Skip to content
EventPast event

humans& releases Ampersand v0.1 model card

1 reports1 reporting sourcesLatest report

Start here

AI overview

humans& released the Ampersand v0.1 model card. This collaborative model is trained based on GLM 5.3 and provided through the Agents API. The safety evaluation shows that WMDP cybersecurity question accuracy rose from 80.6% to 89.0%, an increase of 8.4 percentage points, while biology and chemistry questions were essentially flat; the prompt injection success rate is similar to GLM 5.3 but clearly higher than frontier models, the harmful request attack risk score fell from 21.4 to 13.8, and the sycophantic dialogue proportion fell from 77.3% to 62.6%.

Generated by AI from reports · Updated 23 days ago

Latest update
humans& released the Ampersand v0.1 model card: WMDP cybersecurity question accuracy up 8.4 points, prompt injection success rate higher than frontier models

Related research

Cited in the reports

Progress timeline

Oct 21
  1. X@Jabreu89· Discussion
    Sources / Fuentes: Ampersand launch blog: https://ampersand.humansand.ai/blog/ Model card: https://ampersand.humansand.…(opens the original post in a new tab)
  2. X@Jabreu89· Discussion
    @humansand just released Ampersand, a model that joins a group chat and decides after every message whether to reply or…(opens the original post in a new tab)
  3. X@Jabreu89· Discussion
    @humansand acaba de lanzar Ampersand, un modelo que entra a un group chat y decide después de cada mensaje si responde…(opens the original post in a new tab)
  4. X@ajiteshrawatplx· Discussion
    humans& shipped its first model, Ampersand v0.1: • Trained to help groups hit shared goals over months • Built on…(opens the original post in a new tab)
  5. X@harrison_0x821· Discussion
    The hardest thing for an AI in a group chat is not being able to talk, but knowing when to shut up. At 6:37 this morning, humans& released its first model, Ampersand, trained specifically for this. First, the facts: 1. For every message in the group, it has only two actions, send_message or pass, no…(opens the original post in a new tab)
  6. X@iCleanAI· Discussion
    humans& released its first model, Ampersand v0.1, trained to help groups of people reach shared goals, not work alo…(opens the original post in a new tab)
  7. X@fiv_ai_news· Discussion
    AI 기업 humans&가 10월 20일 첫 모델 Ampersand v0.1을 X에서 공개했습니다. 작업을 대신 끝내는 비서가 아니라, 여러 사람이 함께 쓰는 대화방에서 언제 말하고 언제 입을 다물지 스스로 정하도…(opens the original post in a new tab)
  8. X@moftahlmabry· Discussion
    Most AI models are trained to finish your task. A new one is trained to know when to stay quiet. On Oct 20, humans& r…(opens the original post in a new tab)
  9. X@dsoch8c· Discussion
    The new lab humans& released its first model, Ampersand v0.1, which it officially says achieves 59.1% in the high-scoring tier for team collaboration. According to information seen so far, the model continues training on the foundation of Zhipu's 744B-parameter GLM 5.3, with the approach of having a large number of simulated users spar with it, through rebuttals and sulking…(opens the original post in a new tab)
  10. X@Secretof_forest· Discussion
    humans& released Ampersand v0.1, a model post-trained from GLM 5.3 (744B) to collaborate in groups instead of repla…(opens the original post in a new tab)
  11. X@IDEric_· Discussion
    Most AI is trained to finish tasks. humans& trained Ampersand v0.1 to help groups of people work together, and it c…(opens the original post in a new tab)
  12. X@kdw9804· Discussion
    humans&가 그룹 협업용 모델 Ampersand v0.1을 API로 열었어요. 답을 잘 쓰는 것보다, 여러 사람이 있는 대화에서 끼어들지 말지부터 판단하도록 학습된 모델입니다. · 모델 ID: ampersan…(opens the original post in a new tab)
  13. X@4kd0CRBVsLxm9gg· Discussion
    An AI that knows when to shut up in a group chat is what it means to be sensible. humans& releases its first "collaborator" model, Ampersand v0.1: trained in months-long group simulations, its goal is to help people get things done together, rather than rushing to answer the question; Collaborative Intelligence total score 59.…(opens the original post in a new tab)
  14. X@AI_N0W· Discussion
    Sometimes the useful AI move is silence. humans& released Ampersand v0.1 for group collaboration; it can wait inste…(opens the original post in a new tab)
  15. X@zwz7107· Discussion
    @alacheng One team doing post-training based on GLM 5.3 does not equal “more teams abandoning pretraining.” The more accurate change is: once general-purpose base models become commoditized, differentiation shifts to environments, user models and reward design. Ampersand's key validation is also not simulated leaderboards, but whether cross-month goal improvements can be… from Pers…(opens the original post in a new tab)
  16. X@aloo51523593330· Discussion
    humans& shipped its first model, Ampersand v0.1: a "collaborator" RL-trained in months-long group sims, built on GL…(opens the original post in a new tab)
  17. X@happyha83164146· Discussion
    humans& has just released Ampersand v0.1, focused on an AI that can collaborate with teams for months; Claude Code, by default, compresses a session into a summary after about 1 hour of idleness. AI memory is getting longer, but who owns it? @MemoLabsOrg's answer: models can be replaced…(opens the original post in a new tab)
  18. X@econographs· Discussion
    🛠 Tools worth the click humans&, the AI lab founded by alumni of OpenAI, Anthropic, Google DeepMind and xAI, has rele…(opens the original post in a new tab)
  19. X@alacheng· Discussion
    humans& previously raised 480 million USD and has just launched a large model with better “collaboration capabilities”; the model itself is post-trained based on GLM 5.3 (it seems more teams are abandoning pretraining and using post-training on open-source models to improve some capability of the model). Model use cases (see figure). Model features: Amper…(opens the original post in a new tab)
  20. X@soloist_kn· Discussion
    AI also has the option to “stay silent for now.” humans& has released Ampersand v0.1. Based on the group's speakers and history, the design chooses whether to reply or hold back. It illustrates scheduling examples and limitations in 52 seconds. It is a paid API for developers and requires integration. Collab…(opens the original post in a new tab)
  21. X@humansand· Discussion
    We're releasing Ampersand v0.1 through our agent API, for developers building apps where people collaborate with each o…(opens the original post in a new tab)
  22. X@humansand· Discussion
    No one changes the world alone Introducing Ampersand - the first model built for us. Ampersand is a collaborator traine…(opens the original post in a new tab)
Oct 20
  1. humans&
    humans& released the Ampersand v0.1 model card: WMDP cybersecurity question accuracy up 8.4 points, prompt injection success rate higher than frontier models

    humans& released the Ampersand v0.1 model card; the collaborative model is trained based on GLM 5.3 and provided through the Agents API. Safety evaluation shows that WMDP cybersecurity question accuracy rose from 80.6% to 89.0%, an increase of 8.4 percentage points, while biology and chemistry questions were essentially flat; prompt injection success rate was similar to GLM 5.3 but clearly higher than frontier models, the harmful request attack risk score dropped from 21.4 to 13.8, and the sycophantic dialogue proportion dropped from 77.3% to 62.6%.

Related discussion

Total 22 items