Blockchain, AI and Applied Cryptography | Building Decentralized Compute @gensynai | CS PhD from TU Delft | EE & Math from Bogazici University
Delft, Nederland
Joined December 2011
- Tweets146
- Following334
- Followers651
- Likes2.6K
Last week, I attended Dagstuhl Seminar 26401, "Divided We Stand: Security and Privacy of Collaborative Learning." It was a great opportunity to meet fellow researchers, exchange ideas, and discuss open challenges in the field.
dagstuhl.de/26401
I presented "Attacks and Defenses in Decentralized Collaborative Learning: New Directions in the LLM Era." Drawing on our recent publications, I explained how LLMs change the threat landscape and discussed the implications for securing distributed training.
Thanks to the organisers, Alexandra Dmitrienko, Antonino Nocera, @StjepanPicek_ , and Ahmad-Reza Sadeghi, and all the participants for a stimulating week!
I'm excited that our paper with @ErosFani, “Training-Free Dynamic Upcycling of Expert Language Models,” has been accepted to NeurIPS 2026! 🎉
We introduce DUME: a way to combine already-trained domain experts into one Mixture-of-Experts model without post-merge training. 🧵
We evaluated models from 115M to 9B parameters across language modeling and reasoning tasks. In the 3B reasoning setting, DUME averaged 96.8% of the domain experts’ performance, and in the 9B reasoning setting, it even outperformed them.
Trust open-weight models no more, audit the training yourself!
We introduce OPEN-1B, a fully auditable LLM whose training steps can be replayed and verified bit for bit.
open-1b: the first model you don’t have to trust
Auditable training is the best defense against the future of AI we’re being warned about
open-1b is a milestone toward verifiable AI, a goal that is absolutely necessary for the future of intelligence
gensyn.ai/news/introducing-o…
Inference does not have to look like one centralised gateway routing every prompt to a fixed model catalogue.
With IR3DE-AXL, we built a collective inference network prototype where participants can contribute expert models, use models served by others, or do both.
gensyn.ai/research/ir3de-axl…
🧵1/6
This is still a prototype.
The router is intentionally lightweight, so it is not perfect. Ambiguous prompts or overlapping domains can still be routed incorrectly.
The current version also focuses on the core collective inference loop, not built-in verification, reputation, or incentives.
IR3DE-AXL turns local expert routing into a collective inference network.
The network becomes more useful as participants contribute their own expert models, while each user keeps local control over which providers and expertise areas they rely on.
gensyn.ai/research/ir3de-axl…
Oguzhan Ersoy retweeted
Gensyn at @ICML2026 Seoul July 6-11
Jul 8-9 connect x collaborate x cowork
by Gensyn and @OptiverGlobal
Jul 8 Pitch your Poster by Gensyn
Jul 9 Research to Industry: Fireside Chat & Happy Hour by Gensyn and Optiver
Jul 10-11 Four Gensyn paper presentations
Thread below👇
Oguzhan Ersoy retweeted
Gensyn Research Mega-Thread🧵
Seven Gensyn papers have been accepted to:
@aclmeeting Findings
@icmlconf
@TheOfficialACM EC
The common theme: machine collaboration shouldn't require the same model, hardware or assumptions. Decentralisation as a design constraint, not an addon.
Oguzhan Ersoy retweeted
Wednesday at 12pm ET - join @semihsmile in talking through the recently published paper:
"The Role of Feedback Alignment in Self-Distillation"
Links in the thread below.
nitter.gq/i/spaces/1mGPaanXqBMJN
Oguzhan Ersoy retweeted
Multi-agent LLM systems are everywhere, and agents talk to each other: critiquing, correcting, giving feedback.
A design lever: what style of agent-to-agent communication works best?
At @gensynai, we designed one. New style, same agents, better results.
blog.gensyn.ai/the-role-of-f…
Oguzhan Ersoy retweeted
Look beyond one-size-fits-all LLMs with IR3DE
New research published by Gensyn's @ErosFani and @oguzer90
Read the blog post linked below and follow through to the published paper on @arxiv
Look beyond one-size-fits-all LLMs with IR3DE
Introducing IR3DE: a router that selects the best expert LLM for any prompt.
No complex neural network or gradient-based training—just linear algebra and a linear layer achieving 98.4% in reasoning settings
blog.gensyn.ai/look-beyond-o…
