Zhemal Khamidun
Head of AI at Alpina Digital, CPO of AlpinaGPT (~30-person engineering & AI team). Founder of Vibe Academy & Hamidun Consulting, author of the AI channel «Готовим ИИшницу» (@jemal_hamidun).
A USTC Team's Startup Raised Tens of Millions Three Months After Founding
A USTC team put a world model in a robot that reasons backward from the desired outcome. The startup is three months old and has raised tens of millions.
正行创新 Unveils Physical AI for Retail: Robots Working the Shelves 24/7
At APRCE 2026, 正行创新 showed a robot paired with a management platform for stores: shelf restocking, inspection rounds and cargo carrying, running 24/7.
Doubao adds an infinite canvas for creative work and a 2.1 Lite model
ByteDance's Doubao work mode gets a Canvas feature for complex creative tasks and a lightweight Doubao 2.1 Lite model. Everything sits on one infinite page.
LVSPM Beats VGGT at Camera Pose Estimation on 16–256 Frames and Doesn't Break Down on Large Scenes
TopoGPU Computes Persistent Homology 53x Faster on Average
GHARP animates a 3D head in real time: 13x faster on A100
SAVER Teaches Video Models to Reason From a Few Frames, Using 1,250 Examples
Distilled Diffusion Models Learn to Forget Unwanted Concepts via DPO
Researchers adapted Direct Preference Optimization to few-step diffusion models, removing identities and NSFW content without re-distillation or a loss of speed.
SPLIT-RL Teaches VLMs to See and Reason Separately: +1.4–6.1 pp Over GRPO
SPLIT-RL, a new post-training scheme, splits vision-language model training into perception and reasoning phases and beats GRPO by 1.4–6.1 points on Qwen3-VL and…
VICO Trains VLMs on Rewritable Images: Up to +5.0% on New Tasks
DOGS Beats 9 Baseline Methods at Logo Generation Across Two Renderers
MESSENGER cuts 3D motion forecasting error by 71.6% on nuScenes
LinSlot Teaches a Neural Network to Break a Scene Into Objects and Their Attributes Without Labels
Researchers Prove Why Adversarial Training Makes Saliency Maps Sparse
A new arXiv paper proves that for two-layer ReLU networks, the gradient-norm penalty under ℓ∞ attacks is anisotropic and pushes saliency maps toward sparsity.
Distilling ECG into PPG Memorized Identity: Fixed-Effects More Than Doubled the State Signal
A static "person trait" makes up 40–59% of the ECG teacher's target signal. Fixed-effects distillation subtracts it and more than doubles state agreement.