Nils Höglund
@nilsengineer.bsky.social
Civilingenjör i teknisk fysik
Bästa verktygen byggs utifrån välfungerande redan existerande processer och tillsammans med de som är experter på dessa.
March 8, 2025 at 10:26 AM
Bästa verktygen byggs utifrån välfungerande redan existerande processer och tillsammans med de som är experter på dessa.
Jail for civil contempt of court until following orders?
February 19, 2025 at 1:30 PM
Jail for civil contempt of court until following orders?
M1 Max från år 2021 med 64 GB RAM
January 27, 2025 at 4:12 PM
M1 Max från år 2021 med 64 GB RAM
Ja. Bra kvalitet men lite långsamt köra på min MacBook (14B snabbare än 30B osv)
January 27, 2025 at 4:04 PM
Ja. Bra kvalitet men lite långsamt köra på min MacBook (14B snabbare än 30B osv)
R1 released last week. Took until the weekend for a critical mass of tech experts to evaluate performance (there is so much happening within AI so it takes a while to separate the hype from the real deal).
January 27, 2025 at 3:46 PM
R1 released last week. Took until the weekend for a critical mass of tech experts to evaluate performance (there is so much happening within AI so it takes a while to separate the hype from the real deal).
For the cases I have tested. For a more scientific approach, see the officially released benchmarks (which I believe also are on par with o1).
January 26, 2025 at 8:19 PM
For the cases I have tested. For a more scientific approach, see the officially released benchmarks (which I believe also are on par with o1).
Here is an upvoted recent top comment from someone else on Hacker News:
news.ycombinator.com/item?id=4282...
news.ycombinator.com/item?id=4282...
January 26, 2025 at 7:32 PM
Here is an upvoted recent top comment from someone else on Hacker News:
news.ycombinator.com/item?id=4282...
news.ycombinator.com/item?id=4282...
But there will probably be more open source large language models that are great eventually. I believe the gates have opened now and many people will move off dependency of proprietary plattforms such as OpenAI’s. But only the future can tell 🙂
January 26, 2025 at 7:11 PM
But there will probably be more open source large language models that are great eventually. I believe the gates have opened now and many people will move off dependency of proprietary plattforms such as OpenAI’s. But only the future can tell 🙂
Yep. At least for most stuff. Eventually for all.
January 26, 2025 at 7:07 PM
Yep. At least for most stuff. Eventually for all.
I use it for questions like how to implement/design some parts of systems in dotnet/C#.
January 26, 2025 at 7:05 PM
I use it for questions like how to implement/design some parts of systems in dotnet/C#.
I think something like 5 tokens/s running the R1 30B model locally. So maybe a few minutes for it to answer a difficult/complicated question. The o1 does it in 10 seconds maybe (like the online R1). But if the results keeps being better, I will still prefer the wait for more complicated questions.
January 26, 2025 at 7:03 PM
I think something like 5 tokens/s running the R1 30B model locally. So maybe a few minutes for it to answer a difficult/complicated question. The o1 does it in 10 seconds maybe (like the online R1). But if the results keeps being better, I will still prefer the wait for more complicated questions.
Yes. I thought the answer was better than OpenAI o1. So I will keep using DeepSeek R1 for coding questions. The full version hosted by them in China for non-sensitive stuff and my local version for potentially sensitive stuff (slower but safe). 🙂
January 26, 2025 at 6:57 PM
Yes. I thought the answer was better than OpenAI o1. So I will keep using DeepSeek R1 for coding questions. The full version hosted by them in China for non-sensitive stuff and my local version for potentially sensitive stuff (slower but safe). 🙂
Did you use the 14B?
January 26, 2025 at 6:50 PM
Did you use the 14B?
Much slower, but works.
January 26, 2025 at 6:45 PM
Much slower, but works.
Not in response generation speed but in code quality/solving the actual task it was actually better than o1 in my own test this morning
bsky.app/profile/nils...
bsky.app/profile/nils...
In my own real world coding test earlier today the 32B model was better than o1. 14B on par with o1-mini in that test.
January 26, 2025 at 6:44 PM
Not in response generation speed but in code quality/solving the actual task it was actually better than o1 in my own test this morning
bsky.app/profile/nils...
bsky.app/profile/nils...
Those tests this morning I ran with the help of ollama&ollama-ui on my own MacBook M1 Max from year 2021 (64GB RAM).
You can currently use the full model on their page for free (but I would not submit anything sensitive there) www.deepseek.com
You can currently use the full model on their page for free (but I would not submit anything sensitive there) www.deepseek.com
DeepSeek
DeepSeek, unravel the mystery of AGI with curiosity. Answer the essential question with long-termism.
www.deepseek.com
January 26, 2025 at 6:43 PM
Those tests this morning I ran with the help of ollama&ollama-ui on my own MacBook M1 Max from year 2021 (64GB RAM).
You can currently use the full model on their page for free (but I would not submit anything sensitive there) www.deepseek.com
You can currently use the full model on their page for free (but I would not submit anything sensitive there) www.deepseek.com
In my own real world coding test earlier today the 32B model was better than o1. 14B on par with o1-mini in that test.
January 26, 2025 at 6:38 PM
In my own real world coding test earlier today the 32B model was better than o1. 14B on par with o1-mini in that test.
”As good as gpt3”. lol. Performs in par with o1.
January 26, 2025 at 6:34 PM
”As good as gpt3”. lol. Performs in par with o1.
I run the 30B DeepSeek R1 model locally on my MacBook M1 Max from year 2021. Works.
January 26, 2025 at 6:03 PM
I run the 30B DeepSeek R1 model locally on my MacBook M1 Max from year 2021. Works.
It seems less censored than other models. In addition, if you download the model and use it on your own hardware, DeepSeek can’t take your data.
January 26, 2025 at 6:01 PM
It seems less censored than other models. In addition, if you download the model and use it on your own hardware, DeepSeek can’t take your data.