HNHacker News
TopNewBestAskShowJobs

GaggiX

4,957 karma · joined September 7, 2021

submissionscomments
GaggiX··on 'World Models,' an old idea in AI, mount a comeback
>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 really true, just think of Go.

GaggiX··on Cloudflare Radar: AI Insights
Character.AI is extremely popular among youngers so it's not really surprising.
GaggiX··on Anduril's product engineering machine
What's the acceptable state of mood that you are allow to be in when building killing machines?
GaggiX··on Are OpenAI and Anthropic losing money on inference?
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.

GaggiX··on Are OpenAI and Anthropic losing money on inference?
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.
GaggiX··on Are OpenAI and Anthropic losing money on inference?
> (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 decoding phase.

GaggiX··on Are OpenAI and Anthropic losing money on inference?
That's why they have the batch tier: https://platform.openai.com/docs/guides/batch
GaggiX··on Are OpenAI and Anthropic losing money on inference?
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 on a single batch, it would not make sense even in the decode phase.

GaggiX··on The Math Behind GANs (2020)
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 several losses, usually L1/L2, LPIPS and adversarial.
GaggiX··on Gemini 2.5 Flash Image
"Google AI Studio" and select the model
GaggiX··on Gemini 2.5 Flash Image
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 image" even though on the AI studio it said something different.

GaggiX··on Making games in Go: 3 months without LLMs vs. 3 days with LLMs
>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.

GaggiX··on The AI vibe shift is upon us
>technology that has never proven its worth outside of specious hype

Reading stuff like this makes me question the entirety of the article.

GaggiX··on Europe's crusade against air conditioning is insane
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 not just Germany making a dumb decision before)
GaggiX··on Europe's crusade against air conditioning is insane
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.
GaggiX··on In a first, Google has released data on how much energy an AI prompt uses
>~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're not accounting for batches for the optimal gpu utilization, maybe it can takes 30 seconds but it completed 30 requests.

GaggiX··on Gemma 3 270M re-implemented in pure PyTorch for local tinkering
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)
GaggiX··on What could have been
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?
GaggiX··on What could have been
>"while delivering absolutely nothing of value"

Well maybe for you and not the millions of people that use this technology daily.

GaggiX··on Llama-Scan: Convert PDFs to Text W Local LLMs
>Sub-2010 level OCR

It's not.

GaggiX··on Llama-Scan: Convert PDFs to Text W Local LLMs
>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.

GaggiX··on OpenAI Progress
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 for coding).
GaggiX··on Gemma 3 270M: Compact model for hyper-efficient AI
This model was trained on 6T tokens and has 256k embeddings, quite different than a gpt2 model comparable in size.
GaggiX··on I used to know how to write in Japanese
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 passionate about it, eventually you will pick up the readings of most common ones and you will be able to guess new words automatically, writing them no thank you, I barely write in my own language without a keyboard.
GaggiX··on LLMs aren't world models
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.

GaggiX··on How I code with AI on a budget/free
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 saying: vocabulary, grammar points, nuances).
GaggiX··on I bought a £16 smartwatch just because it used USB-C
You can also use a "dumb" watch that has 10 years battery life.
GaggiX··on Ask HN: How can ChatGPT serve 700M users when I can't run one GPT-4 locally?
Huge batches to find the perfect balance between compute and memory banthwidth, quantized models, speculative decoding or similar techniques, MoE models, routing of requests on smaller models if required, batch processing to fill the GPUs when demand is lower (or electricity is cheaper).
GaggiX··on Vibechart
GPT-5 models are actually great models for the API, the nano model is finally good enough to handle complex structured responses and it's even cheaper than GPT-4.1-nano.
GaggiX··on GPT-5: Key characteristics, pricing and system card
Compared to GPT-4, it is on a completely different level given that it is a reasoning model so on that regard it does delivers and it's not just scaling, but for this I guess the revolution was o1 and GPT-5 is just a much more mature version of the technology.
← PreviousPage 8 of 34Next →