めんふぃむ retweeted
🚨 BREAKING: GEMINI SUBSCRIPTIONS ARE COMING TO HERMES. 🔥
Today, Gemini in Hermes runs on an API key, not the Gemini plan you already pay for.
That's being worked on, the approved way:
▸ Nous is working with Google to bring your Gemini subscription into Hermes
▸ It should be officially supported, with Google's sign-off
▸ Apps plugging in a Gemini plan today aren't doing it the way Google approves
▸ Not live yet, no date
Claude and ChatGPT subscribers can already bring their plan to Hermes. Gemini is next in line.
Microsoft just made Claude Code and Codex cheaper without touching the agent.
TokenEconomizer is a free, open-source plugin that learns from the sessions your agent already ran in a repo:
- Remembers the repo facts your agent keeps rediscovering
- Stops loading tools that never get used, since every tool definition is sent with every request
- Works with Claude Code, Codex, Copilot CLI and Pi
- 18.4% to 30.9% lower cost across 5 agent-model setups on SWE-Bench
If you hit your usage limits every week, this is free money.
Save your tokens.
github.com/microsoft/TokenEc…
めんふぃむ retweeted
まさに自分も自作フィードリーダー作ってました。みんな考えることは同じ
secon.dev/entry/2026/11/12/2…
めんふぃむ retweeted
KagglerでCSバックグランドがない方はこちらでネットワークの勉強しましょう
経路決定で色々面白いアルゴリズムが使われてたりしますよ
youtube.com/watch?v=t0tVQkIA…
📊 耶鲁计算机博士生刘晨把自己发 Nature 子刊和 NeurIPS 的论文配图脚本全开源了,figures4papers 让你的论文图一眼看上去就有顶会的样子。
figures4papers 在 GitHub 揽获 1.1 万 star,也上了 HelloGitHub 的推荐。
以前赶论文 deadline,图往往是最后一晚用默认样式凑出来的,配色、字号、图例都不讲究,审稿人扫一眼就觉得粗糙。
这个仓库里放的是作者真实发表过的图:柱状图、雷达图、折线图、3D 球体图、趋势图,每张都配好生成它的 Python 脚本,照着换上自己的数据就能出同款。
更省事的是,figures4papers 里还附了一套科研绘图 skill,能在 Cursor、Claude Code、Codex 里引用,让 AI 编程助手按这套风格直接帮你画。
论文的门面,往往就是那几张图。
GitHub:github.com/ChenLiu-1996/figu…
めんふぃむ retweeted
OpenAIがロボティクスチーム解散した話。機械学習×ロボティクス学会CoRLでブース取った後に解散した(言語モデルベースのロボ制御に移管された?)らしい。たしかに6月は積極採用してたっぽい。
teapawt.org/sama/status/2061117302… teapawt.org/fortunemagazine/status…
めんふぃむ retweeted
牛逼!谷歌官方刚发的东西,48小时直接干上GitHub热门第一。
AI写前端默认还停在2019年,因为训练数据全是老屎山。
装个技能包直接拉满:通过 56%→88%。
一行命令搞定:npx modern-web-guidance@latest install
212个现代Web特性直接喂给Claude、Codex、Copilot。
仓库地址:github.com/GoogleChrome/mode…
兄弟们,你们觉得下一个该被官方“喂饭”的屎山领域是啥?
めんふぃむ retweeted
Some of my takeaways from CoRL 2026:
1. Deployment is key. The most valuable data comes from real-world deployments. Smarter models need fewer interventions, but get more out of them.
2. World models don't need to be explicitly built into the architecture. With the right data, supervision, and context, models can reason through complex situations.
3. To keep advancing, evals are also a bottleneck. Robotics benchmarks need to improve. We should measure speed, reliability, and robustness, not just task success rates.
4. real2sim2real and code-as-policy are gonna play a key role in overcoming the data bottleneck, enabling deployment in narrow tasks, and collecting data to train future policies. Especially useful for industrial deployments, while home environments offer more opportunities for robot learning.
5. Conferences need to change. We already know most papers and demos from X and arXiv. More events that match people by research interests and create meaningful connections, less focus on traditional paper presentations. Adding tracks to CoRL, for example (?).
6. Tactile sensing continues to be hard. Sensors are heterogeneous, data is noisy and expensive to collect, and there's still no consensus on which sensors to use.
7. A lot of things are still open research. I personally feel we haven't figured out the right recipe for scaling yet. So many questions remain unanswered.
8. More and more people are moving into robotics, and the open-source community is growing. I hope we see more hardware and electronics projects coming out of it.
Summer is coming!
めんふぃむ retweeted
You asked for dsh, I added it
DeepSeek Harness is the open-source coding agent by @deepseek_ai. It's now on the @huggingface Harness Leaderboard, which compares ten coding agents on the same open models.
DSH is statistically tied for first in all five evaluations: DeepSeek-V4.1-Flash and GLM-5.3-Flash at low and high effort, and Flint.
It shines on DeepSWE, where agents change code in open-source repositories: 2nd of 10 averaged over the three models, 0.5 points behind Hermes Agent by @NousResearch. On Terminal-Bench's command-line tasks it ranks 6th.
One habit explains much of that gap: when DSH finishes a task, it shuts down every background program it started.
Three Terminal-Bench tasks ask the agent to start a server, for example a model inference API. The grader tests that server after the agent has finished. With DSH, the server is gone by then.
On DeepSeek-V4.1-Flash and GLM-5.3-Flash, DSH passed 2 of 36 attempts on these tasks. @opencode, @pidotdev and Hermes Agent passed all 36.
Those three tasks cost DSH 1.5 to 1.7 points in every evaluation. On DeepSeek-V4.1-Flash at high effort, that is more than its 1.3-point gap to the leader.
Developer preview: 0.2.2-rc.1. Web search is off because it needs a DeepSeek account key.
It costs $0.53 per solved trial, averaged over the three models.
Leaderboard: huggingface.co/spaces/joelni…
めんふぃむ retweeted
これはopus-5.5にサブエージェントありでAHC074を解かせたときにどういうサブエージェントを用いたかの履歴で、行き詰まってくるとfresh-eyesというサブエージェントをeffort=maxで起動している
めんふぃむ retweeted
Alpamayo 2 Edge x Autowareをリリースしました。VLA基盤モデルによる自動運転AIがすべてオープンソースで手に入ります。
github.com/autowarefoundatio…
huggingface.co/nvidia/Alpama…
E2E自動運転のAIモデルをオープンソース公開しました。地図ありなし、LiDARありなし、GPUありなし、色々選択できるようになりました。
prtimes.jp/main/html/rd/p/00…
めんふぃむ retweeted
We've given all Pro, Max, Ultra, and Team users one free full reset, across all models and limits. Use it anytime before December 10 at 2:00PM PT.
We recently fixed a few small issues that used more limits than intended, and the Fable resets didn't do much for those of you who'd already hit your weekly limits for all models. While these issues didn't affect everyone, we want to make up for the trouble.
You'll find it in Settings > Usage, or at clau.de/reset.
めんふぃむ retweeted
ベンチマーク用にsandboxを作って内部で実行しているので、メモリ・インターネット・作業ディレクトリ以外へのアクセスも出来ない隔離状態。プロンプト(初回)はこんな感じです。時間を指定するとclaudeが残り時間で出来ることがないと言うので、外部から打ち切っています
gist.github.com/wata-orz/75f…
めんふぃむ retweeted
Human in the loop for evaluation -> I'll pass.
The main advantage of Kaggle competitions is to have an automated evaluation procedure. The best score submission wins. Simple.
As soon as you introduce human ins the loop then you introduce bias.
Having some extra prize awarded by a human jury can make sense. But gating submissions with a human evaluation is plain wrong IMHO.
Big news! Today we're launching the competition for Physical-Engineering's Last Exam on @Kaggle, with $400,000 in prizes! 🛠️⚡
LLMs have conquered coding benchmarks—but can your agent engineer the physical world?
Today, the best agent scores under 50%. The headroom is massive. 🚀
Replying to @cradrrat
We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.
Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.
Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.
めんふぃむ retweeted
OpenCode 2 far outscores OpenCode 1.19: from 8th of 9 harnesses to tied for first on all three models.
It replaces 1.19 in the @huggingface Harness Leaderboard, which compares coding agents on the same models. Like most harnesses there, it now uses the Chat Completions API directly, so it's more comparable.
At high effort it solves 65.7% of tasks instead of 43.3% on GLM-5.3-Flash and 69.4% instead of 57.0% on DeepSeek-V4.1-Flash. On Flint, 22.9% instead of 7.1%.
Cost per solved task fell from $3.35 to $0.45.
Part of the old gap was the setup. Harbor's standard route for OpenCode 1.19 is OpenAI's Responses API, which I translated to Chat Completions so it works with Hugging Face Inference Providers.
It ran without errors but silently dropped the model's earlier reasoning. Thanks to @thdxr and @rekram11 from @opencode for spotting it!
Leaderboard: huggingface.co/spaces/joelni…
めんふぃむ retweeted
【OpenAI、AI学習への利用方針を伝える共通シグナルに対応へ】
OpenAIは、コンテンツを生成AIの学習に使ってよいかというWebサイト側の意思を、参加事業者へ一括で伝える共通シグナルを尊重すると発表した。参加は任意。
仕様はIETF(標準化団体)への個人提出の草案。サイト側はrobots.txt(巡回ルールを置くファイル)のUser-Agentに「__usage_ai-training」を指定し、Disallowで拒否、Allowで許容の意思をパス(URLの経路)ごとに示す。
OpenAIは今後数カ月での実装を目指し、従来の巡回ルールも引き続き尊重する。
めんふぃむ retweeted
Having the most mind-numbing experience with #ICLR2027 reviews this year...
Reviewer: Reviews the wrong paper
Me: "I think you reviewed the wrong paper"
Reviewer: Decreases score to a 1, increases confidence to a 4, and pastes in 100% AI generated review
Me: ...
めんふぃむ retweeted
PLaMo 3 Translateの重みが公開されたので、画面の文字をその場で日本語に置き換えるWindowsアプリ「Sonoba」を作ってみました!
今までの翻訳ツールは、テキストを選択したり、コピーしてツールに入れたりする必要があったけど、これは見たまま翻訳できる。
画像の中の文字もそのまま訳せるので、アーキテクチャ図みたいな技術的な画像にも使える。
インストール不要で、全部PCの中で動く。
図1枚で約2秒、字幕は約1秒、1文だけなら約0.2秒。軽い0.8Bなら、図1枚で約1秒、GPUメモリは約0.65GB。
(モデルは初回にアプリからダウンロード、実測はRTX 4070 Ti、1文は760×60の範囲)
リポジトリはリプライに置いておきます。
【ブログ】翻訳特化LLM PLaMo 3 Translateのモデル重みをリリースしました。0.8B / 2B / 8B / 31Bの4パターンで、いずれも世界最高レベルの翻訳能力です。ダウンロードして、ローカルで試せます。開発秘話を描いたTech Blog:
preferred.jp/ja/blog/tech/pl…
Hugging Face:
huggingface.co/collections/p…
























