ynniv

9 ms·

61.
by ynniv 2y ago
How would you test this? Claude will play Battleship with an understanding of ship geometry.
62.
by ynniv 2y ago
In my experience Claude can do this, but can we rephrase it in a way that's more testable?
63.
by ynniv 2y ago
Not Claude
64.
by ynniv 2y ago
They can make SVGs
65.
by ynniv 2y ago
This is a design choice
66.
Ask HN: What can LLMs not do?
16 points by ynniv 2y ago|59 comments
67.
by ynniv 2y ago
Other than we don't tell it how to get the right answer, or understand how it eventually computes correct answers?
68.
by ynniv 2y ago
There's an extension for deletion: https://nostr-nips.com/nip-09 And experimental support for editing. I am personally skeptical of these because once you've probably said something, you can ask people to ignore i
69.
by ynniv 2y ago
Links to Yan: Title: "Objective Driven AI: Towards Machines that can Learn, Reason, and Plan" Lytle Lecture Page: https://ece.uw.edu/news-events/lytle-lecture-series/ Slides: https://drive.go
70.
by ynniv 2y ago
Protocols have: - No accounts - No moderation - No permissions - No rules "My nsec is my passport. Verify... me?" I'm posting straight into the spec because it's honestly that easy to read, but there are ple
71.
Nostr is public key microblogging that works(github.com)
33 points by ynniv 2y ago|14 comments
72.
by ynniv 2y ago
nostr is the hackers microblog. Just sign things and relay them: https://github.com/nostr-protocol/nips/blob/master/01.md https://news.ycombinator.com/item?id=42489954
73.
by ynniv 2y ago
Is it "eerie"? LeCun has been talking about it for some time, and may also be OpenAI's rumored q-star, mentioned shortly after Noam Brown (diplomacybot) joining OpenAI. You can't hill climb tokens, but you can climb mani
74.
by ynniv 2y ago
Please, read the Wolfram blog
75.
by ynniv 2y ago
These kind of hand wavey statements like “practice,” “grok,” and “nonlinear its capabilities will be” are not very constructive as they don’t have solid meaning wrt language models. So, here's my hypothesis, as someone who is adjacen
76.
by ynniv 2y ago
It's overfitting when you train too large a model on too many details. Rote memorization isn't rewarding. The more concepts the model manages to grok, the more nonlinear its capabilities will be: we don't have a data problem,
77.
by ynniv 2y ago
Claude is trained on principles. GPT is trained on billions of edge cases. Which student do you prefer?
78.
by ynniv 2y ago
Everyone's obsessed with new training tokens... It doesn't need to be more knowledgeable, it just needs to practice more. Ask any student: practice is synthetic data.
79.
by ynniv 2y ago
Canoe in the ocean, camp in the swamp
80.
by ynniv 2y ago
I too enjoy defining my default attribute values at the protocol level, so they can never, ever be changed, and then defensively coding library functions in case the arguments are not constructed under the same semantic assumptions as I had
81.
by ynniv 2y ago
"Only"
82.
by ynniv 2y ago
The Java impl for protobuf will never generate a NullPointerException, as calling `get` on a field would just return the default instance of that field. This was a mistake. You still want to check whether it was initialized most of the ti
83.
by ynniv 2y ago
That looks like an AI hallucination
84.
by ynniv 2y ago
I don't think one model is statistically significant. As people have pointed out, it could have chess specific responses that the others do not. There should be at least another one or two, preferably unrelated, "good" data p
85.
by ynniv 2y ago
Not true: opinions carry the weight of their underlying reasons. Not all opinions are equally supported.
86.
by ynniv 2y ago
Quietly
87.
by ynniv 2y ago
https://raspibolt.org
88.
by ynniv 2y ago
Apathy
89.
by ynniv 2y ago
Hosted models have APIs that will be patched, local models tend to use a handful of libraries that will also be patched. Yet Another Nothing Burger.
90.
by ynniv 2y ago
Don't take my word for it, but this crow is delicious.

More ›