@wigglin.bsky.social
We've gone from SOTA on AIME going from 9% to 80% in a month. And then in the next month have 95% cost reduction. These are order of magnitude improvements and if you can't see that, you either have an ulterior motive or just aren't thinking about this clearly.
January 28, 2025 at 1:35 AM
We've gone from SOTA on AIME going from 9% to 80% in a month. And then in the next month have 95% cost reduction. These are order of magnitude improvements and if you can't see that, you either have an ulterior motive or just aren't thinking about this clearly.
I shared a chart comparing it's performance on many benchmarks, but for some reason you won't read it? Deepseek R1 is putting up numbers that were thought to be years out, today. You are not a serious person.
January 27, 2025 at 6:43 PM
I shared a chart comparing it's performance on many benchmarks, but for some reason you won't read it? Deepseek R1 is putting up numbers that were thought to be years out, today. You are not a serious person.
It's literally 95% cheaper to run and has state of the art performance. People are getting it to run locally on their laptops. I honestly don't have any idea what you are talking about.
January 27, 2025 at 6:33 PM
It's literally 95% cheaper to run and has state of the art performance. People are getting it to run locally on their laptops. I honestly don't have any idea what you are talking about.
What you are saying is not true
January 27, 2025 at 2:34 AM
What you are saying is not true
This is like saying in 2000 that the web is a bubble that is going to go away, because none of the of companies are making any money. You are partially right, but at more fundemental level, extremely wrong.
January 15, 2025 at 1:23 PM
This is like saying in 2000 that the web is a bubble that is going to go away, because none of the of companies are making any money. You are partially right, but at more fundemental level, extremely wrong.
Not enough that you got what you wanted... So weird
January 13, 2025 at 2:07 AM
Not enough that you got what you wanted... So weird
IEA estimates global data center (not just AI) power was 460 TWh in 2022. That's about 1 - 2% of global electricity demand. So these numbers seem ballpark accurate. If you have a better estimate, go ahead and share. www.iea.org/reports/elec...
Electricity 2024 – Analysis - IEA
Electricity 2024 - Analysis and key findings. A report by the International Energy Agency.
www.iea.org
January 10, 2025 at 10:09 PM
IEA estimates global data center (not just AI) power was 460 TWh in 2022. That's about 1 - 2% of global electricity demand. So these numbers seem ballpark accurate. If you have a better estimate, go ahead and share. www.iea.org/reports/elec...
Ok, find a better source that gives a different number. The best anyone is going to do is estimate this. If you have a better estimate, share it.
January 10, 2025 at 10:01 PM
Ok, find a better source that gives a different number. The best anyone is going to do is estimate this. If you have a better estimate, share it.
AI probably makes up around 0.1% of global electricity usage heatmap.news/podcast/shif...
A Skeptic’s Take on AI and Energy Growth
Inside episode 10 of Shift Key.
heatmap.news
January 10, 2025 at 2:43 PM
AI probably makes up around 0.1% of global electricity usage heatmap.news/podcast/shif...
If the models are good enough, it doesn't matter that much. If the probabilities align closely enough with the expected output, they can be useful. They are useful today, and getting more and more useful.
December 22, 2024 at 3:53 PM
If the models are good enough, it doesn't matter that much. If the probabilities align closely enough with the expected output, they can be useful. They are useful today, and getting more and more useful.
On the contrary, as these models get better at narrow technical domains, they may stagnate or degrade at more open domains. See what happened with o1 on English AP.
December 22, 2024 at 3:51 PM
On the contrary, as these models get better at narrow technical domains, they may stagnate or degrade at more open domains. See what happened with o1 on English AP.
Right, this should inspire more questions of a good journalist. If its wrong sometimes, why and how is it getting to those wrong answers? Is it improving over time? How will people use this technology even if it's not accurate 100% of the time, etc. etc.
December 21, 2024 at 4:10 PM
Right, this should inspire more questions of a good journalist. If its wrong sometimes, why and how is it getting to those wrong answers? Is it improving over time? How will people use this technology even if it's not accurate 100% of the time, etc. etc.
You asked a question to a chatbot built on top of an LLM, which by its nature is probabilistic, and it gave you a bad answer. Therefore, the technology has no use and is just bad.
December 21, 2024 at 4:06 PM
You asked a question to a chatbot built on top of an LLM, which by its nature is probabilistic, and it gave you a bad answer. Therefore, the technology has no use and is just bad.
1, 2, and 5 are clearly solved. 4 is not a reasonable goal because it's not something that a human can one-shot either.
December 21, 2024 at 4:04 AM
1, 2, and 5 are clearly solved. 4 is not a reasonable goal because it's not something that a human can one-shot either.
1. And yet, they keep getting better and better at reducing confabulations. 2. Three years ago, the consensus opinion was that we were decades out from AI models with the capabilities that are now widespread in 2024.
December 16, 2024 at 5:52 PM
1. And yet, they keep getting better and better at reducing confabulations. 2. Three years ago, the consensus opinion was that we were decades out from AI models with the capabilities that are now widespread in 2024.
Why shouldn't he have some hubris/confidence? His whole career people have been telling him that deep learning won't work. "It'll never solve x." And yet, he has been vindicated over and over and over again.
December 16, 2024 at 5:34 PM
Why shouldn't he have some hubris/confidence? His whole career people have been telling him that deep learning won't work. "It'll never solve x." And yet, he has been vindicated over and over and over again.
Labor theory of the value of art. We enjoy art proportionally to the amount of time we can see an artist has devoted to its creation.
December 14, 2024 at 4:57 PM
Labor theory of the value of art. We enjoy art proportionally to the amount of time we can see an artist has devoted to its creation.
I think you are just too online. You are safe. No one is going to get you.
December 14, 2024 at 3:23 PM
I think you are just too online. You are safe. No one is going to get you.
I think you are okay. You can just block this Jesse guy if you don't like what he writes. You are safe.
December 14, 2024 at 3:20 PM
I think you are okay. You can just block this Jesse guy if you don't like what he writes. You are safe.
This website is so funny. People will say something reasonable, and get dozens of the most hateful toxic histrionic replies imaginable. This site is so much worse than twitter, which is ironic.
December 13, 2024 at 7:03 PM
This website is so funny. People will say something reasonable, and get dozens of the most hateful toxic histrionic replies imaginable. This site is so much worse than twitter, which is ironic.
How are you unsafe?
December 13, 2024 at 6:38 PM
How are you unsafe?
Worse and more toxic than Twitter, in record time. I honestly didn't think it was possible
December 8, 2024 at 2:46 PM
Worse and more toxic than Twitter, in record time. I honestly didn't think it was possible
I see what you mean. Personally, I recall using chatbots like Eliza-clones, or even Google Assistant/ Amazon Alexa, and always being disappointed by them. So when ChatGPT came out, I was truly amazed that they finally worked. Still am amazed.
December 6, 2024 at 2:29 AM
I see what you mean. Personally, I recall using chatbots like Eliza-clones, or even Google Assistant/ Amazon Alexa, and always being disappointed by them. So when ChatGPT came out, I was truly amazed that they finally worked. Still am amazed.
How can it both be a hyped bubble, and a threat to workers? Aren't these mutually exclusive?
December 6, 2024 at 2:22 AM
How can it both be a hyped bubble, and a threat to workers? Aren't these mutually exclusive?