gpm

106 ms·

2071.
by gpm 4y ago
> - Bots were most likely largely operated via the AP This strikes me as extremely unlikely. The bots being discussed were attempting to pretend not to be bots, and there were active attempts to detect and remove them. Using the API woul
2072.
by gpm 4y ago
LLaMa 65B and 30B were trained on 1.4 trillion tokens. This model was trained on 260 billion tokens.
2073.
by gpm 4y ago
The LLaMa paper says 1 trillion for the smaller models (7B, 13B) and 1.4 trillion for the larger models (30B, 65B)
2074.
by gpm 4y ago
Unclear, likely jurisdiction dependent, almost certainly not if you need to operate world wide.
2075.
by gpm 4y ago
Really short summary: LLaMa is better, even smaller LLaMa models. Table format: Benchmark, Cerebras 13B, LLama 7B, LLama 13B, LLama 60B HellaSwag, 51.3, 76.1, 79.2, 84.2 Piqa, 76.6, 79.8, 80.1, 82.8 Wino-Grande, 64.6, 70.1, 73.0, 77.0 Arc-e
2076.
by gpm 4y ago
Is there a benchmark comparing the two that I missed? Edit: The huggingface page has 0-shot benchmarks which you can compare against the llama paper https://huggingface.co/cerebras/Cerebras-GPT-13B https://a
2077.
by gpm 4y ago
From my perspective a large part of the issue is that they were trying to charge the wrong people. What they needed to do was figure out some way to make the people using dockerhub to build things to pay. What they attempted to do was charg
2078.
by gpm 4y ago
It was probably also primarily valuable as a recruiting tool, and if they're not recruiting as heavily it no longer makes as much sense to invest in it.
2079.
by gpm 4y ago
Sure, but your credit card company should be paying people who you direct them to, not paying people who claim that you owe them money.
2080.
by gpm 4y ago
I wonder if part of the solution here might be to put firm limits on the period over which subscriptions can auto-renew. That would limit the upside to these practices. Is there any real societal benefit to subscriptions lasting years witho
2081.
by gpm 4y ago
I've gone to windy.com for weather, which really isn't what is was meant for, but on a large screen it works fine.
2082.
by gpm 4y ago
Per the code linked there is temperature (optional, default=0.1): Controls the randomness of the generated text. Higher values produce more diverse results, while lower values produce more deterministic results. top_p (optional, default=0.7
2083.
by gpm 4y ago
If I think of everything I use electricity for these days, lighting* is pretty far down the list of usefulness. The sun can fill 80% of my lighting needs, giving up on the 20% really doesn't seem like it would be that painful. I actual
2084.
by gpm 4y ago
Someone on lobste.rs mentioned quay.io (maintained by redhat) as an alternative to dockerhub. Anyone have any experience with it?
2085.
by gpm 4y ago
> During that period you will maintain access to any of your public images. Pretty unambiguously means "after that period you may not have access to any of your public images".
2086.
by gpm 4y ago
> All the magic of "7B LLaMA running on a potato" seems to involve lowering precision down to f16 LLaMa weights are f16s to start out with, no lowering necessary to get to there. You can stream weights from RAM to the GPU prett
2087.
by gpm 4y ago
There's a long distance between "the pentagon is involved in a military operation (or even war)" and "the defense production act is invoked". The pentagons role is not solely to prepare for total war scenarios.
2088.
by gpm 4y ago
Quantizing is a lossy process, you can't really claim to be running the 65B model llama at that point (though the 65b qgpt-llama does look like it might be very useful)
2089.
by gpm 4y ago
Yes (I just don't have that much ram) I have a separate branch that streams weights from ram - at which point I think I was only seeing negligible performance loss compared to storing the weights in vram. The bottleneck was compute, no
2090.
by gpm 4y ago
> Brex's offer is collaterized up to 25 cents per dollar. Do you know this? That's a number I (and some other posters) pulled out of a hat as a reasonable thing to do, but I don't think any of us had any sources for it?
2091.
by gpm 4y ago
Huffman encoding the weights (treating each 16bit float a symbol) could reduce the weights size to ~85% the original (I calculated this exactly before, but am going from memory). You could maybe get a bit more than that with arithmetic enco
2092.
by gpm 4y ago
I mean, I'm just going off the FDIC's assertion that > As of December 31, 2022, Silicon Valley Bank had approximately $209.0 billion in total assets and about $175.4 billion in total deposits. And I'm not trying to claim t
2093.
by gpm 4y ago
My code for this is very much not high quality, but I have a CPU + GPU + SSD combination: https://github.com/gmorenz/llama/tree/ssd Usage instructions in the commit message: https://github.com&#x2F
2094.
by gpm 4y ago
Assume your guesstimate of 50% is right. If I lend you 25% of your SVB deposit secured against your SVB deposit, I should still be fine (as long as my claim comes early enough in any ensuing bankruptcy).
2095.
by gpm 4y ago
> Can't fault Brex, there is no downside to anyone except them. Not actually clear who is taking the downside > This credit line is funded by 3rd-party capital (and not Brex directly)
2096.
by gpm 4y ago
All lending is a gamble, but if they only lend 25% of what you have stored with SVB it's (approximately) a gamble that the FDIC will pay out more than 25c on the dollar. Which as lending goes is a pretty safe gamble.
2097.
by gpm 4y ago
> As of December 31, 2022, Silicon Valley Bank had approximately $209.0 billion in total assets and about $175.4 billion in total deposits. Per the FDIC
2098.
by gpm 4y ago
Yes, but the assets of the balance sheet exceed the liabilities by a reasonable percentage. The vast majority of what will be paid out won't be from the FDIC's funds, it will be from those assets.
2099.
by gpm 4y ago
Honestly this makes a ton of sense. The FDIC will be returning most (or all) of the deposits in SVB, so the debt is reasonably safe (at least as far as debt that startups take on ever is). Whoever is funding this is probably not taking on a
2100.
by gpm 4y ago
This title strikes me as incorrect. The user paid for the mailbox as part of the bundle they purchased with their domain name. It was not free.

More ›