer, Nvidia HW is used.
    - poobear22 2y ago
      I re-read the thread topic and it is in response to NVIDIA stock. My response was in response to LLM generators and their moat (not NVIDIA's moat). So, my point was the moat for LLM generators had diminished. Additional competition generating LLMs from new entrants may increase demand for HW. The HW may be used more efficiently but I am still waiting to see if LLM performance continues to improve.
      - Jlagreen 2y ago
        Ah, I'm sorry you're right a misunderstood your comment. I agree and Nvidia positions itself for exactly that. See how fast DeepSeek will come to NIM. People are already wondering how well DeepSeek will run on Digits. Nvidia also offers distilled open models or specific own open models so is indirectly competing in that space as well. But Nvidia isn't in the LLM generator business but in the business of "infrastructure for LLM generators" Everyone is waiting for GPT5 or another big bang. And because it takes so much people start to think that there is a wall. And there is a wall but that wall could be also compute. Blackwell will show if there is a compute wall because simply put, if a training run with large parameter set on GPT5 takes like 4 months then Blackwell might reduce that to under 1 month with the same amount of GPUs. Getting more GPUs can get that down even more. Imagine the speed up in AI frontier model research if your training times come down 4-5x from new GPU generation and another 2x from getting twice as many. The nice part with Nvidia is also that the old GPUs don't become obsolete, OpenAI can continue using them for inferencing or even try to use combined architecture training as long as they don't go FP4. I wouldn't be surprised that at the end of 2025 we will see things which will make DeepSeek and GPT4 as oldschool stuff simply because of the massive compute which Blackwell will deliver this year.

Deepseek's advancements are not going away, even if you ban them. Demand for NVDA chips will go down, and likely will take a long time before Jevon's Paradox kicks in.

aurareturn 2y ago

>Demand for NVDA chips will go down, and likely will take a long time before Jevon's Paradox kicks in. reply I expect Jevons Paradox effect to start immediately. Small businesses and enterprises are likely wanting to order DGX systems for internal inference. OpenAI and larger AI labs will want to put a bigger distance between themselves and Chinese AI models by using compute muscle and advantage. All AI labs see the same efficiency benefit so they will have to continue to compete on compute capacity.

zippyman55 2y ago
Last week the moat was the large size or cost to generate a model. That has gone away. So there should be lots of competition now within the US from smaller companies.
- aurareturn 2y ago
  New AI Lab: Trains a model using Deepseek's techniques on 2,000 GPUs xAI/OpenAI/Anthropic/Google: Trains a model using Deepseek's techniques on 100,000 GPUs I fail to see how this makes smaller companies competitive.
  - GianFabien 2y ago
    Depends on whether smaller companies can deliver sufficient results cheaper than the larger ones. There are some indications that suggest that there are diminishing returns on investing on ever more power. It's like you don't need a 1000HP supercar to get around town, a 55HP sedan is fine for most folks.
    - poobear22 2y ago
      Yep, that is my point. If the large scale LLMs are not sufficiently better than the new crop of startups, I suspect the large firms will need to acquire the startups (that would be their response due to a lack of a moat). Its hard to buy everything and to know where to place your bets.
      - aurareturn 2y ago
        So you’re worried that LLMs have stopped scaling even though the biggest breakthrough from DeepSeek is scaling RF learning without humans?
        poobear22 2y ago
        As I see it: I'm waiting to see improvements in LLM performance. What I see is an improvement in computational efficiency (less hardware needed) If general LLMs do not show continued performance improvement, then there is a lot of excess HW that needs to be utilized somehow. If LLMs continue to show performance improvement, then the hardware can be used more efficiently.
        aurareturn 2y ago
        So how does Deepseek change anything for your view? What you wrote was true before Deepseeek and their non-human RF breakthrough.
        poobear22 2y ago
        Here are my thoughts, I am a little removed from some of these fields so you will not hurt my feelings if you want to be blunt. Ok, “What I wrote was true before Deepseek and their non-human reinforcement learning Breakthrough”. Right, I did make a general statement that could be applied pre- and post-Deepseek. I think I get your point. But, you are stating: that “So you’re worried that LLMs have stopped scaling even though the biggest breakthrough from DeepSeek is scaling RF learning without humans?” I’m not worried about it, but I am waiting to see continued LLM performance improvement due to HW scaling as opposed to algorithmic improvements. This seems to happen in industries, like in weather forecasting, model is developed, sucks up all resources, new model is developed, HPC company goes bankrupt, new super computer purchased, new model rolled out, new performance gains, sucks up all computer resources, rinse and repeat. But it takes a long time to release a new upgrade of a model. Now, regarding algorithmic improvements, the way I think about it is to break it down into two areas: a) throughput/efficiency improvement (faster/more_efficient) and b) performance improvement (better predictions). I’m not following this as close as I would like, but it seems Deepseek is more aligned with faster/more_efficient. My statement I wrote, I think what has changed is the realization that, as applied to LLM, investors have been neglecting the efficiency side of the problem. It seems many of the promises of performance around the corner are more investment/funding driven. The other option, of focusing on efficiency would leave more stakeholders with less wealth, so they have a nature bias to promote this technology. So, with the Deepseek approach/revelation, my thoughts are: Before Deepseek announcement: Established/large LLM producers: Resources: More hardware, data centers, power distribution, tax credits. Challenges: making a profit, establishing a true moat (but not publicly recognized) Moat: Significant cost challenges to new market entrants limited who could compete. New_entrant/small LLM producers: Resources: License large-scale LLMs, tune them for bespoke requests. Moat: Significant cost challenges to new market entrants limited who could compete. HW producers NVIDIA is dominant with CUDA and established HW products After Deepseek announcement: Established LLM producers: Called into question the excess resources: More hardware, data centers, power distribution, tax credits. Need to recognize and acquire new entrants that may pose a down-stream challenge. This will be tough as many companies will spring up. New_entrant/small LLM producers: Resources: New companies should sprout up as barrier to entry is reduced. Moat: Small vendors can specialize, or build business to be acquired by large LLM vendor. HW producers NVIDIA is dominant with CUDA and established HW products Increased GPU demand and high margins of established HW players will bring in new market entrants (competition). Numerous small LLM producers will increase HW sales. Margins are high with NVIDIA, so other HW suppliers will see this as an opportunity to enter the market. Also, since NVIDIA margins are so high, large-scale LLM vendors will welcome competition.
- Jlagreen 2y ago
  That was never the moat. Nvidia's moat is controlling compute for 80-90% of workloads. Why do you think Steam has more RTX 4090 than 4080 in their survey? Ever since RTX 2080TI you could buy multi server consumer GPUs for ML. We sell PCs with RTX cards regularly to customer for small local AI applications. Project Digits is the next thing. It is not only a Nvidia GPU in some PC, it's the real AI PC. The real deal and we're already considering switchting to that as a system instead of a PC since it's size is perfect for our Vision application. Do you think Nvidia cares if you buy 1 Blackwell DC GPU, 13x Digits or 20x RTX 5090? In the end it's all the same turnover for Nvidia. Nvidia's goal is to spread and dominate workloads worldwide and that no matter if DC, enterprise or consumer, Nvidia HW is used.
  - poobear22 2y ago
    I re-read the thread topic and it is in response to NVIDIA stock. My response was in response to LLM generators and their moat (not NVIDIA's moat). So, my point was the moat for LLM generators had diminished. Additional competition generating LLMs from new entrants may increase demand for HW. The HW may be used more efficiently but I am still waiting to see if LLM performance continues to improve.
    - Jlagreen 2y ago
      Ah, I'm sorry you're right a misunderstood your comment. I agree and Nvidia positions itself for exactly that. See how fast DeepSeek will come to NIM. People are already wondering how well DeepSeek will run on Digits. Nvidia also offers distilled open models or specific own open models so is indirectly competing in that space as well. But Nvidia isn't in the LLM generator business but in the business of "infrastructure for LLM generators" Everyone is waiting for GPT5 or another big bang. And because it takes so much people start to think that there is a wall. And there is a wall but that wall could be also compute. Blackwell will show if there is a compute wall because simply put, if a training run with large parameter set on GPT5 takes like 4 months then Blackwell might reduce that to under 1 month with the same amount of GPUs. Getting more GPUs can get that down even more. Imagine the speed up in AI frontier model research if your training times come down 4-5x from new GPU generation and another 2x from getting twice as many. The nice part with Nvidia is also that the old GPUs don't become obsolete, OpenAI can continue using them for inferencing or even try to use combined architecture training as long as they don't go FP4. I wouldn't be surprised that at the end of 2025 we will see things which will make DeepSeek and GPT4 as oldschool stuff simply because of the massive compute which Blackwell will deliver this year.