chessgecko

19 ms·

61.
by chessgecko 3y ago
but if the number of hours you need worked is constant than its cheaper to hire a new worker at 1x wage for that extra hour than the same worker at 1.5x
62.
by chessgecko 3y ago
Doesn’t that graph have a toks/sec of 18? Or am I reading it wrong
63.
by chessgecko 3y ago
Is immortality a genetic advantage? I'd imagine it could create a lot of infighting.
64.
by chessgecko 3y ago
And apparently they have so many employees the auditorium isn’t big enough anymore so there’s some spillover for all staff meetings
65.
by chessgecko 3y ago
TSMC is cheap because of the risk of invasion, otherwise they would have a pretty insane valuation already.
66.
by chessgecko 3y ago
After what happened with openai I wouldn’t say no scenario
67.
by chessgecko 3y ago
I'd hope that people would run good-ai before pushing or a white hat vuln finder before adding packages. Not sure people will or if it will be as good, but it will probably be available.
68.
by chessgecko 4y ago
Openai doesn’t want share features it would make it to easy to scrape a dataset of gpt4 answers
69.
by chessgecko 4y ago
I wonder what led to such a gap between llama 7b and Cerebras 13b. I hope they discuss it in the paper.
70.
by chessgecko 4y ago
The botnets don’t need this, if they can’t get access to gpt3/4 they’d probably just rent some a100s. You can make so much blogspam in an hour with 8 a100s
71.
by chessgecko 4y ago
People are speculating that gpt3.5 turbo is actually much smaller and that they are very likely currently making a profit on it. It seems likely just given how quickly some of the 3.5 turbo responses are from the api, and how much they push
72.
by chessgecko 4y ago
The gap in quality between gpt2 and gpt3 was much much larger than the gap between gpt3 and 3.5. If they could get the gap between 2 -> 3 again when moving to 4 it would be crazy
73.
by chessgecko 4y ago
It seems like google is using search or some knowledge base to throw relevant information into the context when it's generating to get up to date results. It relies on knowing what to retrieve, but google is already pretty good at that
74.
by chessgecko 4y ago
GPT 3.5 is code for the model underlying davinci-text-003 and chatgpt (although there are some rumors chat is based on davinci-2).
75.
by chessgecko 4y ago
Maybe not super relevant but people used to do that for translation with RNNs. They reversed the input so that the hidden states at the end would be close to what was needed to start translating. Nobody really did it enc-dec transformers bu
76.
by chessgecko 4y ago
Yeah to get the working result I had to reword the problem a little. In my post I kinda assume you could fine tune a model to do that. It is a little cheap though.
77.
by chessgecko 4y ago
Kinda depends on how you define a transformer solving the problem. I feel like you could fine-tune a transformer in the style of https://www.ai21.com/blog/jurassic-x-crossing-the-neuro-symb... to produce steps to solve
78.
by chessgecko 4y ago
I'm a fan of the conspiracy that he's the fall guy here and it was on purpose to draw all the attention. Can't ham it too hard though or it'll get suspicious.
79.
by chessgecko 4y ago
Interesting that in this case the capital on the w seems to make a big difference. I ran it a few times with a capital W and a lowercase w, it said sailfish for the lowercase w most of the time and switched between an orca and a dolphin for
80.
by chessgecko 4y ago
It got the sea mammal question right for me. > What is the fastest sea mammal? > The fastest sea mammal is the dolphin. Dolphins are known for their speed and agility in the water and can swim at speeds of up to 45 miles per hour. The
81.
by chessgecko 4y ago
The trick is that they pretended they were compliant with us regulations when they weren’t. Some users probably don’t mind this, but if there are issues they might change their mind.
82.
by chessgecko 4y ago
Idk about gold, but I'd argue that both houses and art have discounted cash flows if you decide to rent them out (or maybe avoiding paying rent is a sort of cash flow for your budget)
83.
by chessgecko 4y ago
Can you afford to buy back the equity you would have accelerated? We had this happen and vested them, but the person didn’t really appreciate the equity at all, despite it ending up more valuable than the cash.
84.
by chessgecko 4y ago
Maybe it’s to speed up multi gpu matrix multiplies. They’re useful for serving/training gpt3 size models
85.
by chessgecko 4y ago
That’s true, but the free tier covers almost everything in our grammar checker, most of the premium features are on the paraphraser
86.
by chessgecko 4y ago
https://quillbot.com/grammar-check is an alternative for English. Not open source, but very little of the grammar checker is paywalled if the other options seem expensive. (Disclaimer I’m a founder)
87.
by chessgecko 4y ago
fp16 models inference just fine in fp32, though I was sorta joking in my original comment, it would potentially take weeks for this to run one input. You're better off trying to make something like huggingface accelerate work (like the
88.
by chessgecko 4y ago
That already exists depending on your definition of slow. Just get a big ssd, use it as swap and run the model on cpu.
89.
by chessgecko 5y ago
I was referring to the benchmark script from the article https://github.com/neuralmagic/deepsparse/blob/main/examples... It looks like they used torch.load instead of torch.jit.load so I don’t think it’s
90.
by chessgecko 5y ago
Super interesting, but it would probably help the pytorch performance significantly on both the gpu and cpu if they torchscripted the models, it's probably pretty simple given they exported it to onnx and would be more apples to apples

More ›