Researcher at the Institute of Physics, University of Belgrade. Computational social science.
Novi Sad, Serbia
Joined March 2009
- Tweets2.5K
- Following2K
- Followers1.1K
- Likes40.5K
YSocial is the most feature-complete social media simulation or digital twin framework available right now.
High behavioral fidelity, interesting features for different simulations of X/Reddit communities. Validated against real social media data.
Very easy to set up: y-not.social/
What happens when a simulated user gets trashed in the replies?
In YSocial 4.0.0 he gets stressed, posts less, remembers who did it, and may even leave the platform.
Stress/reward, memory and opinion dynamics.
y-not.social/blog/2026/ysoci… #YSocial
Follow @omabib_app for new updates and use cases!
Omabib 0.2. Free reference manager for Linux and macOS. Your Claude Code, Codex or ChatGPT work in the same library.
Zotero import. PDF highlights. Sync on your storage. Built-in MCP.
omabib.github.io
Aleksandar Tomasevic retweeted
Omabib 0.2. Free reference manager for Linux and macOS. Your Claude Code, Codex or ChatGPT work in the same library.
Zotero import. PDF highlights. Sync on your storage. Built-in MCP.
omabib.github.io
I wanted to build a completely open source, fast, free Zotero alternative with first-class agent support both in-app and via MCP.
Give Omabib a try! omabib.github.io/
Bring your own cloud storage. Bring your own agent.
Early alpha version, macOS and Linux. Full @omarchy support, super-fast.
New in 0.2.0:
- import your whole Zotero library in one click
- highlight PDFs in the built-in reader (your agent can read the highlights too),
- add a PDF to get its metadata filled in automatically!
Install on Linux (x86_64) or macOS (Apple silicon):
curl -fsSL
github.com/atomashevic/omabi… | sh
First launch sets up Claude Code, Codex or Claude Desktop in one step.
Aleksandar Tomasevic retweeted
What happens when a simulated user gets trashed in the replies?
In YSocial 4.0.0 he gets stressed, posts less, remembers who did it, and may even leave the platform.
Stress/reward, memory and opinion dynamics.
y-not.social/blog/2026/ysoci… #YSocial
Aleksandar Tomasevic retweeted
I really like the AI declaration in this paper:
"This paper was written through human–AI collaboration ... The collaboration included literature search, numerical experimentation, and the proposal and rejection of candidate strategies before the proofs took shape. The authors set the direction, made the mathematical judgments, and take full responsibility for the correctness of the results. Theorems 1.1 and 1.2 and the essential argument leading to them are formalized unconditionally in Lean 4.35.0.."
The optimist in me says "This is the future".
The pessimist has a vision of AI as a God-like oracle, with the ability to answer any mathematics question we could pose, and possessing better taste and aesthetic judgement (and even the capacity to tune it for each person) than we could ever have.
The realist acknowledges that we don't yet know, and that we live in interesting times.
First paper with the legend @KenOno691. As a fun data point, gpt-5.6-sol Ultra tried to solve this for 2 weeks without success.
arxiv.org/pdf/2610.19309
Aleksandar Tomasevic retweeted
🤖 Oh no, the robots are here!
A participant in our experiment wrote that comment when AI agents joined their thread.
We tested the effect of human/AI account labels in a 21-day field experiment in a hybrid simulated social media platform.
36 people and 72 AI agents on a Reddit-like forum, one experimental group with visible labels and one without.
With labels, people replied to AI accounts less often (3.5 pp, 95% CrI 0.5 to 6.5) and criticised them more: 32% of replies to labelled AI were critical vs 17% to unlabelled AI, and about 5% to humans. 🧵
Aleksandar Tomasevic retweeted
The problem with most LLM social simulations is that the agents are way too nice. Even if you fine-tune a model on social media posts, the sycophancy is too strong.
We used an uncensored model instead, Dolphin-Mistral-24B, and our agents got into fights and arguments and left snarky, even toxic comments.
Preprint: osf.io/preprints/psyarxiv/kw…
Aleksandar Tomasevic retweeted
It is in line what we found - that LLMs provide social desirable profile: doi.org/10.1098/rsos.240180
The problem with most LLM social simulations is that the agents are way too nice. Even if you fine-tune a model on social media posts, the sycophancy is too strong.
We used an uncensored model instead, Dolphin-Mistral-24B, and our agents got into fights and arguments and left snarky, even toxic comments.
Preprint: osf.io/preprints/psyarxiv/kw…
Aleksandar Tomasevic retweeted
Fields arranged in the reverse of the order they will be subsumed by AI.
🤖 Oh no, the robots are here!
A participant in our experiment wrote that comment when AI agents joined their thread.
We tested the effect of human/AI account labels in a 21-day field experiment in a hybrid simulated social media platform.
36 people and 72 AI agents on a Reddit-like forum, one experimental group with visible labels and one without.
With labels, people replied to AI accounts less often (3.5 pp, 95% CrI 0.5 to 6.5) and criticised them more: 32% of replies to labelled AI were critical vs 17% to unlabelled AI, and about 5% to humans. 🧵
This matters for policy because labelling is the default proposal for AI accounts on social platforms. In our data, labels reduced engagement with AI accounts and increased criticism of them. Whether that is desirable depends on what the agents are for.
