GaggiX

13 ms·

211.
by GaggiX 1y ago
A model is capable of learning the "calibration" during the reinforcement learning phase, in this "old" post from OpenAI: https://openai.com/index/introducing-simpleqa/ you can see the positive
212.
by GaggiX 1y ago
They also beat Apple to market a smartwatch with microLED.
213.
by GaggiX 1y ago
I know nothing about this technology/service but I do hope it would help avoiding the censorship from Mastercard/Visa.
214.
by GaggiX 1y ago
>Fwiw nothing beats ‘implement the game logic in full (huge amounts of work) and with pruning on some heuristics look 50 moves ahead’. This is how chess engines work and how all good turn based game ai works. Just read the parent comment
215.
by GaggiX 1y ago
>This is how chess engines work All strongest chess engine have at least one neural network to evaluate positions, including Stockfish, and this impact the searching window. >how all good turn based game ai works That's not reall
216.
by GaggiX 1y ago
Character.AI is extremely popular among youngers so it's not really surprising.
217.
by GaggiX 1y ago
What's the acceptable state of mood that you are allow to be in when building killing machines?
218.
by GaggiX 1y ago
https://openrouter.ai/deepseek/deepseek-chat-v3.1 They are dirt cheap. Same model architecture for the comparison: $0.30/M $1.00/M. Or even $0.20-$0.80 from another provider.
219.
by GaggiX 1y ago
The "efficiency" meantioned in blog post you have linked is the price difference between Deepseek and o1, it doesn't mean that GPT-5 or other SOTA models are less efficient.
220.
by GaggiX 1y ago
> (And actually during generation you can use speculative decoding to do better than this roofline anyways). And more importantly batches, so taking the example from the blog post, it would be 32 tokens per each forward pass in the decod
221.
by GaggiX 1y ago
That's why they have the batch tier: https://platform.openai.com/docs/guides/batch
222.
by GaggiX 1y ago
Your calculations make no sense. Why are you loading the model for each token independently? You can process all the input tokens at the same time as long as they can fit in memory. You are doing the calculation as they were output tokens o
223.
by GaggiX 1y ago
Adversarial loss is still use on most image generators, diffusion/autoregressive models work on a latent space (they don't have to, but it would incredibly inefficient) created by an autoencoder, these autoencoders are trained on
224.
by GaggiX 1y ago
"Google AI Studio" and select the model
225.
by GaggiX 1y ago
An image seems to be 256 tokens looking the AIstudio tab, so you can generate 3906,25 images per 1M tokens, that seems a lot if I'm not wrong in some ways. Edit: the blog post is now loading and reports "1290 output tokens per ima
226.
by GaggiX 1y ago
>As an LLM hater I thought this was the start of a joke or something, I guess if you use LLMs you are a "LLM lover" then.
227.
by GaggiX 1y ago
>technology that has never proven its worth outside of specious hype Reading stuff like this makes me question the entirety of the article.
228.
by GaggiX 1y ago
Germany also was the biggest country for solar energy, what it's essentially making a huge chunk of the energy production in summer in Europe, what started the industrialization of solar panels, making them dirt cheap. (Also Europe is
229.
by GaggiX 1y ago
This article is so stupid for not mentioning energy costs. It's like Europeans don't use AC because they want to degrowth or something.
230.
by GaggiX 1y ago
>~If they're running on, say, two RTX 6000s for a total draw of ~600 watts, that would be a response time of 1.44 seconds. So obviously the median prompt doesn't go to some high-end thinking model users have to pay for. You&#x2
231.
by GaggiX 1y ago
I imagine you and your team have finetuned the model on different tasks, can you share some results? (I have only seen the alien NPC finetuning)
232.
by GaggiX 1y ago
So you're implying that ChatGPT and similar are so popular because of "errors, defects, inefficiencies, or poor service delivery", how does this make any sense in the context?
233.
by GaggiX 1y ago
>"while delivering absolutely nothing of value" Well maybe for you and not the millions of people that use this technology daily.
234.
by GaggiX 1y ago
>Sub-2010 level OCR It's not.
235.
by GaggiX 1y ago
>are cheap and strong enough to make this practical. It all depends on the scale you need them, with the API it's easy to generate millions of tokens without thinking.
236.
by GaggiX 1y ago
the actual major leap was o1, going from 3.5 to 4 is just scaling, o1 is a different paradigm that skyrocketed its performance on math/physics problems (or reasoning more generally), it also made the model much more precise (essential
237.
by GaggiX 1y ago
This model was trained on 6T tokens and has 256k embeddings, quite different than a gpt2 model comparable in size.
238.
by GaggiX 1y ago
That's why I focus on learning words rather than individual kanji characters. Kanji characters have so many different readings, and the Japanese language is full of exceptions that it's not worth the time unless you're passio
239.
by GaggiX 1y ago
https://www.youtube.com/watch?v=LtG0ACIbmHw Sota LLMs do play legal moves in chess, I don't why the article seem to say otherwise.
240.
by GaggiX 1y ago
OpenAI offering 2.5M free tokens daily small models and 250k for big ones (tier 1-2) is so useful for random projects, I use them to learn japanese for example (by having a program that list informations about what the characters are just s

More ›