Yuichi Ishikawa
@yuichiis.bsky.social
Pinned
Yuichi Ishikawa
@yuichiis.bsky.social
· Mar 22
Neural Machine Translation with Transformer Models in PHP : R...
rindow.github.io
I finally published one new document.
"Neural Machine Translation Using Transformers in PHP"
There should be 16 more pages of new content 😭
rindow.github.io/neuralnetwor...
"Neural Machine Translation Using Transformers in PHP"
There should be 16 more pages of new content 😭
rindow.github.io/neuralnetwor...
gSDE isn't working well, so I'm taking a break from development work. However, I've realized that gSDE is necessary😅
October 2, 2025 at 10:34 PM
gSDE isn't working well, so I'm taking a break from development work. However, I've realized that gSDE is necessary😅
I've successfully implemented SAC for ContinuousMountainCar. Now that I've organized the code, I'm trying to decide what to do next: either add GPU support or create a parallel execution runner.
August 27, 2025 at 1:20 PM
I've successfully implemented SAC for ContinuousMountainCar. Now that I've organized the code, I'm trying to decide what to do next: either add GPU support or create a parallel execution runner.
However, GPT-5 is becoming good for using as a second opinion on complex theories, not for code generation. It's quite good because it hallucinates less. The only downside is the 10-message-a-day limit.😅
August 26, 2025 at 12:45 AM
However, GPT-5 is becoming good for using as a second opinion on complex theories, not for code generation. It's quite good because it hallucinates less. The only downside is the 10-message-a-day limit.😅
I normally use Gemini 2.5 Pro for coding. It's great for open-source development with no sensitive information because it's free and unlimited. Free versions of ChatGPT and Claude are practically useless.
August 26, 2025 at 12:40 AM
I normally use Gemini 2.5 Pro for coding. It's great for open-source development with no sensitive information because it's free and unlimited. Free versions of ChatGPT and Claude are practically useless.
I'm having trouble with my SAC implementation. I have to sample from a Gaussian distribution within the pipeline. I was intentionally avoiding this with PPO, but it seems unavoidable in my SAC implementation.
August 25, 2025 at 9:17 AM
I'm having trouble with my SAC implementation. I have to sample from a Gaussian distribution within the pipeline. I was intentionally avoiding this with PPO, but it seems unavoidable in my SAC implementation.
I implemented masked actions with PPO. I also managed to implement continuous actions for A2C, which I had previously given up on.
August 23, 2025 at 7:59 AM
I implemented masked actions with PPO. I also managed to implement continuous actions for A2C, which I had previously given up on.
I've completed the implementation for continuous actions in PPO and am now working on masked actions. I'm starting my investigation there because masking isn't working properly with REINFORCE.
August 13, 2025 at 11:01 PM
I've completed the implementation for continuous actions in PPO and am now working on masked actions. I'm starting my investigation there because masking isn't working properly with REINFORCE.
I've almost finished the PPO implementation. It's been trained on discrete actions, but I know that PPO struggles with continuous actions.
August 2, 2025 at 12:36 PM
I've almost finished the PPO implementation. It's been trained on discrete actions, but I know that PPO struggles with continuous actions.
I'm revamping the Runner for Proximal Policy Optimization (PPO) in reinforcement learning.
July 29, 2025 at 10:50 PM
I'm revamping the Runner for Proximal Policy Optimization (PPO) in reinforcement learning.
I've given up on learning continuous actions with A2C😭
July 26, 2025 at 2:39 AM
I've given up on learning continuous actions with A2C😭
Gemini CLI. It's been thinking for an hour with no response, so it's probably okay to cancel, right? 😂
July 24, 2025 at 10:33 PM
Gemini CLI. It's been thinking for an hour with no response, so it's probably okay to cancel, right? 😂
I implemented A2C for discrete actions, but I'm not sure how to correctly handle continuous actions. None of the Python A2C samples for training Pendulum seem to work properly.
July 17, 2025 at 11:37 PM
I implemented A2C for discrete actions, but I'm not sure how to correctly handle continuous actions. None of the Python A2C samples for training Pendulum seem to work properly.
I implemented DDPG in my reinforcement learning framework. I'm just one step away from getting it back to its state three years ago.
July 3, 2025 at 9:00 PM
I implemented DDPG in my reinforcement learning framework. I'm just one step away from getting it back to its state three years ago.
Original Japanese emojis, the origin of emojis, are gone.
www.docomo.ne.jp/info/notice/...
They lacked skin tone concepts; faces/hands were abstract characters.
iPhone introduced skin tones, affirming division. What is diversity?
www.docomo.ne.jp/info/notice/...
They lacked skin tone concepts; faces/hands were abstract characters.
iPhone introduced skin tones, affirming division. What is diversity?
ドコモからのお知らせ : ドコモ絵文字の提供終了について | お知らせ | NTTドコモ
昨今の端末の絵文字の利用状況を鑑みドコモ絵文字の提供を2025年6月下旬以降に発売する機種から終了いたします。
www.docomo.ne.jp
June 15, 2025 at 3:45 PM
Original Japanese emojis, the origin of emojis, are gone.
www.docomo.ne.jp/info/notice/...
They lacked skin tone concepts; faces/hands were abstract characters.
iPhone introduced skin tones, affirming division. What is diversity?
www.docomo.ne.jp/info/notice/...
They lacked skin tone concepts; faces/hands were abstract characters.
iPhone introduced skin tones, affirming division. What is diversity?
To be honest, today's announcement was essentially 'The user interface will look different. That's all.'🍎
June 9, 2025 at 11:06 PM
To be honest, today's announcement was essentially 'The user interface will look different. That's all.'🍎
Each reinforcement learning algorithm has subtle differences in network models, making abstraction quite difficult.
May 25, 2025 at 8:15 PM
Each reinforcement learning algorithm has subtle differences in network models, making abstraction quite difficult.
I'm reading reinforcement learning code I haven't touched in over 2 years. The surrounding environment has changed so much that fundamental revisions are required.
May 10, 2025 at 12:13 AM
I'm reading reinforcement learning code I haven't touched in over 2 years. The surrounding environment has changed so much that fundamental revisions are required.
Feeling exhausted, I ended up completely relying on AI to finish writing. Gemini 2.5 Pro is just too good, I can't help but depend on it.
May 8, 2025 at 9:29 AM
Feeling exhausted, I ended up completely relying on AI to finish writing. Gemini 2.5 Pro is just too good, I can't help but depend on it.
I've uploaded a significantly rewritten mathematics section document. There are probably still various mistakes. And the GPU part remains untouched😭.
May 6, 2025 at 5:53 AM
I've uploaded a significantly rewritten mathematics section document. There are probably still various mistakes. And the GPU part remains untouched😭.
Considering officially releasing a Tensor type that allows operator overriding.
Is there a need for a matrix operation that can be written as $c = $a + $b$?
Is there a need for a matrix operation that can be written as $c = $a + $b$?
May 3, 2025 at 9:24 AM
Considering officially releasing a Tensor type that allows operator overriding.
Is there a need for a matrix operation that can be written as $c = $a + $b$?
Is there a need for a matrix operation that can be written as $c = $a + $b$?
I was slacking off on document creation and instead created an automatic build for PHP extensions 😅
Successfully achieved automatic building of 20 different types of binaries for various PHP versions and OS environments 🤩
Successfully achieved automatic building of 20 different types of binaries for various PHP versions and OS environments 🤩
May 2, 2025 at 8:03 PM
I was slacking off on document creation and instead created an automatic build for PHP extensions 😅
Successfully achieved automatic building of 20 different types of binaries for various PHP versions and OS environments 🤩
Successfully achieved automatic building of 20 different types of binaries for various PHP versions and OS environments 🤩
Let's move on for now and leave the system of complex equations for later.
April 23, 2025 at 1:17 AM
Let's move on for now and leave the system of complex equations for later.
Writing math library docs is harder than I thought. Deciding which functions to publish is tough, and I even misunderstood one! Re-studying now. Building a fundamental library is really challenging. 😅
April 20, 2025 at 11:18 PM
Writing math library docs is harder than I thought. Deciding which functions to publish is tough, and I even misunderstood one! Re-studying now. Building a fundamental library is really challenging. 😅
Using Gemini 2.5 Pro for free right now and its coding ability is amazing. Wondering how much better OpenAI's o3 is. Will I end up needing a paid subscription? 😅
April 17, 2025 at 1:18 AM
Using Gemini 2.5 Pro for free right now and its coding ability is amazing. Wondering how much better OpenAI's o3 is. Will I end up needing a paid subscription? 😅