Who's Who of AI

Eugene Vinitsky

Reinforcement learning and autonomy researcher

733 trust researcher @eugenevinitsky.bsky.social · 12,775 followers @eugenevinitsky · 0 followers
Anti-cynic. Towards a weirder future. Reinforcement Learning, Autonomous Vehicles, transportation systems, the works. Asst. Prof at NYU https://emerge-lab.github.io https://www.admonymous.co/eugenevinitsky
What they're sharing

Articles & links

New Paper: arxiv.org/abs/2606.19370 Self-play yields capabilities but requires frustrating cost-function tuning. Surprisingly, just 30 minutes of demonstration data produces much more human-like driving policies! Led by @daphne-cornelisse.bsky.social Website: spiced-self-play.com

Human-like autonomy emerges from self-play and a pinch of human data arxiv.org
AI Weekly's analysis →
  • A new method trains driving AI using only 30 minutes of human demonstrations, 2,500 times fewer than comparable imitation learning approaches.
  • Human demonstrations serve as a regularization signal on top of a basic goal-reaching reward, not as the primary training objective.
  • Resulting policies coordinate with held-out human trajectories and finish training in 15 hours on a single consumer-grade GPU.
Read full analysis →
View on Bluesky · ♥ 104 ↻ 18 ↩ 1 · 2 from the directory shared this · 19d ago
↻ Eugene Vinitsky reposted
Ida Momennejad @neuroai.bsky.social

New work with Roberta Raileanu: A Compositional Framework for Open-ended Intelligence Open-ended intelligence is the capacity to adapt to novel problems & environments that are substantially different from those seen in training. But most models of open-mindedness don't have c…

A Compositional Framework for Open-ended Intelligence arxiv.org
AI Weekly's analysis →
  • Momennejad and Raileanu define open-ended intelligence as compositional closure L(P,C) from minimal primitives and operators, not behavioral diversity.
  • The framework proposes 'next primitive prediction' training, targeting reuse of algorithmic primitives rather than next-token or latent-state prediction.
  • Three evaluation metrics (PRI, CDG, TaR) are proposed to measure compositional generalization, but the paper presents no empirical benchmark results.
Read full analysis →
 View on Bluesky →
↻ Eugene Vinitsky reposted
@bernhard-jaeger.bsky.social

🔬 This week's KE:SAI research highlight is World Engine, a RL post-training simulator and pipeline for end-to-end driving. 📜 arxiv.org/abs/2606.19836

World Engine: Towards the Era of Post-Training for Autonomous Driving arxiv.org
AI Weekly's analysis →
  • World Engine raised rare-scenario closed-loop success on nuPlan from 73.66% to 88.89%, a gain of 15.23 percentage points.
  • Production deployment at Huawei ADS cut cut-in collisions 45.5% and pedestrian/cyclist collisions 15.8%.
  • Post-training gains proved equivalent to roughly 14 times more pre-training data, per the paper's own analysis.
Read full analysis →
 View on Bluesky →
↻ Eugene Vinitsky reposted
Kristin Branson @kristinmbranson.bsky.social

Agentic coding is genuinely useful now, and there are some impressive reports of AI agents doing science. But how well and how reliably can they handle tasks scientists actually want to hand off, ones that bottleneck progress? How do we even measure that?? New paper🧵 arxiv.org…

A case study of evaluating AI agents on a neuroscience data-to-discovery pipeline arxiv.org
AI Weekly's analysis →
  • Coding agents solved individual stages of a fly optogenetics pipeline but could not complete the full end-to-end discovery run.
  • Agents struggle most when there is no predefined criterion to iterate on and must use their own scientific judgment, a key open challenge.
  • The study flags challenges absent from standard benchmarks: computational resource management and generalization to large held-out datasets.
Read full analysis →
 View on Bluesky →

At Percepta, we've helped Summa Health automate the process of fairly scheduling nurse shifts. This is the type of work I'm really proud of, clearly using ML to make folks lives easier and better www.percepta.ai/blog/buildin...

Building the AI-native hospital | Percepta percepta.ai
View on Bluesky · ♥ 48 ↻ 4 ↩ 2 · 2 from the directory shared this · 8d ago
Their own posts

Recent commentary

Given the concerns about AI writing, be aware that Pangram has a super high false positive rate! It frequently rates my personal writing as 100% LLM written!

View on Bluesky · ♥ 95 ↻ 7 ↩ 3 · 50d ago

A lot of my fears about LLM usage come from using Stackoverflow when I was a wee programmer. It was easy to fall into copy pasting behavior that long term made you worse

View on Bluesky · ♥ 289 ↻ 22 ↩ 7 · 1d ago

The median AI safety person is a weirdo who was into it 10 years before there was any system that appeared to match their concerns. It’s more often correct to treat them as honest than disingenuous

View on Bluesky · ♥ 221 ↻ 14 ↩ 6 · 34d ago

arxiv paper now on hold for almost a week. To all the people submitting LLM slop that brought us to this place, please know that you've appreciably made the world worse

View on Bluesky · ♥ 197 ↻ 15 ↩ 5 · 21d ago

Google's AI search hitting new heights. Unsure how this one happens tbh

View on Bluesky · ♥ 157 ↻ 32 ↩ 8 · 39d ago

Proposed norm: you're allowed to respond to an unedited LLM document with an unedited LLM document that is twice as long

View on Bluesky · ♥ 172 ↻ 17 ↩ 7 · 14d ago

I have already met some number of people who are basically a front-end to an LLM

View on Bluesky · ♥ 157 ↻ 18 ↩ 8 · 4d ago

Tech has become increasing dismissive of academia, but I feel like I'm orders of magnitude more useful to society right now producing open work than if I was helping slightly improve an LLM metric

View on Bluesky · ♥ 180 ↻ 8 ↩ 1 · 18d ago

Every person who is deceptively forcing me to read LLM text, I want you to know that I don't like you. This isn't an opinion, it's a full-fledged declaration.

View on Bluesky · ♥ 138 ↻ 6 ↩ 3 · 30d ago

Everything is sports now, with professional commentators and fandoms. Personally I think Karpathy is going to have a great season this year after being picked up by Anthropic, yielding at least 2 points on LLM benchmarks

View on Bluesky · ♥ 119 ↻ 9 ↩ 8 · 15d ago
Their network

In Eugene Vinitsky's orbit

Center = Eugene Vinitsky. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.