HNHacker News
TopNewBestAskShowJobs

GaggiX

4,957 karma · joined September 7, 2021

submissionscomments
GaggiX··on Pocket TTS: A high quality TTS that gives your CPU a voice
I love that everyone is making their own TTS model as they are not as expensive as many other models to train. Also there are plenty of different architecture.

Another recent example: https://github.com/supertone-inc/supertonic

GaggiX··on Iran official says 2k people have been killed in unrest
For a comparison Iran reported "200 killed" in the Mahsa Amini protests.
GaggiX··on Iran's complete Internet shutdown reaches 24 hours
I agree, this really shows how bad things have gotten for the regime.

Iran relies heavily on digital transactions because of the hyperinflation, there is not enough physical cash anymore. Soon people are going to the street simply because they can't buy food anymore.

GaggiX··on Iran Goes Into IPv6 Blackout
When the internet will be working again I will show your comment to my friend in Iran, I think he will find it hilarious.
GaggiX··on Iran Goes Into IPv6 Blackout
Iran has shut down the internet nationwide several times before, this is nothing new when there are protests.
GaggiX··on Iran Goes Into IPv6 Blackout
They are almost completely inaccessible to the average Iranian. A friend of mine who has come a long way to fight Iranian censorship told me that they essentially don't exist.
GaggiX··on Lua 5.5
I managed to run Balatro on a Trimui Brick on a very minimalist Linux distro, steamos should be easy.
GaggiX··on What Is (AI) Glaze?
From what I remember Glaze is using some small CLIP model and LPIPS (based on VGG) for their adversarial loss, that's why it's so ineffective to large, better trained model.

It use SD to do a style transfer on the image using image-to-image, then it use gradient descent on the image itself to lower the difference between CLIP embeddings of the original and style transfer image + trying to maintain LPIPS, then every step is normalized to not exceed a certain threshold from the original image.

So essentially it's an adversarial attack against a small CLIP model, even though today's models are much robust than that.

GaggiX··on On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs
The people talking about semantics in the comment section seems to completely ignore the positive correlation of LLMs between accuracy and stated confidence, this is called calibration and this "old" blog post from a year ago already showed it, LLMs can know what they know: https://openai.com/index/introducing-simpleqa/
GaggiX··on Measuring AI Ability to Complete Long Tasks
Opus: "an artistic work, especially one on a large scale."

The names Haiku, Sonnet, and Opus have not been chosen randomly.

GaggiX··on Making Google Sans Flex
I never saw the cursor changing size to fit the button you are hovering on, it's pretty cool, I don't know if it's better but it's cool.
GaggiX··on T5Gemma 2: The next generation of encoder-decoder models
≥In general encoder+decoder models are much more efficient at infererence than decoder-only models because they run over the entire input all at once (which leverages parallel compute more effectively).

Decoder-only models also do this, the only difference is that they use a masked attention.

GaggiX··on Gemini 3 Flash: Frontier intelligence built for speed
>or some sort of cracked way to pack a ton of parametric knowledge into a Flash Model.

More experts with a lower pertentage of active ones -> more sparsity.

GaggiX··on Gemini 3 Flash: Frontier intelligence built for speed
It still supports the legacy mode of setting the budget, you can set it to 0 and it would be equivalent to none reasoning effort like gpt 5.1/5.2
GaggiX··on Gemini 3 Flash: Frontier intelligence built for speed
They went too far, now the Flash model is competing with their Pro version. Better SWE-bench, better ARC-AGI 2 than 3.0 Pro. I imagine they are going to improve 3.0 Pro before it's no more in Preview.

Also I don't see it written in the blog post but Flash supports more granular settings for reasoning: minimal, low, medium, high (like openai models), while pro is only low and high.

GaggiX··on Japan to revise romanization rules for first time in 70 years
Japan has an extremely high literacy rate.
GaggiX··on Hashcards: A plain-text spaced repetition system
Also Anki is not ugly at all in general, the interface of Hashcards looks much uglier to me.
GaggiX··on Hashcards: A plain-text spaced repetition system
Anki is amazing, "anal_reactor" I think you're wrong on this.

Also "half an hour" != "slow learner", everything depends on the quantity, the difficulty and the chosen desired retention.

GaggiX··on OpenAI are quietly adopting skills, now available in ChatGPT and Codex CLI
When there is no alt text do you have like a solution for that? Like VLMs are really powerful, I imagine they can be used to parse through the unlabeled images automatically if needed.
GaggiX··on Auto-grading decade-old Hacker News discussions with hindsight
It became a unicorn in December 2019 tho, 4 years later.
GaggiX··on Auto-grading decade-old Hacker News discussions with hindsight
I was reading the Anki article on 2015-12-13, and the best prediction was by markm248 saying: "Remember that you read it here first, there will be a unicorn built on the concept of SRS"

They were right, Duolingo.

GaggiX··on Auto-grading decade-old Hacker News discussions with hindsight
I think the most fun thing is to go to: https://karpathy.ai/hncapsule/hall-of-fame.html

And scroll down to the bottom.

GaggiX··on Show HN: Gemini Pro 3 imagines the HN front page 10 years from now
You haven't used Gemini 3.0 Pro if so, this is very much in its range of abilities.
GaggiX··on Show HN: Gemini Pro 3 imagines the HN front page 10 years from now
The gary marcus post about AI progress stalling made me laugh. I think that would be accurate.
GaggiX··on Horses: AI progress is steady. Human equivalence is sudden
ChatGPT was released 3 years ago and that was complete ass compared to what we have today.
GaggiX··on Z-Image: Powerful and highly efficient image generation model with 6B parameters
>So what's the nitpicking about?

Idk I just thought it was funny to read the ignorant comment that called the Chinese model useless because it rendered Chinese text and calling it Japanese. The model is trained to render English or Chinese text.

GaggiX··on Z-Image: Powerful and highly efficient image generation model with 6B parameters
>manga-influenced comic strips with kanji. Useless.

Are you sure it was Japanese? Because the model is Chinese so it's likely to output Chinese (it happened in my testing).

GaggiX··on Kilauea erupts, destroying webcam [video]
>A synthesized text-to-video voiceover was used in the narration for this story.

I wasn't even realizing it without reading this in the description.

GaggiX··on Z-Image: Powerful and highly efficient image generation model with 6B parameters
I didn't even know seedream 4.5 has been released, things move fast, I have used seedream 4 a lot through their API.
GaggiX··on The writing is on the wall for handwriting recognition
Are you sure to have used the Gemini 3.0 pro model? Maybe try increasing the media resolution on the AI studio if the text is small
← PreviousPage 6 of 34Next →