S. Ota
LightNews LightNews
banner
ota.bsky.social
S. Ota
@ota.bsky.social
470 followers 200 following 440 posts
Interests: Reinforcement Learning, Natural Language Processing and Artificial General Intelligence.

arXiv papers bot: @paper.bsky.social
Posts Replies Media Videos
ota.bsky.social
"Designed by Jr.Mars
Inspired by the sleek and elegant Light Mode of Apple's MacOS system, bring a touch of classic nostalgia to your keyboard."

keyreative.store/products/kap...
KAP_Retro Lights R2 Keycaps
Inspired by the iconic Light Mode of Apple's MacOS, KAP_Retro Lights R2 bring a touch of nostalgia with their charming dot icons. Upgrade your keyboard now!
keyreative.store
October 25, 2025 at 12:54 PM
ota.bsky.social
How to burn firmware to the Swagkeys Eave65

1. Install dfu-util
2. Plug in the USB-C cable
3. Press the reset button on the back of the PCB
4. Run `dfu-util -l`. It should show 3 targets (alt=0, 1, 2)
5. Run `dfu-util -d 1eaf:0003 -a 2 -D eave65.bin` (alt=2)
6. Unplug and plug the USB-C cable
September 26, 2025 at 6:55 PM
Reposted by S. Ota
paper.bsky.social
[30/30] 198 Likes, 6 Comments, 3 Posts
2509.06160, cs․AI | cs․CL, 07 Sep 2025

🆕Reverse-Engineered Reasoning for Open-Ended Generation

Haozhe Wang, Haoran Que, Qixin Xu, Minghao Liu, Wangchunshu Zhou, Jiazhan Feng, Wanjun Zhong, Wei Ye, Tong Yang, Wenhao Huang, Ge Zhang, Fangzhen Lin
While the ``deep reasoning'' paradigm has spurred significant advances in verifiable domains like mathematics, its application to open-ended, creative generation remains a critical challenge. The two dominant methods for instilling reasoning -- reinforcement learning (RL) and instruction distillation -- falter in this area; RL struggles with the absence of clear reward signals and high-quality reward models, while distillation is prohibitively expensive and capped by the teacher model's capabilities. To overcome these limitations, we introduce REverse-Engineered Reasoning (REER), a new paradigm that fundamentally shifts the approach. Instead of building a reasoning process ``forwards'' through trial-and-error or imitation, REER works ``backwards'' from known-good solutions to computationally discover the latent, step-by-step deep reasoning process that could have produced them. Using this scalable, gradient-free approach, we curate and open-source DeepWriting-20K, a large-scale dataset of 20,000 deep reasoning trajectories for open-ended tasks. Our model, DeepWriter-8B, trained on this data, not only surpasses strong open-source baselines but also achieves performance competitive with, and at times superior to, leading proprietary models like GPT-4o and Claude 3.5.
September 14, 2025 at 12:06 AM
ota.bsky.social
"chat.fontSize" and "chat.fontFamily" for GitHub Copilot Chat.

github.com/microsoft/vs...
Chat - add fontSize and fontFamily settings by lszomoru · Pull Request #263299 · microsoft/vscode
github.com
September 1, 2025 at 4:57 PM
ota.bsky.social
GRPO for gpt-oss-20b with verl and sglang

github.com/volcengine/v...
github.com
August 29, 2025 at 3:21 PM
ota.bsky.social
A useful table to convert Slurm scripts to ABCI (PBS) and TSUBAME (SGE/AGE).

"This table lists the most common command, environment variables, and job specification options used by the major workload management systems: PBS/Torque, Slurm, LSF, SGE and LoadLeveler."

slurm.schedmd.com/rosetta.html
Slurm Workload Manager - Rosetta Stone of Workload Managers
slurm.schedmd.com
August 29, 2025 at 3:04 PM
ota.bsky.social
"Unsloth gpt-oss fine-tuning is 1.5x faster, uses 70% less VRAM, and supports 10x longer context lengths. gpt-oss-20b LoRA training fits on a 14GB VRAM, and gpt-oss-120b works on 65GB VRAM."

docs.unsloth.ai/basics/gpt-o...
gpt-oss: How to Run & Fine-tune | Unsloth Documentation
Run & fine-tune OpenAI's new open-source models!
docs.unsloth.ai
August 10, 2025 at 4:50 AM
ota.bsky.social
"Collection of scripts demonstrating different optimization and fine-tuning techniques for OpenAI's GPT-OSS models (20B and 120B parameters).

...

For full-parameter training on one node of 8 GPUs, ..."

github.com/huggingface/...
GitHub - huggingface/gpt-oss-recipes: Collection of scripts and notebooks for OpenAI's latest GPT OSS models
Collection of scripts and notebooks for OpenAI's latest GPT OSS models - huggingface/gpt-oss-recipes
github.com
August 10, 2025 at 4:46 AM
ota.bsky.social
"The Agar Mini is available in two distinct versions ...

Wired Edition: Powered by QMK firmware, ... is fully compatible with VIA, VIAL, ...

Dual-Mode Wireless Edition: Built on ZMK firmware, ... is fully customizable via the zmk.studio editor."

kbdfans.com/products/aga...
Agar Mini
Agar mini The Agar Mini represents the latest evolution in the esteemed Agar series, distilling its signature design language into a new compact form factor. This model preserves the elegant, curved a...
kbdfans.com
July 25, 2025 at 2:09 PM
Reposted by S. Ota
simonwillison.net
My notes on Gemini CLI, including poking around in their system prompt which I've extracted into a more readable rendered Gist simonwillison.net/2025/Jun/25/...
Gemini CLI
First there was Claude Code in February, then OpenAI Codex (CLI) in April, and now Gemini CLI in June. All three of the largest AI labs now have their own …
simonwillison.net
June 25, 2025 at 5:55 PM
Reposted by S. Ota
sciortino.bsky.social
The "secret project" I've been working on at my job has gone public (and open source) today! Check it out!
Gemini CLI: your open-source AI agent
Free and open source, Gemini CLI brings Gemini directly into developers’ terminals — with unmatched access for individuals.
blog.google
June 25, 2025 at 3:36 PM
ota.bsky.social
Since I had already used the Gemini API, I had to unset the GEMINI_API_KEY in order to authenticate with my Google account.

GEMINI_API_KEY="" gemini
June 25, 2025 at 3:15 PM
ota.bsky.social
I expanded the Claude code feed to include the Gemini CLI.

bsky.app/profile/did:...
Skyfeed settings for the Claude Code | Gemini CLI feed.
June 25, 2025 at 2:25 PM
ota.bsky.social
bsky.app/profile/did:...

This feed will be moved to

bsky.app/profile/did:...
June 23, 2025 at 3:16 PM
ota.bsky.social
Posts related to custom keyboard, DIY keyboard, mechanical keyboard, key switch, keycap, etc.
自作キーボード, キースイッチ, キーキャップなどを含むポスト.

bsky.app/profile/did:...
June 23, 2025 at 10:49 AM
ota.bsky.social
Posts related to `Claude Code`.
Claude Code を含むポスト。

bsky.app/profile/did:...
June 23, 2025 at 10:46 AM
Reposted by S. Ota
matsuu.bsky.social
Jupyter Notebookと生成AIの組み合わせ、あると思います。チャット用のサイドウィンドウを表示したり、%%aiでNotebookの中から生成AIに問い合わせできたり。各種AI利用の他、Ollamaなどにも対応してるのでローカルLLMも利用可。
GitHub - jupyterlab/jupyter-ai: A generative AI extension for JupyterLab
A generative AI extension for JupyterLab. Contribute to jupyterlab/jupyter-ai development by creating an account on GitHub.
github.com
June 1, 2025 at 6:59 AM
Reposted by S. Ota
paper.bsky.social
[5/30] 396 Likes, 96 Comments, 4 Posts
2505.03335, cs․LG | cs․AI | cs․CL, 06 May 2025

🆕Absolute Zero: Reinforced Self-play Reasoning with Zero Data

Andrew Zhao, Yiran Wu, Yang Yue, Tong Wu, Quentin Xu, Yang Yue, Matthieu Lin, Shenzhi Wang, Qingyun Wu, Zilong Zheng, Gao Huang
Reinforcement learning with verifiable rewards (RLVR) has shown promise in enhancing the reasoning capabilities of large language models by learning directly from outcome-based rewards. Recent RLVR works that operate under the zero setting avoid supervision in labeling the reasoning process, but still depend on manually curated collections of questions and answers for training. The scarcity of high-quality, human-produced examples raises concerns about the long-term scalability of relying on human supervision, a challenge already evident in the domain of language model pretraining. Furthermore, in a hypothetical future where AI surpasses human intelligence, tasks provided by humans may offer limited learning potential for a superintelligent system. To address these concerns, we propose a new RLVR paradigm called Absolute Zero, in which a single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them, without relying on any external data. Under this paradigm, we introduce the Absolute Zero Reasoner (AZR), a system that self-evolves its training curriculum and reasoning ability by using a code executor to both validate proposed code reasoning tasks and verify answers, serving as an unified source of verifiable reward to guide open-ended yet grounded learning. Despite being trained entirely without external data, AZR achieves overall SOTA performance on coding and mathematical reasoning tasks, outperforming existing zero-setting models that rely on tens of thousands of in-domain human-curated examples. Furthermore, we demonstrate that AZR can be effectively applied across different model scales and is compatible with various model classes.
May 8, 2025 at 12:07 AM
ota.bsky.social
Karabiner-Elements でマルチレイヤのキーマップを作った。karabiner.ts というライブラリを使ったらレイヤーが簡単に実装できた！

layer('japanese_eisuu', '英数 + ijkl').manipulators([
map('i').to('↑'),
map('j').to('←'),
map('k').to('↓'),
map('l').to('→'),
]),

それと久しぶりに Deno を使ってみたが、こういう簡単なプログラムならかなり楽。

github.com/susumuota/ka...
Screenshot of layers and keymap.
April 12, 2025 at 2:36 PM
ota.bsky.social
"By serving as an intermediary in user interactions, it can autonomously generate context-aware responses, prefill required information, and facilitate seamless communication with external systems, significantly reducing cognitive load and interaction friction."
paper.bsky.social Paper @paper.bsky.social · Apr 9
[28/30] 54 Likes, 12 Comments, 3 Posts
2503.08102, cs․AI | cs․CL | cs․HC, 12 Mar 2025

🆕AI-native Memory 2.0: Second Me

Jiale Wei, Xiang Ying, Tao Gao, Fangyi Bao, Felix Tao, Jingbo Shang
Human interaction with the external world fundamentally involves the exchange of personal memory, whether with other individuals, websites, applications, or, in the future, AI agents. A significant portion of this interaction is redundant, requiring users to repeatedly provide the same information across different contexts. Existing solutions, such as browser-stored credentials, autofill mechanisms, and unified authentication systems, have aimed to mitigate this redundancy by serving as intermediaries that store and retrieve commonly used user data. The advent of large language models (LLMs) presents an opportunity to redefine memory management through an AI-native paradigm: SECOND ME. SECOND ME acts as an intelligent, persistent memory offload system that retains, organizes, and dynamically utilizes user-specific knowledge. By serving as an intermediary in user interactions, it can autonomously generate context-aware responses, prefill required information, and facilitate seamless communication with external systems, significantly reducing cognitive load and interaction friction. Unlike traditional memory storage solutions, SECOND ME extends beyond static data retention by leveraging LLM-based memory parameterization. This enables structured organization, contextual reasoning, and adaptive knowledge retrieval, facilitating a more systematic and intelligent approach to memory management. As AI-driven personal agents like SECOND ME become increasingly integrated into digital ecosystems, SECOND ME further represents a critical step toward augmenting human-world interaction with persistent, contextually aware, and self-optimizing memory systems. We have open-sourced the fully localizable deployment system at GitHub: https://github.com/Mindverse/Second-Me.
April 11, 2025 at 2:57 AM
ota.bsky.social
"Visual Studio Codeのエージェントモードを全ユーザーに提供します。このモードはMCPをサポートしており、必要なあらゆるコンテキストや機能へのアクセスを可能にします。... エージェントモードのモデルは、Claude 3.5と3.7 Sonnet、Google Gemini 2.0 Flash、OpenAI GPT-4oから選択できます。現在、エージェントモードはClaude 3.7 Sonnetを使用した場合、SWE-bench Verifiedで56.0%の合格率を達成しています。"

github.blog/jp/2025-04-0...
GitHub Copilotでバイブコーディング：エージェントモードとMCPサポートがVS Codeユーザーに提供開始
MCPをサポートしたエージェントモードをすべてのVS Codeユーザーに展開します。また、新しい GitHub Copilot Pro+ プラン (プレミアム リクエスト付き)、Anthropic、Google、OpenAI のモデルの一般提供開始、Next Editコード補完提案、GitHub Copilot コード レビュー エージェントについても発表します。
github.blog
April 7, 2025 at 8:51 AM
ota.bsky.social
"... if a task can only be done by a handful of those most educated, that task is considered intellectual. One example is writing, the physical act of copying words onto paper. In the past, when only a small portion of the population was literate, writing was considered intellectual."
The End of Programming as We Know It
www.oreilly.com
April 7, 2025 at 8:05 AM
Reposted by S. Ota
dorialexander.bsky.social
Well the latest DeepSeek is very satisfying from an humanities perspective. The trick to generalize RL is replacing scalar grades with… source criticism (qualitative principles and critiques). arxiv.org/pdf/2504.02495
April 4, 2025 at 8:01 AM
ota.bsky.social
"A key challenge of RL is to obtain accurate reward signals for LLMs in various domains beyond verifiable questions or artificial rules. ... we investigate how to improve reward modeling (RM) with more inference compute for general queries, i.e. the inference-time scalability of generalist RM"
paper.bsky.social Paper @paper.bsky.social · Apr 5
[14/30] 130 Likes, 21 Comments, 2 Posts
2504.02495, cs․CL | cs․AI | cs․LG, 03 Apr 2025

🆕Inference-Time Scaling for Generalist Reward Modeling

Zijun Liu, Peiyi Wang, Runxin Xu, Shirong Ma, Chong Ruan, Peng Li, Yang Liu, Yu Wu
Reinforcement learning (RL) has been widely adopted in post-training for large language models (LLMs) at scale. Recently, the incentivization of reasoning capabilities in LLMs from RL indicates that $\textit{proper learning methods could enable effective inference-time scalability}$. A key chal...
April 5, 2025 at 4:05 AM
ota.bsky.social
I just realised that image alt text can be 2000 characters long.

I have fixed @paper.bsky.social to allow 2000 chars.
image alt text can be 2000 characters long.
April 5, 2025 at 3:58 AM