
Assistant Professor @bengurionu | Fundamental Research Lead @DecartAI | Building efficient, robust foundation models, world models & real-time generative AI
Israel
Joined November 2019
- Tweets292
- Following2.9K
- Followers716
- Likes4.5K
1/7
Some of science's greatest missed opportunities aren't failed experiments. They're discoveries that never get the chance to change the world.
Yesterday, at Nevo Labs and Yad Hanadiv's Bridge to Impact event in Jerusalem, I shared why that matters. ๐งต
6/7
That's the bridge at its best: practical challenges reveal new scientific questions, while discoveries open new possibilities for real systems.
It was inspiring to see Nevo Labs and Yad Hanadiv actively supporting this vision through the Bridge to Impact Grant.
1
81
7/7
Grateful to Scott Tobin for moderating, to fellow panelists Lihi Adler-Abramovich and @RosenblumSerge Serge Rosenblum, and to everyone who made the evening possible.
We should be as ambitious about what discoveries can become as we are about making them.
76
6/7
That's the bridge at its best: practical challenges reveal new scientific questions, while discoveries open new possibilities for real systems.
It was inspiring to see Nevo Labs and Yad Hanadiv actively supporting this vision through the Bridge to Impact Grant.
1
32
7/7 Grateful to Scott Tobin for moderating, to fellow panelists Lihi Adler-Abramovich and
@RosenblumSerge
, and to everyone who made the evening possible. We should be as ambitious about what discoveries can become as we are about making them.
23
Chaim Baskin retweeted
ืืืื ืืืืื ืืืื ืืืช ืืืฉ, ืืจืฉืื ืืืื - ืจืง $0.02 ืืฉื ืื.
ืืืื Lucy 2.6 ืฉื Decart ืืืคืฉืจ ืืขืจืื ืืืืื ืืืื ืืืช ืืืืฆืขืืช ืืงืกื, ืชืืื ืช ืจืคืจื ืก ืื ืฉืืืื ืฉืืื.
ืืขืืจ ืืืืชื ืืื ื ืชืืืืืื ืืืืื ืขื ืืืืื SAM ืืกืืื ืืฆืื ืืืขืงื ืืืจื ืืืืืืงืืื, ืืืฉืื ืืืชื ืขื ืืืื ื ืืกืคืื ืืขืจืืื ืืืืจืืื ืฉื ืืชืืฆืื. ืืื ืขืืืื ืืืชืืฆืื ืืืืชื ืกืืืจื ืืื ืืืืชืจ.
ืืื ืืืื ืื ืกืืจืื ืฉื ืืืืื ืืืงืืืื ืกืืจืื ืขืจืื ืืืืจื. ืืืืืืจ ื ืขืฉื ืืจื WebRTC, ืขื ืชืืืื ื-720P ืื-30 ืคืจืืืืื ืืฉื ืืื. ืืคืฉืจ ืืฉื ืืช ืืช ืืื ืืืืช ืชืื ืืื ืืขืืืื.
ืืคื ืกืจืืื ื ืืืืืื ืืืื ืืืืืืืช ืฉืจืืืชื ืืชืืฆืืืช ื ืจืืืช ืืขืืืืช. ืื ื ืืืืื ืื ืกื ืืขืฆืื. ืื ืืื ืืืง ืืืื ืืืคืงืืช, ืืจืืืช ืชืืฆืืืช ืืืื ืืืช ืืื ืืคืืกื ืคืจืืืงืฉื ืฉืืืงื ืฉืืืขืืช.
ืืืขืื ืืื ืืืืื ืืืืื ืืขืื ืฉืื ืืคืฉืจืืืืช ืฉื ืืชืงืคืืช ืกืืืืจ ืืชืืืืืืช. ืขืื ืืื ืืืืฉืื ืืืจ ืืืื ืืืชื ืืคืฉืจื ืืืืืื. ืขืืฉืื ืื ืืืื ืืืช.
ืืคื Decart, ืืฉืืคืืจ ืืืจืืื ืืืจืกื ืืืืฉื ืืื ืื ืืื ืืช ืืจืคืจื ืก: ืคืจืื ืคื ืื, ืืืืื ืืืืฆืจืื ืฉื ืฉืืจืื ืืื ืืืชืจ ืืืืจื ืืชื ืืขื.
ื-API ืืืื ืืืจ ืืืืืื ืคืืืคืืจืืืช. ืขืืฉืื ื ืืื ืืจืื ืืคืืืงืฆืื ืื ืจืื ืื ืืคืฉืจ ืืขืฉืืช ืืืชื.
Chaim Baskin retweeted
Over the past few days, I dug into the geometry of LoRA to understand how exactly it constraints the geometry of the search space. There's a beautiful connection to the *intrinsic dimension* of a task.
Also learnt I'm saying "difficulty" wrong... ๐คก
0:00 MNIST vs the walking humanoid
1:05 Do we still need fine-tuning?
3:05 What is LoRA?
6:55 What does the rank constraint actually do? (geometric intuition)
9:50 Intrinsic dimension: training along random directions
13:30 Intrinsic dimension of pre-trained LLMs
15:10 Demo: fine-tuning a query router with Qwen3.8-27B
Our spatial memory system helps video world models reconstruct familiar places.
94x less storage. 44ms retrieval from a 1.65-hour archive.
Learn more โ
research.decart.ai/spatial-mโฆ
Chaim Baskin retweeted
Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI.
Perhaps counterintuitively, I think it's a great time to do so.
I wrote up some thoughts on this here: web.mit.edu/phillipi/www/wriโฆ
What if an AI model could take back a harmful answer before it finishes it?
Diffusion language models can.
New Decart research accepted to #EMNLP2026 found harmful responses in just 9โ21% of attacks, vs. 48โ62% for comparable autoregressive models.
We turned that behavior into SRI Guard - detecting harmful generations with just 0.01โ0.04% added latency.
Safety shouldnโt come at the cost of speed.
Read the paper: research.decart.ai/step-wiseโฆ
@INSIGHTLabBGU @TechnionLive @bengurionu @CS_Uni_of_Haifa





