Hugsun

12 ms·

121.
by Hugsun 2y ago
In the article, I make the case that they can reason but it is in many ways distinct from the way we do it. It is also learned from very different fundamentals. Our reasoning fundamentals come from sensing, and experiencing. LLMs reasoning
122.
by Hugsun 2y ago
You might enjoy the analysis in the article. https://www.arnaldur.be/writing/about/large-language-model-r... There is an expandable box with all the questions, colored by correctness, and question 6 has a bunch of
123.
by Hugsun 2y ago
That's very true. I thought about speaking more to that issue but the post was already longer than I wanted. You might find this short analysis interesting. https://www.arnaldur.be/experimenting/with/large-lan
124.
by Hugsun 2y ago
I know about the Chinese Room. The pivotal difference is that the intelligence is performing a mechanical task and none of its intelligence is inserted into the inputs. No synthesis is performed. In the API example, it's literally peop
125.
by Hugsun 2y ago
Look at question six and its responses. That is what I look into there. From the article: > I asked the model further about its 6th response. It realized its error at the slightest hint of disagreement, but just [asking it to elaborate]
126.
by Hugsun 2y ago
> But just ask ChatGPT (4o) how a farmer and a sheep would cross a river on a boat. For certain formulations, it reliably hallucinates a wolf or a piece of cabbage into the problem. Why? Because it pattern matches the solution to a well-
127.
by Hugsun 2y ago
That is exactly the point I'm trying to make in the chapter They don't reason like us: https://www.arnaldur.be/writing/about/large-language-model-r...
128.
by Hugsun 2y ago
Interesting concept, it's especially apt, as GEB has been sitting, unread, on my desk for about a month now. I'll take the hint and read it ;)
129.
by Hugsun 2y ago
Why is that API, as specified, not intelligent? If your calculator could do everything you can do, as long as it fits through an API, I would call it intelligent, would you not?
130.
by Hugsun 2y ago
I don't rigorously define reason in the article and I state that it is hard to draw clear boundaries. I'm more relying on the readers intuitive understanding of the idea, which is perhaps not a good thing to do. I wouldn't sa
131.
by Hugsun 2y ago
It would certainly be lossy. We don't know how to make them reliably know any particular facts. Neither training or in context learning does this without fail.
132.
Can large language models reason?(arnaldur.be)
14 points by Hugsun 2y ago|60 comments
133.
by Hugsun 2y ago
I'm surprised they don't mention grouped query attention at all.
134.
by Hugsun 2y ago
It will be very interesting to see which tools become the de-facto for each of the common LLM use cases. The landscape is so fractured, I feel like I haven't even heard of most tools. I encountered Microsoft's olive the other day
135.
by Hugsun 2y ago
It's funny. I have really wanted a product like this for a while. There is absolutely no way I would consider a non-FOSS program to do this though.
136.
by Hugsun 2y ago
I think you vastly overestimate how much people care about model censorship. There are a bunch of open models that aren't censored. Llama 3 is still way more popular because it's just smarter.
137.
by Hugsun 2y ago
Lmao
138.
by Hugsun 2y ago
That is not true. Transformers are being applied all over because they work better than what was used before in so many cases.
139.
by Hugsun 2y ago
I never played myst or riven. I loved The Witness and the Talos Principle. What concessions were made to the internet age?
140.
by Hugsun 2y ago
I was thinking about this in the context of the graffiti subculture around the German rail system. It's pretty interesting and different from random tagging. I somewhat agree with you if we are talking about low effort tagging, especia
141.
by Hugsun 2y ago
Two of my replies just said that graffiti was worse than bank robberies. That's hilarious and crazy to me. I get your point though, on reading the comment I replied to, after reading yours, it's clear to me that the equivalence im
142.
by Hugsun 2y ago
I was thinking about societies in general. Bank robberies are absolutely worse. They can involve multiple lives being at much greater risk and a large amount of resources allocated to criminals. The fact that it doesn't come from the b
143.
by Hugsun 2y ago
Architecturally, LLMs are very simple compared to many software projects. The crux of their behavior comes from their learned weights which are gigabytes and can cost millions to obtain via training.
144.
by Hugsun 2y ago
Implying that graffiti is as bad as bank robberies is wild.
145.
by Hugsun 2y ago
Hah! That's also sad. I don't want to in the US. I'm spoiled by walkability and quality cycling infrastructure.
146.
by Hugsun 2y ago
Cool! I'm amused that it's on the Steelseries website and it's not so clear how these two are related.
147.
by Hugsun 2y ago
That's such a frustrating thing. There are so many opportunities in SF alone. But I really don't want to live there.
148.
by Hugsun 2y ago
According to the paper[1], Gemini Nano-1 is 1.8B, and (Nano-2) is 3.25B parameters. They are quantized to 4-bits which makes them 0.9GB and 1.6GB respectively. That's a juicy binary blob to ship with Chrome. As this will be on device,
149.
New SOTA 4-bit weight and KV quantization with optimized GPU performance(arxiv.org)
3 points by Hugsun 2y ago|0 comments
150.
by Hugsun 2y ago
I believe model distillation does this. SparseGPT was a big one, managing to remove 50% of parameters without loosing much accuracy IIRC. I saw a more recent paper citing the SparseGPT one that managed around 70-80% sparsity, pretty impress

More ›