gabriel666smith

22 ms·

91.
by gabriel666smith 1y ago
Love this! One thing I'd find frustrating if using this when working on music: "When beat is playing, get new sample for drum plays as soon as it generates". If "this sample is programmed to play in playing beat", t
92.
by gabriel666smith 1y ago
Thanks! That's really helpful. > What contradictions do you see? I don't see any. I guess us seeing very different things is also what a settlement might be for :-). But I think I was wrong. I think others in the thread are deb
93.
by gabriel666smith 1y ago
Thanks!
94.
by gabriel666smith 1y ago
Would it actually set any kind of legal precedent, or just establish a sort of cultural vibe baseline? I know Anthropic doesn't have to admit fault, and I don't know if that establishes anything in either direction. But I'm n
95.
by gabriel666smith 1y ago
This is a 2025 snapshot of Libgen, rather than the dataset Anthropic trained on, or the 'reduced to 500k books' dataset of books by authors who are deemed legally due some cash from Anthropic because Anthropic downloaded the ebook
96.
by gabriel666smith 1y ago
Someone I know (not SWIM, it actually wasn't me) did a little bit of work for a high-school-level tutoring company. The company got all their leads from TikTok. They did numbers on TikTok. The company was tutoring English Literature as
97.
by gabriel666smith 1y ago
The idea of giving an LLM with a tool any kind of control over an actual user's account remains (though you put this more elegantly) batshit insane to me. Even assuming you've correctly auth'd the user contacting you (big a
98.
by gabriel666smith 1y ago
I MVP'd one of these (a basic sequence of LLM customer support 'agents') at my last job, I guess spring 2024. So much has changed since then! 'Routing through increasingly specialised agents' was my approach, and th
99.
by gabriel666smith 1y ago
Thank you for your good-faith explanation. > Empathy like all emotions require effort and cognitive load, and without things being mutual you will eventually slowly become drained, bitter and resentful because of empathy fatigue. Do you
100.
by gabriel666smith 1y ago
Maybe I don’t understand well enough. Could anyone highlight what the problems are with this fix? 1. If ‘bad topic’ detected, even when model believes it is in ‘roleplay’ mode, pass partial logs, attempting to remove initial roleplay framin
101.
by gabriel666smith 1y ago
>You cannot be empathetic to complete strangers. Why not? I’m not trying to inflame this further, I’m genuinely interested in your logic for this statement.
102.
Show HN: Create a clone of your old self from your old iOS backup; talk to it(github.com)
3 points by gabriel666smith 1y ago|0 comments
103.
by gabriel666smith 1y ago
On it! Thanks! Will be fun to learn how to do 'definitely random'.
104.
by gabriel666smith 1y ago
Yep, the latter is far more interesting I think! Those results I got in the other file in the repo were fascinating to me
105.
by gabriel666smith 1y ago
For interim adjacent results, I published https://github.com/henrygabriels/FMLLM/blob/main/improving_l... also, but I'm this moment working on a similar suggestion (comparing coherence 'fib-ass
106.
by gabriel666smith 1y ago
I've just added the model generation code! I hope that's helpful. 1. Yes — that’s exactly right. It counts. If "Hacker" appeared 2 places before "News" multiple times, the counts would reflect that. Later, when
107.
by gabriel666smith 1y ago
Thank you, I agree, I think it'd be helpful to publish aspects of it. > Are you saying the training task is to ask for the (fib_i)th token rather than the next token? Yes, functionally - I explained in more detail in another comment
108.
by gabriel666smith 1y ago
> Where are you going with this? I don't know! Adding code now. Is this more clear, for at least the initial 'word set' generation? I can add to repo if so: Concept: The system predicts and generates words based on Fibonac
109.
by gabriel666smith 1y ago
Thank you for all this! I ran some controls in a separate study more recently - the (tiny!) results seemed interesting to me: https://github.com/henrygabriels/FMLLM/blob/main/improving_l... This compares
110.
by gabriel666smith 1y ago
I found some weird results when messing around with different embeddings in text generation. I'm not sure if this is meaningful, and - if anyone on here is interested - I could use some help figuring out what's going on. I was inv
111.
Fmllm: 4mb training data, 100mb model, Fibonacci embeddings, near-coherent. WTF?(github.com)
37 points by gabriel666smith 1y ago|27 comments
112.
by gabriel666smith 1y ago
I genuinely appreciate the more specific feedback! I thought I'd give it to the model to handle: Input: "thank you for the more substantive but negative feedback" Output: "thank you for the more substantive but negative
113.
by gabriel666smith 1y ago
This feels a little unkind, but also functionally inaccurate. The previous readme I posted was for an Apple Music client. It contained a download link for the client. Other Apple Music clients shared in the same way were heavily-upvoted by
114.
FMLLM: 4mb training data, 100mb model, Fibonacci embeddings, near-coherent. WTF?(github.com)
2 points by gabriel666smith 1y ago|5 comments
115.
by gabriel666smith 1y ago
I found some weird results when messing around with different embeddings in text generation. I'm not sure if this is meaningful, and - if anyone on here is interested - I could use some help figuring out what's going on.
116.
Show HN: Scratch Radio Ft Doctor Von Peel: An LLM / Cat-Based Apple Music Client(standard-default-and-grace.github.io)
2 points by gabriel666smith 1y ago|0 comments
117.
by gabriel666smith 1y ago
Agree. There are probably more than a million tiny contextual data points that make a person look at something (whether it’s a tech product or a musician) and go: “cool, man.” But those millions of data points can (rarely, briefly,) coalesc
118.
by gabriel666smith 1y ago
This is cool! I've recently been working on a similar project as a way to "walk around" in my own writing, understand my characters more, etc. I like the three-choice mechanism, with free-text, which I also use. I do feel lik
119.
by gabriel666smith 1y ago
I've been tracking usage in my first month of "20x" Max (which was, unfortunately, this month). Depending on how this usage is amortised (working days, which is what matters to me, or 5 hour periods, or I guess, now weeks..?)
120.
by gabriel666smith 1y ago
No, quite the opposite, apologies if I was unclear. I think that LLMs are a tool, and a tool that is still in the process of being iterated on. I think that how this new tool could be applied and iterated on by humans who were, I think, uni

More ›