namaria

14 ms·

211.
by namaria 2y ago
Hackers and Founders meetup near me has become mostly just founders and they are all high on the LLM fumes. It's very weird.
212.
by namaria 2y ago
I metrics fetish might be quite widespread. Why would anyone need an app to tell them how well they've slept? I can feel how well I've slept when I wake up. But this can be used to great effect. Make an argument and people will ag
213.
by namaria 2y ago
I noticed that I am often audited when I am unshaven and wear hoodies and almost never when I have a sports jacket on...
214.
by namaria 2y ago
I understand your position. But I think you're underestimating just how much training data is used and how much information can be encoded in hundreds of billions of parameters. But this is the crux of the disagreement. I think the mod
215.
by namaria 2y ago
Good man I never said anything about AGI. Why do you keep responding to things I never said? This whole exchange was you having knee-jerk reactions to things you imagined I said. It has been incredibly frustrating. And at the end you shrug
216.
by namaria 2y ago
If do not you think past research by OpenAI and Anthropic on how to use LLMs to generate code is relevant to how Anthropic LLMs generate code 3 years later I really don't think it is possible to have a reasonable conversation about thi
217.
by namaria 2y ago
It's overkill. The models do not capture knowledge about coding. They overfit to the dataset. When one distills data into a useful model the model can be used to predict future behavior of the system. That is the premise of LLM-as-AI.
218.
by namaria 2y ago
> "they can only produce code that's almost identical to what's in their training data" Again, you're misinterpreting in a way that seems like you are reacting to the perception that someone attacked some of your
219.
by namaria 2y ago
I am sorry but that's nonsense. I quoted the paper "Evolution through Large Models" written in collaboration between OpenAI and Anthropic researchers "In other words, the model learns to predict plausible changes to code
220.
by namaria 2y ago
All due respect to Simon but I would love to see some of that groundbreaking code that the LLMs are coming up with. I am sure that the functionalities implemented are novel but do you really think the training data cannot possibly have had
221.
by namaria 2y ago
> The idea that models can only write code if they've seen code that does the exact same thing in the past is deeply uninformed in my opinion. This is a conceited interpretation of what I said.
222.
by namaria 2y ago
> ... which explains why some models are better at code than others. No. It explains why models seem better at code in given situations. When your prompt mapped to diffs in the training data that are useful to you they seem great.
223.
by namaria 2y ago
It's overfitting. Some people say they find LLMs very helpful for coding, some people say they are incredibly bad. I often see people wondering if the some coding task is performed well or not because of availability of code examples i
224.
by namaria 2y ago
I don't think it's so clear cut. The wikipedia article cites a 1971 book and an article from 1991 as sources. A paper from 2014 is not so sure this is a settled matter: "Infant object permanence is still an enigma after four
225.
by namaria 2y ago
I for one look forward to LLMs automating the work of MBAs.
226.
by namaria 2y ago
> the most cracked CS students I know deliberately avoid using these tools for programming work. They understand the value in the struggle I think they are in the right path here > they've automated their humanities, social scien
227.
by namaria 2y ago
> can ultimately understand the underlying reasoning This is at the root of the Dunnin-Kruger effect. When you read an explanation you feel like you understand it. But it's an illusion, because you never developed the underlying cog
228.
by namaria 2y ago
Thank you for bringing light to this. I think it makes a very relevant point to us as well. The value of doing the work yourself is in internalizing and developing one's own cognition. The argument of offloading to the LLM to me sounds
229.
by namaria 2y ago
I feel conflicted about this. On the one hand, it seems like an enormous waste and a threat to the effort to contain climate change. On the other hand, this would seem in line with the idea that life has evolved to more efficiently dissipat
230.
by namaria 2y ago
> there are literal hundreds of billions at stake and people are lying through their teeth for money, as they always do. That's pretty much how I've felt about life in this world ever since I can remember.
231.
by namaria 2y ago
Thanks for the perspective. I don't feel love or hate, I am just perplexed (haha) about the discourse around it sometimes.
232.
by namaria 2y ago
I thought the value proposition of using LLMs to code is the lesser cognitive load of just describing what you want in natural language. But if it turns out writing the prompt is so involved, you end up trading snippets on forums and you of
233.
by namaria 2y ago
Doesn't trading prompt patches trying to get around undefined behavior from the model make you wonder if this is a net positive?
234.
by namaria 2y ago
If you're gonna formulate this conversation as a satisfiability problem you should be aware that this is an NP-complete problem (and actually working on that problem is the source of the insight that there is such as thing as NP-comple
235.
by namaria 2y ago
Undefined behavior as a service is truly a bizarre proposition to my ears. Layering undefined behavior (agents) and gaming undefined behavior in hopes it comes out as you need (prompting) sounds insane and sometimes I have to wonder if I am
236.
by namaria 2y ago
Not to mention the variations of "you need to prompt better" including now "rules files" which begs the question: wouldn't just writing code be a much better way to exercise control over the machine?
237.
by namaria 2y ago
> Or maybe I just suck at prompting hehe Every time someone argues for the utility of LLMs in software development by saying you need to be better at prompting, or add more rules for the LLM on the repository, they are making an argument
238.
by namaria 2y ago
> I think it's your responsibility to control the LLM. Yes. The issue here is control and NLP is a poor interface to exercise control over the computer. Code on the other hand is a great way. That is the whole point of skepticism ar
239.
by namaria 2y ago
Also I think this has been a long time dream of business types. They have always resented domain experts, because they need them for their businesses to be successful. They hate the leverage the domain experts have and they think these LLMs
240.
by namaria 2y ago
Do they do anything or are they an expensive toy in the shape of a humanoid robot?

More ›