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).
DAGFormer Plans Connections Between Layers and Cuts Perplexity by 17–25%
DAGFormer adds direct cross-layer connections to the transformer and plans them with a topology predictor. On models up to 1B, WikiText-2 perplexity drops by 17–25%.
Researchers Find a 0.4–1.1% Window in Pythia Pretraining Where the Readout Reorganizes
A model's hidden states encode a task before its output shows it. The gap traces back to the unembedding matrix, which reorganizes during 0.4–1.1% of Pythia's…
semCER Weights OCR Errors by Severity: 61M LLM-Judge Ratings
The new semCER metric keeps CER's alignment but has an LLM judge score every recognition error. The authors released 61M rated edits across 16 languages.
Modal raises $750M at $15.75B valuation and reaches $1B run rate
Agent teams for $155 instead of $510: Cohere finds where delegation actually pays off
WinARDet fixes looping in MLLM detectors: F1 rises from 32.16 to 39.36
GAFU pulls knowledge out of generation in UMMs and lifts VStar-Bench by 20%
RTAformer Teaches a Model to Notice When Relationships Between Objects Change in Video
Researchers propose RTAformer, a one-stage transformer for video scene graphs. It models object persistence and relationship changes separately, and improves recall…
MolLean Makes LLMs Check Their Chemistry: Wins on 12 of 12 Tasks
MolLean, a new Lean-based language, verifies every step of a model's chemical reasoning. On Qwen3.6-27B it beat plain chain-of-thought on all 12 tasks across 8,800…
FLOMSS: flow matching splits music into 4 stems and takes 2nd place in listening tests
SliderEdit++ Smooths Out Sliders in AI Image Editors: +18.5% and Up to 20x Faster
JockeyBench: 28 LLM Agent Configurations All Lost Money on Japanese Horse Racing
KANLab: A TensorFlow Playground Fork That Shows KAN Training in the Browser
Isomorphic Labs Uses Gemini 4 Argon to Parse 500-Page Patents in a Day Instead of a Month
Isomorphic Labs researchers apply the multimodal Gemini 4 Argon to chemical patents: a structured report takes one day instead of a month of manual reading.
Habr Author Tests How Many Edits It Takes to Erase an LLM Watermark
A Habr author tests whether Claude and OpenAI watermarks survive rewrites, error fixes and code. Four questions are posed; the results are still to come.