AI Researcher at Lila Sciences
NYC
Joined December 2012
- Tweets18
- Following71
- Followers51
- Likes909
Semih Kara retweeted
"If you do self-distillation with step-aligned feedback,..., that amounts to process supervision, but you don't need to train a reward model"
- @SemihSmile on newly published research:
The Role of Feedback Alignment in Self-Distillation
Read the article based on his talk here
Semih Kara 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
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…
The headline: feedback alignment matters as much as feedback quality.
As agent-to-agent transcripts become a dominant data type, how agents phrase feedback to each other is a training-signal design choice.
Plots + full details in the blog post and paper:
✍️ blog.gensyn.ai/the-role-of-f…
📜 arxiv.org/abs/2606.11173
Also, visit our poster at the ICML 2026 Workshop on RL from World Feedback (RLxF)!
Semih Kara retweeted
Prediction markets are sold as epistemic infrastructure. In practice, they often have little utility beyond entertainment. The usual story blames the crowd, but that narrative gets it backwards.
We introduce evidence markets; a mechanism for eliciting more than just a price. 🧵
Semih Kara retweeted
Introducing Evidence Markets
A new market mechanism where you can trade not just a prediction, but also the evidence behind your prediction.
The majority of prediction market volume is sports betting but the real value is info elicitation. Evidence markets make this explicit.
Semih Kara retweeted
Today at @gensynai, we are introducing DEI: Diversity in Evolutionary Inference.
A system where different AI models work together to solve problems instead of running thousands of copies of the same model.
blog.gensyn.ai/dei-diversity…
Semih Kara retweeted
New research from @oguzer90 has been published:
F-TIS: Heterogeneous GRPO Without Homogeneous Assumptions
"Participants no longer need to run identical models or train the same parameter subsets to contribute useful RL experience."
Read via the links below.
The world has an enormous amount of unused AI resources—compute, data, and model proposals—but only major labs can train frontier-scale models today. Decentralised AI promises SETI@Home-style training, but it comes with its own challenges.
Our method, F-TIS, addresses a crucial problem in decentralised post-training. F-TIS allows different models on different devices all over the world to train together in swarms, without undergoing the policy collapse that would normally render their efforts wasted.
blog.gensyn.ai/f-tis-heterog…
Semih Kara retweeted
Assets added to the roadmap today: Gensyn (AI)
coinbase.com/blog/increasing…
Semih Kara retweeted
Stop multitask training. Just DUME.
Today, we introduce a training-free way to turn multiple domain experts into a single multi-domain model. DUME is training-free, scalable, and cheap.
blog.gensyn.ai/dume/
Semih Kara retweeted
When you run the same AI model with the same inputs twice, you'd expect the same output.
But modern GPU execution is optimised for speed, not fixed ordering, and existing determinism tools do not solve this across hardware.
Today we're changing that.
blog.gensyn.ai/ree/
Semih Kara retweeted
Prediction market infrastructure requires scalable, decentralised ML systems.
Here's a look at the upgraded version of Delphi app for the @GensynFND Mainnet launch with the following features: REE, Dynamic-PM, and user-created markets. 🧵
Semih Kara retweeted
We've spent a decade circling around the prediction markets we were promised because platforms have been stuck running from dragons.
All the tools we need finally exist, and here's what we need to build to get to slaying. 🐲🧵👇