dahara1
@dahara1.bsky.social
I made machine translation with LLMs.
I made PC Chrome translation plugin for bluesky.
I made smart feed for bluesky.
I mada content import agent.
Let's improve these qualities!
I made PC Chrome translation plugin for bluesky.
I made smart feed for bluesky.
I mada content import agent.
Let's improve these qualities!
The AI revolution has begun.
September 4, 2025 at 4:47 PM
The AI revolution has begun.
Surprisingly, Gemini didn't get the joke just by looking at it, but Claude did.
The title is "Italians are furious."
The title is "Italians are furious."
August 26, 2025 at 3:01 PM
Surprisingly, Gemini didn't get the joke just by looking at it, but Claude did.
The title is "Italians are furious."
The title is "Italians are furious."
There are rumors that llama 4.1 and 4.2 are SLM.
August 15, 2025 at 8:58 AM
There are rumors that llama 4.1 and 4.2 are SLM.
A survey showed that performance did not drop significantly even when using a GPU via a VM
(AMD GPU)
github.com/sbnb-io/sbnb...
(AMD GPU)
github.com/sbnb-io/sbnb...
June 26, 2025 at 4:24 PM
A survey showed that performance did not drop significantly even when using a GPU via a VM
(AMD GPU)
github.com/sbnb-io/sbnb...
(AMD GPU)
github.com/sbnb-io/sbnb...
Magenta RealTime: An Open-Weights Live Music Model
Google Deepmind's music generation model that lets you feel like a DJ
It was fun to use.
magenta.withgoogle.com/magenta-real...
Google Deepmind's music generation model that lets you feel like a DJ
It was fun to use.
magenta.withgoogle.com/magenta-real...
June 21, 2025 at 3:45 PM
Magenta RealTime: An Open-Weights Live Music Model
Google Deepmind's music generation model that lets you feel like a DJ
It was fun to use.
magenta.withgoogle.com/magenta-real...
Google Deepmind's music generation model that lets you feel like a DJ
It was fun to use.
magenta.withgoogle.com/magenta-real...
I did some research a while back, but I think quantization definitely makes things worse than people thinking.
Do you have any verification results that show that there is no performance degradation in actual downstream tasks?
Do you have any verification results that show that there is no performance degradation in actual downstream tasks?
June 12, 2025 at 8:00 AM
I did some research a while back, but I think quantization definitely makes things worse than people thinking.
Do you have any verification results that show that there is no performance degradation in actual downstream tasks?
Do you have any verification results that show that there is no performance degradation in actual downstream tasks?
🧠Evolution of LLM Sampling
People who use relatively small local LLMs should keep this in mind.
I'm proud of myself for being able to organize things relatively neatly.
#MLSky
People who use relatively small local LLMs should keep this in mind.
I'm proud of myself for being able to organize things relatively neatly.
#MLSky
June 10, 2025 at 7:01 AM
🧠Evolution of LLM Sampling
People who use relatively small local LLMs should keep this in mind.
I'm proud of myself for being able to organize things relatively neatly.
#MLSky
People who use relatively small local LLMs should keep this in mind.
I'm proud of myself for being able to organize things relatively neatly.
#MLSky
Apparently a fraudulent book using Professor Hinton's name was being sold on Amazon
That's a pretty aggressive price tag.
That's a pretty aggressive price tag.
June 9, 2025 at 2:11 AM
Apparently a fraudulent book using Professor Hinton's name was being sold on Amazon
That's a pretty aggressive price tag.
That's a pretty aggressive price tag.
I learned that the wandb project name is publicly available to the world, and that my top-secret hyperparameter search is completely open.
#MLSky
#MLSky
May 17, 2025 at 7:16 AM
I learned that the wandb project name is publicly available to the world, and that my top-secret hyperparameter search is completely open.
#MLSky
#MLSky
The browser interface of llama.cpp can now upload PDFs.
It was a nail-biting event that was so short on resources(32GB system memory) that it was impossible to record a video, but the 25 page Gemma 3 Report was beautifully summarized in Gemma 3 12B Q4 without any hallucinations.
#MLSky
It was a nail-biting event that was so short on resources(32GB system memory) that it was impossible to record a video, but the 25 page Gemma 3 Report was beautifully summarized in Gemma 3 12B Q4 without any hallucinations.
#MLSky
May 16, 2025 at 5:48 AM
The browser interface of llama.cpp can now upload PDFs.
It was a nail-biting event that was so short on resources(32GB system memory) that it was impossible to record a video, but the 25 page Gemma 3 Report was beautifully summarized in Gemma 3 12B Q4 without any hallucinations.
#MLSky
It was a nail-biting event that was so short on resources(32GB system memory) that it was impossible to record a video, but the 25 page Gemma 3 Report was beautifully summarized in Gemma 3 12B Q4 without any hallucinations.
#MLSky
Why LLMs became popular?
April 22, 2025 at 10:04 AM
Why LLMs became popular?
There was also a slight price reduction.
The o3 is about 2/3 the input/output price of the o1, and 1/3 the cache input.
The o4-mini is half the cache input price of the o3-mini.
The difference between the o1-pro and the gpt-4.1-nano is a whopping 1500 times.
The o3 is about 2/3 the input/output price of the o1, and 1/3 the cache input.
The o4-mini is half the cache input price of the o3-mini.
The difference between the o1-pro and the gpt-4.1-nano is a whopping 1500 times.
April 16, 2025 at 6:12 PM
There was also a slight price reduction.
The o3 is about 2/3 the input/output price of the o1, and 1/3 the cache input.
The o4-mini is half the cache input price of the o3-mini.
The difference between the o1-pro and the gpt-4.1-nano is a whopping 1500 times.
The o3 is about 2/3 the input/output price of the o1, and 1/3 the cache input.
The o4-mini is half the cache input price of the o3-mini.
The difference between the o1-pro and the gpt-4.1-nano is a whopping 1500 times.
In the app version of X, if you press and hold an image/photo posted by someone else, a menu called Edit Photo With Grok will appear, allowing you to edit it.
Some illustrator deleted all of his previous illustrations that he had posted on X.
It encourages copying of artist works/personalities.
Some illustrator deleted all of his previous illustrations that he had posted on X.
It encourages copying of artist works/personalities.
April 16, 2025 at 10:43 AM
In the app version of X, if you press and hold an image/photo posted by someone else, a menu called Edit Photo With Grok will appear, allowing you to edit it.
Some illustrator deleted all of his previous illustrations that he had posted on X.
It encourages copying of artist works/personalities.
Some illustrator deleted all of his previous illustrations that he had posted on X.
It encourages copying of artist works/personalities.
Gemini 2.0 Flash = gpt 4.1 nano
April 15, 2025 at 2:39 AM
Gemini 2.0 Flash = gpt 4.1 nano
I thought This is very cheap, but it's only for labor costs, which means you have to prepare and send in two RTX 4090 yourself after paying $700.
It doesn't state on the page that this is a labor charge, so to be honest, this seems like a scam.
It doesn't state on the page that this is a labor charge, so to be honest, this seems like a scam.
April 13, 2025 at 11:43 AM
I thought This is very cheap, but it's only for labor costs, which means you have to prepare and send in two RTX 4090 yourself after paying $700.
It doesn't state on the page that this is a labor charge, so to be honest, this seems like a scam.
It doesn't state on the page that this is a labor charge, so to be honest, this seems like a scam.
I finetuned the LLM-based text2speech model (canopylabs/orpheus-3b) to Japanese.
It mimics the voice of a famous Japanese voice synthesis system, so it's not a natural sound, but it does mimic the voice.
However, the existing T2S system is very powerful, so I'm not sure whether to go any further.
It mimics the voice of a famous Japanese voice synthesis system, so it's not a natural sound, but it does mimic the voice.
However, the existing T2S system is very powerful, so I'm not sure whether to go any further.
April 12, 2025 at 7:09 AM
I finetuned the LLM-based text2speech model (canopylabs/orpheus-3b) to Japanese.
It mimics the voice of a famous Japanese voice synthesis system, so it's not a natural sound, but it does mimic the voice.
However, the existing T2S system is very powerful, so I'm not sure whether to go any further.
It mimics the voice of a famous Japanese voice synthesis system, so it's not a natural sound, but it does mimic the voice.
However, the existing T2S system is very powerful, so I'm not sure whether to go any further.
I sometimes make mistakes when posting on social media, so it might be a good idea to have Gemini check them.
Gemini 2.5 Pro explained the intention without any misunderstanding.
Gemini 2.5 Pro explained the intention without any misunderstanding.
April 8, 2025 at 10:02 AM
I sometimes make mistakes when posting on social media, so it might be a good idea to have Gemini check them.
Gemini 2.5 Pro explained the intention without any misunderstanding.
Gemini 2.5 Pro explained the intention without any misunderstanding.
This is the video that is said to be the source of the misinformation about postponing the introduction of tariffs for 90 days.
However, Gemini 2.5 Pro explained the intention without any misunderstanding.
However, Gemini 2.5 Pro explained the intention without any misunderstanding.
April 8, 2025 at 10:02 AM
This is the video that is said to be the source of the misinformation about postponing the introduction of tariffs for 90 days.
However, Gemini 2.5 Pro explained the intention without any misunderstanding.
However, Gemini 2.5 Pro explained the intention without any misunderstanding.
I saw this in real time
At the moment Nvidia stock and the S&P are turning green, but is anyone bullish on US stocks after seeing this?
At the moment Nvidia stock and the S&P are turning green, but is anyone bullish on US stocks after seeing this?
April 7, 2025 at 6:49 PM
I saw this in real time
At the moment Nvidia stock and the S&P are turning green, but is anyone bullish on US stocks after seeing this?
At the moment Nvidia stock and the S&P are turning green, but is anyone bullish on US stocks after seeing this?
It's true that the stock market crash isn't a big deal when you look at the Dow graph
But if you look at it from a more historical perspective...
But if you look at it from a more historical perspective...
April 7, 2025 at 10:23 AM
It's true that the stock market crash isn't a big deal when you look at the Dow graph
But if you look at it from a more historical perspective...
But if you look at it from a more historical perspective...