renonce

15 ms·

181.
by renonce 3y ago
If people can somehow keep cryptocurrency values high, they can take USD margin loans out of their Bitcoin holdings, which increases the amount of total dollars laying around. That’s what happened in 2021 that bumped crypto asset prices due
182.
by renonce 3y ago
That’s more like an inherent limitation of autoregressive prediction than that of LLM. Maybe LLMs can be trained or finetuned in other ways that allows it to think before answering.
183.
by renonce 3y ago
So is AGI AI? Why use such a broad term that puts AGI, a term defined by science fiction, with Transformer, a practical next token predictor based on gradient descent and attention mechanism, in the same basket?
184.
by renonce 3y ago
It’s Hacker News so you can expect something interesting even if it looks like a clickbait.
185.
by renonce 3y ago
The letters have to be in order so maybe this will not work?
186.
by renonce 3y ago
> Technical info: Booleans are stored as one-and-a-half bits. No you can’t store three choices in one-and-a-half bits. Two booleans would be 9 choices whereas 3 bits can store only 8 choices.
187.
by renonce 3y ago
Most Bitcoin miners have been using ASIC chips that compute nothing but Bitcoin format SHA256 since long ago so it’s not increasing demand for GPUs. Ethereum is but it has already switched to PoS.
188.
by renonce 3y ago
Large scale sets are only needed for training. For inference, 8x NVIDIA A100 80G will allow inference for 300b models (GPT-3 is 175b) or 1200b models with 4-bit quantization (quantization impact is negligible for large models), so a single
189.
by renonce 3y ago
Apart from poe.com there is also nat.dev. It even supports Claude-100K. Just pay $5 and it will bill by API pricing, proportional to number of tokens.
190.
by renonce 3y ago
Doesn’t matter if you only use it yourself. No one will know.
191.
by renonce 3y ago
A big chunk of developments based on Facebook’s LLaMA model are by small-time developers and individuals, not large players. Facebook has already shown a viable way to release models in the way you described.
192.
by renonce 3y ago
Does the author refer to GPT-4 API or the ChatGPT version of GPT-4? The recent discussions on GPT-4 quality deterioration seem to focus on the ChatGPT version rather than the API. Also, since ChatGPT version of GPT-4 is now supporting web b
193.
by renonce 3y ago
What is the difference between a supercomputer and a million ordinary servers connected together?
194.
by renonce 3y ago
If you bought NVDA stock at ChatGPT launch you would be super rich today. Not what the GPT influencers would tell you, though.
195.
by renonce 3y ago
You can’t quantize it for training due to numerical instability. For inference you don’t usually use such a big cluster.
196.
by renonce 3y ago
I suppose it would be 12 bytes? 4 bytes for base model, 4 bytes for optimizer momentum and 4 bytes for optimizer second moment EWA.
197.
by renonce 3y ago
For many applications there is no need to fully model the human brain. An approximation of a particular aspect would be sufficient in most cases. We didn’t build aeroplanes by fully modeling a bird, we just need aerodynamics.
198.
by renonce 3y ago
If TPUv4 pods were so powerful, why are most new models trained with NVIDIA cards rather than TPU?
199.
by renonce 3y ago
Are you sure your app is used just by people in your city rather than abused by projects like gpt4free (where they use your API for purposes other than chatting)?
200.
by renonce 3y ago
Is that really possible? I don't understand the field exactly but at least if you use deep learning, you'll be able to recover inputs from gradients, and if you want to program it such that no other party is able to recover your i
201.
by renonce 3y ago
I don't exactly see how to define "3x as intelligent as a single person" so I'll conveniently define it as something that thinks 3x faster than a single person. A single such thing can talk to 3 persons simultaneously, a
202.
by renonce 3y ago
"We intend to offer a free tier for low-volume users." I would expect that users using below a certain amount, like $3, probably do not need to pay at all.
203.
by renonce 3y ago
Does that work in batch so I can upload, like, 1000 photos at a time? That's what I expect from an API.
204.
by renonce 3y ago
My computers are behind firewalls and I need Tailscale to do the NAT punching. I don't see another tool that does the job as well as Tailscale.
205.
by renonce 3y ago
Has binary neural network been implemented for Transformers yet?
206.
by renonce 3y ago
Only if you use gpt-4. gpt-3.5-turbo is much faster, and gpt-4 is only going to get faster as GPUs get faster.
207.
by renonce 3y ago
Everyone would know it and the attestation key would be obsolete. New wallets will be made with another key, and for old wallets users already know they are genuine anyway.
208.
by renonce 3y ago
It does uncover a vulnerability about Trezor that allows attackers to fake a Trezor without the user knowing it. It should have been defended via attestation, and software downloaded from the official website should have checked the attesta
209.
by renonce 3y ago
How can a mobile phone's network be as stable as that of a datacenter? Without that stability it's not going to be practically useful.
210.
by renonce 3y ago
At least it’s a fair game here. It could have been unfair if, for example, the audience cannot verify the truthfulness of the article, which would promote cheap fake but interesting articles.

More ›