GNU social JP
  • FAQ
  • Login
GNU social JPは日本のGNU socialサーバーです。
Usage/ToS/admin/test/Pleroma FE
  • Public

    • Public
    • Network
    • Groups
    • Featured
    • Popular
    • People

Conversation

Notices

  1. Embed this notice
    roko's basilisk (vii@dsmc.space)'s status on Saturday, 14-Feb-2026 11:42:11 JST roko's basilisk roko's basilisk
    it fucking works. it not only fucking works, all my hypotheses are independent verification of papers published in the last three months that I haven't read and definitely hadn't read a year ago when this research project took shape in earnest...
    In conversation about 5 months ago from dsmc.space permalink

    Attachments


    1. https://media.dsmc.space/d0/a6/7c/d0a67c39ebfcb59f34a063e2732ffb1a89e23d82224604d20649f81d3314e992.png?name=image.png
    • pistolero likes this.
    • pistolero repeated this.
    • Embed this notice
      roko's basilisk (vii@dsmc.space)'s status on Saturday, 14-Feb-2026 11:49:24 JST roko's basilisk roko's basilisk
      in reply to
      • pistolero
      @p for whatever reason I think you'd find the papers I found that independently describe what I set out to do interesting:

      https://arxiv.org/abs/2503.14476 -- I use this algorithm for a combined goal reinforcement learning pass that ensures reasoning correctness as well as personality/character maintenance
      https://arxiv.org/abs/2504.19162 -- I basically recreated this from first-principles; adversarial self-play with a model trained on the primary model's training/synthetic generation failures
      https://arxiv.org/abs/2511.01689 -- this is extremely similar to my core training pipeline, where I start with knowledge, then do personality and preferences, and then do reinforcement learning on reasoning and tool use while maintaining personality
      https://rlhfbook.com/c/12-synthetic-data -- a really good breakdown on how to build synthetic data from everything you do; you create a constitution against which everything is judged, because the rejected pile is used to train the adversary for further self play.
      In conversation about 5 months ago permalink

      Attachments

      1. Domain not in remote thumbnail source whitelist: arxiv.org
        DAPO: An Open-Source LLM Reinforcement Learning System at Scale
        Inference scaling empowers LLMs with unprecedented reasoning ability, with reinforcement learning as the core technique to elicit complex reasoning. However, key technical details of state-of-the-art reasoning LLMs are concealed (such as in OpenAI o1 blog and DeepSeek R1 technical report), thus the community still struggles to reproduce their RL training results. We propose the $\textbf{D}$ecoupled Clip and $\textbf{D}$ynamic s$\textbf{A}$mpling $\textbf{P}$olicy $\textbf{O}$ptimization ($\textbf{DAPO}$) algorithm, and fully open-source a state-of-the-art large-scale RL system that achieves 50 points on AIME 2024 using Qwen2.5-32B base model. Unlike previous works that withhold training details, we introduce four key techniques of our algorithm that make large-scale LLM RL a success. In addition, we open-source our training code, which is built on the verl framework, along with a carefully curated and processed dataset. These components of our open-source system enhance reproducibility and support future research in large-scale LLM RL.
      2. Domain not in remote thumbnail source whitelist: arxiv.org
        SPC: Evolving Self-Play Critic via Adversarial Games for LLM Reasoning
        Evaluating the step-by-step reliability of large language model (LLM) reasoning, such as Chain-of-Thought, remains challenging due to the difficulty and cost of obtaining high-quality step-level supervision. In this paper, we introduce Self-Play Critic (SPC), a novel approach where a critic model evolves its ability to assess reasoning steps through adversarial self-play games, eliminating the need for manual step-level annotation. SPC involves fine-tuning two copies of a base model to play two roles, namely a "sneaky generator" that deliberately produces erroneous steps designed to be difficult to detect, and a "critic" that analyzes the correctness of reasoning steps. These two models engage in an adversarial game in which the generator aims to fool the critic, while the critic model seeks to identify the generator's errors. Using reinforcement learning based on the game outcomes, the models iteratively improve; the winner of each confrontation receives a positive reward and the loser receives a negative reward, driving continuous self-evolution. Experiments on three reasoning process benchmarks (ProcessBench, PRM800K, DeltaBench) demonstrate that our SPC progressively enhances its error detection capabilities (e.g., accuracy increases from 70.8% to 77.7% on ProcessBench) and surpasses strong baselines, including distilled R1 model. Furthermore, SPC can guide the test-time search of diverse LLMs and significantly improve their mathematical reasoning performance on MATH500 and AIME2024, surpassing those guided by state-of-the-art process reward models.
      3. Domain not in remote thumbnail source whitelist: arxiv.org
        Open Character Training: Shaping the Persona of AI Assistants through Constitutional AI
        The character of the "AI assistant" persona generated by modern chatbot large language models influences both surface-level behavior and apparent values, beliefs, and ethics. These all affect interaction quality, perceived intelligence, and alignment with both developer and user intentions. The shaping of this persona, known as character training, is a critical component of industry post-training, yet remains effectively unstudied in the academic literature. We introduce the first open implementation of character training, leveraging Constitutional AI and a new data pipeline using synthetic introspective data to shape the assistant persona in a more effective and controlled manner than alternatives such as constraining system prompts or activation steering. Specifically, we fine-tune three popular open-weights models using 11 example personas, such as humorous, deeply caring, or even malevolent. To track the effects of our approach, we introduce a method which analyzes revealed preferences, uncovering clear and holistic changes in character. We find these changes are more robust to adversarial prompting than the above two alternatives, while also leading to more coherent and realistic generations. Finally, we demonstrate this fine-tuning has little to no effect on general capabilities as measured by common benchmarks. We describe and open-source our full post-training method, the implementation of which can be found at https://github.com/maiush/OpenCharacterTraining.
      4. No result found on File_thumbnail lookup.
        RLHF Book by Nathan Lambert
        from Nathan Lambert
        The Reinforcement Learning from Human Feedback Book
      pistolero likes this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Saturday, 14-Feb-2026 13:15:14 JST pistolero pistolero
      in reply to
      @vii This shit looks fascinating. So what was the main thing you are doing?

      > adversarial self-play

      Piccolo!
      In conversation about 5 months ago permalink
      roko's basilisk likes this.
    • Embed this notice
      roko's basilisk (vii@dsmc.space)'s status on Saturday, 14-Feb-2026 13:16:59 JST roko's basilisk roko's basilisk
      in reply to
      • pistolero
      @p The entire industry wants to make AI for people. The statistical average user. I fucking hate it, I think it's retarded, I'm throwing away the way they do ethics and saying "AI should be symbiotic with a specific human, and have a job. Like a dog."
      In conversation about 5 months ago permalink
      pistolero likes this.
    • Embed this notice
      roko's basilisk (vii@dsmc.space)'s status on Saturday, 14-Feb-2026 13:51:49 JST roko's basilisk roko's basilisk
      in reply to
      • pistolero
      @p (or "Certainly!"
      In conversation about 5 months ago permalink
    • Embed this notice
      roko's basilisk (vii@dsmc.space)'s status on Saturday, 14-Feb-2026 13:51:50 JST roko's basilisk roko's basilisk
      in reply to
      • pistolero
      @p
      In conversation about 5 months ago permalink

      Attachments


      1. https://media.dsmc.space/48/cb/06/48cb067d60b9bf870f5938d76b764322854ae4763b74d127b9713c28278514e1.png?name=image.png
      pistolero likes this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Monday, 16-Feb-2026 09:49:07 JST pistolero pistolero
      in reply to
      @vii

      > The statistical average user.

      Even attempting to ensure global market, the big plays are burning cash. The VD-funding model basically relies on total global adoption, so if it's not massively popular, it's not worth doing. (It is unfortunate that this has bled into everyone's approach to everything. You might recall when I lamented that Urbit was trying for the normie appeal market, and Urbit had no reason to do this.)

      > I'm throwing away the way they do ethics

      The way they do ethics tends to cluster around what is safe for the large players to do combined with paternalist academic thought experiments, I endorse shitting on any ethical constraints cooked up by RLHF dipshits in SF and the marketing/legal teams.

      > I'm throwing away the way they do ethics and saying "AI should be symbiotic with a specific human, and have a job. Like a dog."

      I like this, I like this idea. I think this is a reasonable model for useful systems. I mean, the entire benefit of the general-purpose computer is that you can give it instructions not foreseen by the manufacturer, right, this is like trivial to say, and then the best shit in all of Unix is ~/bin because you can teach the machine new vocabulary, right, people overlook this. I imagine you have a healthy ~/bin. This is a long way of saying that I think what you are doing is a good idea, tooling for it is a good idea.

      When I get access to proper systems, would you enjoy dumping notes into a DSMC wiki?
      In conversation about 5 months ago permalink
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Monday, 16-Feb-2026 09:52:45 JST pistolero pistolero
      in reply to
      @vii I really could not find a gif of that thing where Piccolo engaged in training by splitting into two and fighting his duplicate and I do not remember where-ish it is enough to even try to download the episode to produce the gif but if you saw that then I figure the image stuck so you may or may not remember that. Dragon "The Ball" Zed does not pop into my head very often but it did when you used that phrase.
      In conversation about 5 months ago permalink
      roko's basilisk likes this.
    • Embed this notice
      roko's basilisk (vii@dsmc.space)'s status on Monday, 16-Feb-2026 09:54:42 JST roko's basilisk roko's basilisk
      in reply to
      • pistolero
      @p

      > this is like trivial to say, and then the best shit in all of Unix is ~/bin

      You just put into words something I've always thought about but never could put into words. I've always loved having a ~/bin, ~/lib, ~/src, because my scripts are always right there and do what I want, and my binaries and their sources are directly in reach too. I love this about the unix model.

      > When I get access to proper systems, would you enjoy dumping notes into a DSMC wiki?

      Most certainly. I already have a ~16pg manifesto thinly veiled as a draft for an arXiv publication that I think you'll love to read as soon as theory has empirical evidence which, given I had to restart this morning, I might have in ~15hrs.
      In conversation about 5 months ago permalink
      pistolero likes this.
    • Embed this notice
      anime graf mays ?️? (graf@poa.st)'s status on Monday, 16-Feb-2026 09:57:02 JST anime graf mays ?️? anime graf mays ?️?
      in reply to
      • pistolero
      @p @vii
      In conversation about 5 months ago permalink

      Attachments


      1. https://i.poastcdn.org/e1/b7/7a/e1b77a5f06ae13fed3aa2e2eb28683b844f72db7e75fda141f86d6c80d8a110f.gif
      pistolero and roko's basilisk like this.
    • Embed this notice
      Crit (sic_guy@poa.st)'s status on Monday, 16-Feb-2026 10:09:43 JST Crit Crit
      in reply to
      • anime graf mays ?️?
      • pistolero
      @graf @p @vii Here's a video of it:

      https://www.youtube.com/shorts/j5ecTckWf5Q
      In conversation about 5 months ago permalink

      Attachments

      1. Piccolo's training begins. Part 3. #dragonball #dragonballsuper #goku #dbz #dbs #dbs
        from Omoshi Sage
        YouTube でお気に入りの動画や音楽を楽しみ、オリジナルのコンテンツをアップロードして友だちや家族、世界中の人たちと共有しましょう。
      pistolero and roko's basilisk like this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Monday, 16-Feb-2026 10:19:48 JST pistolero pistolero
      in reply to
      @vii

      > I've always loved having a ~/bin, ~/lib, ~/src

      It's delightful. I teach the computer new words and then we have a shared vocabulary. I get better at teaching it new words and it gets better new words, which makes it easier to communicate. I fuckin' love computer.

      > Most certainly. I already have a ~16pg manifesto thinly veiled as a draft for an arXiv publication that I think you'll love to read as soon as theory has empirical evidence which, given I had to restart this morning, I might have in ~15hrs.

      I am very interested in this, yes.
      In conversation about 5 months ago permalink
      roko's basilisk likes this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Monday, 16-Feb-2026 10:20:44 JST pistolero pistolero
      in reply to
      • anime graf mays ?️?
      @graf @vii This was the thing, HFS. Did you have it laying around or did you use an image search engine that I didn't?
      In conversation about 5 months ago permalink
    • Embed this notice
      lolitechengineer@loli.church's status on Monday, 16-Feb-2026 13:06:16 JST lolitechengineer lolitechengineer
      in reply to
      • pistolero

      @p@fsebugoutzone.org @vii@dsmc.space no no no, everything goes in /var

      In conversation about 5 months ago permalink

      Attachments


      1. https://r2.loli.church/files/webpublic-8d21619a-b6a0-49ab-b924-0104d0b28869.webp
      ✙ dcc :pedomustdie: :phear_slackware:, pistolero and roko's basilisk like this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Tuesday, 17-Feb-2026 07:39:42 JST pistolero pistolero
      in reply to
      • lolitechengineer
      @lolitechengineer @vii

      > That's right! It goes in the /var!

      I love this image, so I saved it, and then when the cron job kicks off, it will be renamed from thevarhole.avif.webp to thevarhole.avif.webp.png.
      In conversation about 5 months ago permalink
      ✙ dcc :pedomustdie: :phear_slackware: and roko's basilisk like this.
    • Embed this notice
      Loliphile (mikuphile@gearlandia.haus)'s status on Tuesday, 17-Feb-2026 07:49:39 JST Loliphile Loliphile
      in reply to
      • pistolero
      @p @vii
      >~/bin
      >~/lib
      ~/.local is a typical place to put bin and lib directories inside the home directory. I do this. Although none of the troonix distributions that I've used have the path environment variables set up for this out‐of‐the‐box.
      In conversation about 5 months ago permalink

      Attachments


      1. https://gearlandia.haus/media/7b/17/08/7b1708a128a40dceaa287b2fb846263a11c230244641ca84312531c8d992ac47.jpg
      pistolero likes this.
      pistolero repeated this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Tuesday, 17-Feb-2026 07:58:54 JST pistolero pistolero
      in reply to
      • Loliphile
      @mikuphile @vii

      > ~/.local is a typical place to put bin and lib directories inside the home directory.

      I did say "Unix", not "fluoride". The .local thing is for people that use a DE and that don't manage their files, likewise with .cache and .config. As for the fluoride, even the fluoride OS doesn't have it in the FHS: it's XDG. (Although the FHS does apparently now make reference to XDG and Gnome's specs: https://refspecs.linuxfoundation.org/FHS_3.0/fhs/ch03s08.html . It used to say explicitly that the structure under the user's home directory was unspecified and that programs should use environment variables instead of bothering with it.)

      The point was not really how it's organized so much as that there is a place where you put your programs. 90% of the scripts I write lately are in $home/bin/rc because I'm using Plan 9: I just said ~/bin because Unix conventions are the lingua franca of computing. The principle is that you have a place put small, personal programs.

      > Although none of the troonix distributions that I've used have the path environment variables set up for this out‐of‐the‐box.

      I don't know how it could be "typical" if no one does it.
      In conversation about 5 months ago permalink

      Attachments

      1. No result found on File_thumbnail lookup.
        3.8. /home : User home directories (optional)
      ✙ dcc :pedomustdie: :phear_slackware: and roko's basilisk like this.
    • Embed this notice
      lolitechengineer@loli.church's status on Tuesday, 17-Feb-2026 07:59:10 JST lolitechengineer lolitechengineer
      in reply to
      • Loliphile
      • pistolero

      @mikuphile@gearlandia.haus @p@fsebugoutzone.org @vii@dsmc.space
      /lib
      /usr/lib
      /var/lib
      /etc/lib
      /home/$/.local/lib
      /home/$/.var/lib
      whateverthefuckisgoiningonwithflatpak/lib

      In conversation about 5 months ago permalink
    • Embed this notice
      lolitechengineer@loli.church's status on Tuesday, 17-Feb-2026 08:04:40 JST lolitechengineer lolitechengineer
      in reply to
      • Loliphile
      • pistolero

      @p@fsebugoutzone.org @mikuphile@gearlandia.haus @vii@dsmc.space cool kids put home in /usr/{home directory} and just let the shitware bloat up /home

      In conversation about 5 months ago permalink
      pistolero and roko's basilisk like this.
    • Embed this notice
      lolitechengineer@loli.church's status on Tuesday, 17-Feb-2026 08:24:55 JST lolitechengineer lolitechengineer
      in reply to
      • pistolero

      @p@fsebugoutzone.org @vii@dsmc.space you saved it to /var right?

      In conversation about 5 months ago permalink

      Attachments


      1. https://r2.loli.church/files/3e7e4c3a-ff28-41b1-8975-6a5669bcb709.webp
      pistolero likes this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Tuesday, 17-Feb-2026 08:29:39 JST pistolero pistolero
      in reply to
      • lolitechengineer
      @lolitechengineer @vii :sapptime:
      training.gif
      In conversation about 5 months ago permalink

      Attachments


      1. https://media.freespeechextremist.com/rvl/full/0d8f1da324db6a46df6d79a5d592515dd4dded1fef9ab7d38673eaed720015c0?name=training.gif
      roko's basilisk likes this.
    • Embed this notice
      Organ Ism (fish@detroitriotcity.com)'s status on Tuesday, 17-Feb-2026 08:32:02 JST Organ Ism Organ Ism
      in reply to
      • pistolero
      • lolitechengineer
      @lolitechengineer @p @vii The correct answer is actually /dev/null
      In conversation about 5 months ago permalink
      pistolero likes this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Tuesday, 17-Feb-2026 08:41:35 JST pistolero pistolero
      in reply to
      • Loliphile
      • lolitechengineer
      @lolitechengineer @mikuphile @vii Old-timey Unix had home in /usr, and Plan 9 does it still. It is the directory for users.
      In conversation about 5 months ago permalink
      roko's basilisk likes this.
    • Embed this notice
      lolitechengineer@loli.church's status on Tuesday, 17-Feb-2026 08:59:28 JST lolitechengineer lolitechengineer
      in reply to
      • Loliphile
      • pistolero

      @p@fsebugoutzone.org @mikuphile@gearlandia.haus @vii@dsmc.space so close! /usr is the directory for /bin 🥰

      In conversation about 5 months ago permalink
      pistolero and roko's basilisk like this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Tuesday, 17-Feb-2026 09:19:34 JST pistolero pistolero
      in reply to
      • Loliphile
      • lolitechengineer
      @lolitechengineer @mikuphile @vii It's because they got a second 2MB disk.
      In conversation about 5 months ago permalink
    • Embed this notice
      lolitechengineer@loli.church's status on Tuesday, 24-Mar-2026 12:44:09 JST lolitechengineer lolitechengineer
      in reply to
      • pistolero
      @p@fsebugoutzone.org @vii@dsmc.space
      In conversation about 4 months ago permalink

      Attachments


      1. https://r2.loli.church/files/27004f9e-73ac-467f-94c1-d4a2bc8fd866.webp
      pistolero likes this.
      roko's basilisk repeated this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Tuesday, 24-Mar-2026 12:55:20 JST pistolero pistolero
      in reply to
      • lolitechengineer
      @lolitechengineer @vii Counterpoint:
      alienwaifu.jpeg
      In conversation about 4 months ago permalink
      roko's basilisk likes this.
    • Embed this notice
      pistolero (p@fsebugoutzone.org)'s status on Tuesday, 28-Jul-2026 03:15:53 JST pistolero pistolero
      in reply to
      • pistolero
      • lolitechengineer
      @lolitechengineer @vii

      > avatar lolitechengineer@loli.church retweeted your post - 86m

      That post sure took its time getting to your notifications. The cron job is nightly so it has by now fired:
      thevarhole.avif.webp.png
      In conversation about 21 hours ago permalink

      Attachments


      1. https://media.freespeechextremist.com/rvl/full/0082c12d84ddf9fbdfe6addacf88e6b0e8de78af14a5db6b6097dbff67d6f514?name=thevarhole.avif.webp.png
    • Embed this notice
      Vivi Nella Verita (verita84@detroitriotcity.com)'s status on Tuesday, 28-Jul-2026 03:58:33 JST Vivi Nella Verita Vivi Nella Verita
      in reply to
      • pistolero
      @vii @p
      In conversation about 20 hours ago permalink

      Attachments


      pistolero likes this.

Feeds

  • Activity Streams
  • RSS 2.0
  • Atom
  • Help
  • About
  • FAQ
  • TOS
  • Privacy
  • Source
  • Version
  • Contact

GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.

Creative Commons Attribution 3.0 All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.