matusp

16 ms·

271.
by matusp 2y ago
ELI5 what is the G2 situation?
272.
by matusp 2y ago
The gradients are noisy as they are, this additional noise probably does not hurt that much overall
273.
by matusp 2y ago
The parallels to ML you drew are on point. ML has this tendency to oversimplify complex phenomena with a easy to produce datasets, because that's what ML folks do, they find a smart and easy way to create a dataset and then they focus
274.
by matusp 2y ago
Please explain how Hopfield network influenced modern deep learning models based on supervised differentiable training. All the "impactful" architectures such as MLP, CNN, Attention, come from a completely different paradigm, a pa
275.
by matusp 2y ago
The techniques highlighted in this prize are not really that useful for deep learning.
276.
by matusp 2y ago
True, but I can easily see that the same role - short range airstrikes - could be played by drone teams in the future. They also have the capability to observe and strike targets over the treelines. And in top of that, they are cheaper, eas
277.
by matusp 2y ago
I am not surprised, drone warfare today requires a lot of people and a lot of manhours, and it does not scale that well. Several guys are often moving the equipment on foot to get as close to the frontline as possible, so that the pilot can
278.
by matusp 2y ago
Tbh Perun often just takes a report released by RUSI, ISW, or other war think tank and does some narration over it. On top of that, he often extrapolates from very small or noisy data.
279.
by matusp 2y ago
Attack helis were never that important compared to tanks of mechanized infantry. On top of that, it really seems that a significant part of their role can be fulfilled by drones.
280.
by matusp 2y ago
How is this phenomenon of very weakly grounded research called? Science washing or something?
281.
by matusp 2y ago
We already got an LLM generated meta review that was very clearly just summarization of reviews. There were some pretty egregious cases of borderline hallucinated remarks. This was ACL Rolling Review, so basically the most prestigious NLP v
282.
by matusp 2y ago
> I also think LLMs are wildly over hyped, and they will be shunted to a niche in the next few years, essentially providing autocomplete or search for business tasks for the most part, with strong inroads on video analysis, automated tra
283.
by matusp 2y ago
Yeah, this should be discussed more in my opinion. This entire business is just exploiting kids. I'm pretty worried about how my kids will behave when they get older and they will start to get bombarded by the Algorithm with all this &
284.
by matusp 2y ago
I remember when GPT-2 was "too dangerous" to release. I am confused why people still take these clown claims seriously.
285.
by matusp 2y ago
BERT is still the most downloaded LM at huggingface with 46M downloads last month. XLM Roberta has 24M and Distilbert is at 15M. I feel like BERTs are doing okay.
286.
by matusp 2y ago
I don't think that the design is that old. I was working as a web designed ~11 years back and a lot of these design ideas were "in" back then. The subtle dark and light 1px lines that create this embedded effect, some shadows
287.
by matusp 2y ago
This echoes our experience with LangChain, although we have abandoned it before putting it into production. We found out that for simple use cases it's too complex (as mentioned in the blog), and for complex use cases it's too dif
288.
by matusp 2y ago
"fix the English in the following prompt: {prompt}"
289.
by matusp 2y ago
You could add logs with timestamps. It is common that players forgot what actions they did, e.g. "Did I forgot to add a TR for this ocean?"
290.
by matusp 2y ago
It is not about the backs of cabinets. It is about telling your customers how big of a deal the backs of cabinets really are.
291.
by matusp 2y ago
They are getting ready for the inevitable copyright wars.
292.
by matusp 2y ago
As L2 English speaker I use it to check grammar - in/definite articles are very hard for me. To find words with specific meaning when I cannot recall them. During programming to get code snippets for staff that I do not use that often
293.
by matusp 2y ago
Substack is a good example as well. Literally cannot render simple text and image articles without visible and annoying lags. If there are several hundred comments it can take tens of seconds to finish rendering. With functionality that was
294.
Did we just receive an AI-generated meta-review?(opensamizdat.com)
3 points by matusp 2y ago|0 comments
295.
by matusp 3y ago
Yeah, I feel like the article is too quick with its conclusions. Many other problems can be made arbitrary complex and difficult with additional requirements. But there are still data structure and standard libraries to provide good enough
296.
by matusp 3y ago
If the company fires hundreds or thousands of people and there is no visible dip in the productivity, is it really a bad thing? One thing that is baffling to me is that these companies can fire 10% of their workforce and they just keep on c
297.
Abstraction Layers for LLM Security(opensamizdat.com)
2 points by matusp 3y ago|0 comments
298.
Disinformation Capabilities of Large Language Models(arxiv.org)
5 points by matusp 3y ago|1 comments
299.
by matusp 3y ago
nice project, but the model that was being studied is really just a toy-model (both in size and training data). as such, this model can indeed be approximated by simpler models (I would suspect even n-gram LMs), but it might not be represen
300.
by matusp 3y ago
Yeah, or villains building really edgy decor on their fortresses. Who was Sauron's decorator? They surely loved their spiky walls.

More ›