The Genie project is very close to my heart, having started my career making AI for simulation games, and studying memory & imagination in the brain, for me it brings all those elements together. Also reminds me of the dream sequences in Inception – science fiction made real…
RESEARCH
-
New video generation from world models before Holodeck
By
–
Before the Holodeck really arrives, we are going to see a new kind of video generation that will come from these new world models. This is the beginning of a new kind of media.
-
Spark: Dynamic Branching for Long-Horizon Agent Learning
By
–
Spark Strategic Policy-Aware Exploration via Dynamic Branching for Long-Horizon Agentic Learning
-

Advancing Open-source World Models
By
–
Advancing Open-source World Models pic.twitter.com/jDjv2k0Tft
— AK (@_akhaliq) 29 janvier 2026Advancing Open-source World Models
-

Self-Distillation Fine-Tuning: On-Policy Continual Learning from Expert Demonstrations
By
–
2026 is the year of continual learning And we are getting some amazing papers towards that This paper introduces Self-Distillation Fine-Tuning (SDFT): on-policy continual learning from expert demonstrations, with no explicit reward inference or engineering The trick here is:
-

World Models Replacing Video Generation in 2026
By
–
World models are going to take over in 2026. While video generation models are impressive dreamers, they aren't world simulators. Generative video models can produce stunning clips, but they hallucinate pixel transitions based on statistical correlations. This leads to
-
Subjective Time Compression: A Conceptual Model Beyond Predictions
By
–
Good question. This is not a predictive model but a conceptual one. The result is highly sensitive to assumptions like expected lifespan, novelty of experiences, health, and cognitive bias. The point is not the exact percentage, but the intuition that subjective time compresses
-
Fine-tuning Resources: SFT, GRPO, and CPT Notebooks
By
–
Text SFT: https://
colab.research.google.com/drive/1vGRg4ks
Rj__6OLvXkHhvji_Pamv801Ss?usp=sharing
…
Vision SFT: https://
colab.research.google.com/drive/1FaR2HSe
91YDe88TG97-JVxMygl-rL6vB?usp=sharing
…
Text GRPO: https://
colab.research.google.com/drive/1mIikXFa
GvcW4vXOZXLbVTxfBRw_XsXa5?usp=sharing
…
CPT (text completion): https://
colab.research.google.com/drive/10fm7eNM
ezs-DSn36mF7vAsNYlOsx9YZO?usp=sharing
…
CPT (translation): https://
colab.research.google.com/drive/1gaP8yTl
e2_v35Um8Gpu9239fqbU7UgY8?usp=sharing
… -

LFM2.5 notebooks update: SFT, GRPO, CPT with vision
By
–
We worked with @UnslothAI to bring up-to-date notebooks for LFM2.5 It covers SFT, GRPO, and CPT with text and vision (SFT-only). Continued pre-training is particularly nice with the base model for text completion and translation examples.
-

SSM Recurrence Security: Mamba State Propagation vs Attention Safety
By
–
3/5 SSMs are recurrent, each token depends on previous state. Mamba decode reads state → updates it. Misclassified request reads garbage → propagates recursively through all tokens. Attention writes K/V first, then attends. Safe even if misclassified.