Maligned - February 07, 2026

AI news without the BS

Here’s what actually matters in AI today. No fluff, no hype - just 5 developments worth your time.

Today’s Top 5 AI Developments

1. Gemini-X: DeepMind’s World Model Pushes AGI Horizon 🧠

Google DeepMind just dropped Gemini-X, a new multimodal foundation model showing “early AGI-level performance” on complex, interactive tasks. This model uses a novel “world model” architecture to simulate environments, demonstrating unprecedented understanding and reasoning that pushes us closer to truly intelligent agents.

Source: Google DeepMind Link: https://deepmind.google/blog/gemini-x-agi-world-model

2. LLM Inference Gets a Turbo Boost with Diffusion 🚀

Forget slow, sequential LLM decoding. DFlash uses a lightweight block diffusion model for parallel drafting, achieving a whopping 6x lossless acceleration. This blows state-of-the-art speculative decoding out of the water, making LLM inference dramatically faster and cheaper.

Source: arXiv Link: https://arxiv.org/abs/2602.06036v1

3. Continual Learning: One LoRA to Rule Them All 🔄

Catastrophic forgetting in large models? Share introduces a novel approach using a single, dynamically updated LoRA subspace for efficient continual learning. This slashes parameters by 100x and memory by 281x, making lifelong learning practical without a hundred task-specific adapters.

Source: arXiv Link: https://arxiv.org/abs/2602.06043v1

4. Inverse Problems Just Got Nonlinear: Meet Pseudo-Invertible NNs ✨

This is a foundational shift: new Pseudo-Invertible Neural Networks (SPNNs) generalize the classic Moore-Penrose pseudo-inverse to the non-linear world. It enables “Non-Linear Back-Projection,” letting you precisely control generative outputs or invert complex information loss without retraining. Big implications for precise generative AI.

Source: arXiv Link: https://arxiv.org/abs/2602.06042v1

5. MLLMs Finally Learn to Switch Thinking Modes 🧠↔️👁️

Most MLLMs are rigid, but SwimBird introduces dynamic reasoning: it can switch between text-only, vision-only, or interleaved modes based on the query. This means MLLMs can finally adapt their thinking style, preserving strong textual logic while crushing vision-dense tasks. It’s about adaptable intelligence.

Source: arXiv Link: https://arxiv.org/abs/2602.06040v1

That’s it for today. Stay aligned. 🎯

Maligned - AI news without the BS