3 ms·

Because the difference between a model that costs 10 million to train and a model that costs 10 billion to train is 6 months. Deepseek R1 is something that you

Because the difference between a model that costs 10 million to train and a model that costs 10 billion to train is 6 months. Deepseek R1 is something that you can run in a garage on hardware that the average software engineer can buy with a months salary and when it came out last month it was better than _every_ other model.

cedws 2y ago
What about third world programmers? They can’t necessarily afford a $5000 GPU. If it weren’t for the “generosity” of tech companies like GitHub granting free LLM usage, they might be locked out entirely. This would put them at a disadvantage, we can argue to what degree but it’s still a disadvantage. Depending on the curve we’re on, LLMs may grow more resource hungry while becoming closer to human performance in software engineering tasks. It’s not unimaginable this would concentrate productivity in the upper class of software engineers that can afford the hardware and/or licenses.
- llm_trw 2y ago
  $5000 monthly salary? My heart goes out to all the oppressed programmers in the EU.
  - ben_w 2y ago
    EU isn't "third world". Kenya is, and at current exchange rates the range in this citation is $558-$1876/month: https://www.glassdoor.com/Salaries/kenya-software-engineer-salary-SRCH_IL.0,5_IN130_KO6,23.htm?countryRedirect=true https://www.glassdoor.com/Salaries/kenya-software-engineer-s...
    - llm_trw 2y ago
      With salaries like that it's not first world either. If op is trying to catch up to frontier models locally on a budget 1/5th of what you can get in the west then I can see why she would feel the way she does about Ai.
      - ben_w 2y ago
        What are you even talking about? You said a month's salary; cedws said "third world" devs can't afford $5000 — and this is correct, third world devs can't afford that. cedws did not say EU, they said third world, at least not here. You said EU, not them. When you reply to me with "With salaries like that it's not first world either", who is "they"? The country I said explicitly was third world? Because that's a tautology.
- Fnoord 2y ago
  Deepseek R1 performs well on a 600 EUR Jetson, and a 700 EUR AMD GPU. Both bought during COVID crisis. It is that quick. However, don't ask it about certain sensitive topics. You can bet your ass Musk is using his AI tools as propaganda tool for his advancement, just like he does with X. We already seen prompt leak of Grok, it wasn't neutral.
  - mrtesthah 2y ago
    This is exactly correct. Grok is already slandering specific journalistic outlets such as “The Information” that publish negative stories on Musk.
mwigdahl 2y ago
No, it was not. It was not better than o1, nor o1-pro. Yes, it was _cheaper_ than those models, and superior in price/performance if the performance was acceptable. But in terms of raw performance it was behind them.
- johnfn 2y ago
  Chatbot arena leaderboard[1] says otherwise. 4o is ahead now, but that version of 4o was released after r1. r1 was ahead of all version of 4o and o1 at the time of release. https://huggingface.co/spaces/lmarena-ai/chatbot-arena-leaderboard https://huggingface.co/spaces/lmarena-ai/chatbot-arena-leade...
  - mwigdahl 2y ago
    Chatbot arena leaderboard is a good test for vibes and style of response, but not much else. R1's performance in objective benchmarks (coding, etc.) showed very good performance, granted, but inferior to the full o1 and o1-pro models. It's still a very impressive feat, but it wasn't frontier-pushing.
_heimdall 2y ago
Is it your expectation that as models get cheaper we won't be developing much more powerful models at the higher price range? Its already worth noting that we already ran into the self hosting at scale problem. People don't want to run a web server and instead accept all the problems that come along with the convenience of social media. Why would LLMs, or any future AI product, be different?

cedws 2y ago

What about third world programmers? They can’t necessarily afford a $5000 GPU. If it weren’t for the “generosity” of tech companies like GitHub granting free LLM usage, they might be locked out entirely. This would put them at a disadvantage, we can argue to what degree but it’s still a disadvantage. Depending on the curve we’re on, LLMs may grow more resource hungry while becoming closer to human performance in software engineering tasks. It’s not unimaginable this would concentrate productivity in the upper class of software engineers that can afford the hardware and/or licenses.

llm_trw 2y ago
$5000 monthly salary? My heart goes out to all the oppressed programmers in the EU.
- ben_w 2y ago
  EU isn't "third world". Kenya is, and at current exchange rates the range in this citation is $558-$1876/month: https://www.glassdoor.com/Salaries/kenya-software-engineer-salary-SRCH_IL.0,5_IN130_KO6,23.htm?countryRedirect=true https://www.glassdoor.com/Salaries/kenya-software-engineer-s...
  - llm_trw 2y ago
    With salaries like that it's not first world either. If op is trying to catch up to frontier models locally on a budget 1/5th of what you can get in the west then I can see why she would feel the way she does about Ai.
    - ben_w 2y ago
      What are you even talking about? You said a month's salary; cedws said "third world" devs can't afford $5000 — and this is correct, third world devs can't afford that. cedws did not say EU, they said third world, at least not here. You said EU, not them. When you reply to me with "With salaries like that it's not first world either", who is "they"? The country I said explicitly was third world? Because that's a tautology.
Fnoord 2y ago
Deepseek R1 performs well on a 600 EUR Jetson, and a 700 EUR AMD GPU. Both bought during COVID crisis. It is that quick. However, don't ask it about certain sensitive topics. You can bet your ass Musk is using his AI tools as propaganda tool for his advancement, just like he does with X. We already seen prompt leak of Grok, it wasn't neutral.
- mrtesthah 2y ago
  This is exactly correct. Grok is already slandering specific journalistic outlets such as “The Information” that publish negative stories on Musk.

mwigdahl 2y ago

No, it was not. It was not better than o1, nor o1-pro. Yes, it was _cheaper_ than those models, and superior in price/performance if the performance was acceptable. But in terms of raw performance it was behind them.

johnfn 2y ago
Chatbot arena leaderboard[1] says otherwise. 4o is ahead now, but that version of 4o was released after r1. r1 was ahead of all version of 4o and o1 at the time of release. https://huggingface.co/spaces/lmarena-ai/chatbot-arena-leaderboard https://huggingface.co/spaces/lmarena-ai/chatbot-arena-leade...
- mwigdahl 2y ago
  Chatbot arena leaderboard is a good test for vibes and style of response, but not much else. R1's performance in objective benchmarks (coding, etc.) showed very good performance, granted, but inferior to the full o1 and o1-pro models. It's still a very impressive feat, but it wasn't frontier-pushing.

_heimdall 2y ago

Is it your expectation that as models get cheaper we won't be developing much more powerful models at the higher price range? Its already worth noting that we already ran into the self hosting at scale problem. People don't want to run a web server and instead accept all the problems that come along with the convenience of social media. Why would LLMs, or any future AI product, be different?