Kenneth Chiu
@kjw-chiu.bsky.social
Assoc. Prof. of Computer Science at Binghamton University. Science groupie. Also parallel computing, AI4Science. I mainly reply, haphazardly and eclectically. Expect silly jokes, random things that I found interesting, etc.
If you can use LaTeX, then Overleaf. If not, MS Sharepoint for Word, or Google Docs for Word. I'm a bit suspicious of Google Docs for Word, though, given that Word is not a Google product.
May 18, 2025 at 6:35 PM
If you can use LaTeX, then Overleaf. If not, MS Sharepoint for Word, or Google Docs for Word. I'm a bit suspicious of Google Docs for Word, though, given that Word is not a Google product.
The overall point remains, but as a technical point, the copies could be in encrypted form.
May 2, 2025 at 5:56 PM
The overall point remains, but as a technical point, the copies could be in encrypted form.
Is there a paper or an example of using Orb, or some other NNP to simulate a reaction, such as 2 H2 + O2 -> 2 H2O? Would it be possible to do it simply by giving the reactants some initial collision velocity > activation energy, and just running time steps forward?
March 20, 2025 at 10:15 PM
Is there a paper or an example of using Orb, or some other NNP to simulate a reaction, such as 2 H2 + O2 -> 2 H2O? Would it be possible to do it simply by giving the reactants some initial collision velocity > activation energy, and just running time steps forward?
For what it's worth, a bit of context is that Andreesen got his start at UIUC/NCSA. Perhaps he would argue that universities were good back then, though.
March 20, 2025 at 1:42 PM
For what it's worth, a bit of context is that Andreesen got his start at UIUC/NCSA. Perhaps he would argue that universities were good back then, though.
Someone probably googled "spectrum". Saw something that looked like a rainbow.
March 20, 2025 at 1:36 PM
Someone probably googled "spectrum". Saw something that looked like a rainbow.
But we have different subsystems. Our visual processing, for example, is highly parallel and fast. For example, if we flash an image of a dog or cat, we can classify it very quickly.
March 18, 2025 at 3:23 PM
But we have different subsystems. Our visual processing, for example, is highly parallel and fast. For example, if we flash an image of a dog or cat, we can classify it very quickly.
Hm...I would say that you could compare them, but you would have to lay out exactly how you define the terms. For example, you could approximately define cognitive RAM by short-term memory. An FPGA-like device might be a better analogy for a human brain.
March 18, 2025 at 3:18 PM
Hm...I would say that you could compare them, but you would have to lay out exactly how you define the terms. For example, you could approximately define cognitive RAM by short-term memory. An FPGA-like device might be a better analogy for a human brain.
A guess is because they want to use the interactions as training data, and figure that the potential annoyance is worth it.
March 17, 2025 at 3:06 AM
A guess is because they want to use the interactions as training data, and figure that the potential annoyance is worth it.
If they made the in-car video available, it would be interesting to try some optical-flow depth estimation model.
March 16, 2025 at 9:42 PM
If they made the in-car video available, it would be interesting to try some optical-flow depth estimation model.
Lack of motion parallax may also be enough to detect something fishy, for models that can analyze a sequence of frames.
March 16, 2025 at 9:38 PM
Lack of motion parallax may also be enough to detect something fishy, for models that can analyze a sequence of frames.
(And I do generally think that lidar is a good thing.)
March 16, 2025 at 9:30 PM
(And I do generally think that lidar is a good thing.)
I wonder whether or not this would have also fooled a human driver? I suspect that if they were attentive, they would have noticed discrepancies such as lack of motion parallax. But it's possible that it would not become super-apparent until too late.
March 16, 2025 at 9:29 PM
I wonder whether or not this would have also fooled a human driver? I suspect that if they were attentive, they would have noticed discrepancies such as lack of motion parallax. But it's possible that it would not become super-apparent until too late.
How do you define "smarter"? I've found R1 to often do better than o1 on reasoning problems, but would not call it smarter.
January 28, 2025 at 5:37 PM
How do you define "smarter"? I've found R1 to often do better than o1 on reasoning problems, but would not call it smarter.
I've had ChatGPT Plus for a while now, and so far, I've been finding R1 to be at least as good as o1, and often better. But mostly only do math/logic word problems, and coding problems.
January 28, 2025 at 3:38 PM
I've had ChatGPT Plus for a while now, and so far, I've been finding R1 to be at least as good as o1, and often better. But mostly only do math/logic word problems, and coding problems.
Google Lens and Google Translate, to name two. Many people don't realize that the underlying technology is the same (deep neural networks, transformers, etc.).
January 27, 2025 at 4:20 PM
Google Lens and Google Translate, to name two. Many people don't realize that the underlying technology is the same (deep neural networks, transformers, etc.).
It's really impressive (to me at least) that they trained it for so little cost.
January 23, 2025 at 3:58 PM
It's really impressive (to me at least) that they trained it for so little cost.
If you haven't yet, try R1, at least for questions that require more reasoning. I've been impressed by it, and for the problems that I've tested it on, it's done as good as or better than o1.
January 23, 2025 at 3:11 PM
If you haven't yet, try R1, at least for questions that require more reasoning. I've been impressed by it, and for the problems that I've tested it on, it's done as good as or better than o1.
I've actually found it outperform o1. Here, R1 is right, o1 is wrong. (Apparently no way to link to a DeepSeek chat, so just screenshot of the answer.)
o1: chatgpt.com/share/679063...
o1: chatgpt.com/share/679063...
January 23, 2025 at 6:06 AM
I've actually found it outperform o1. Here, R1 is right, o1 is wrong. (Apparently no way to link to a DeepSeek chat, so just screenshot of the answer.)
o1: chatgpt.com/share/679063...
o1: chatgpt.com/share/679063...
Are you rubbing it in?
January 23, 2025 at 5:47 AM
Are you rubbing it in?