Gemini Omni turns any input into video, reasoning halved

Google unveiled Gemini Omni at I/O 2026, a multimodal model that generates video from any combination of image, text, or audio inputs. The system embeds a SynthID watermark for provenance and promises future support for broader media outputs, signaling a major step toward universal AI content creation.

Gemma‑4 models collapse into endless repetition on long enumeration prompts, a glitch traced to a handful of MLP neurons. Flipping the sign of a single neuron's weight eradicates the loops while preserving benchmark scores, but the deeper “doom loops” persist, revealing limits of weight surgery.

NULLs embed source-specific knowledge into isolated sink neurons, letting you erase an individual data source by simply disabling its sinks. The method scales to millions of Wikipedia articles and preserves shared facts and overall language performance, matching full retraining without extra gradients.

SuperThoughts packs pairs of consecutive CoT tokens into one latent vector and decodes two tokens per step via a lightweight Multi‑Token Prediction module, cutting inference length by 20‑30% with only 1‑2 point accuracy loss on math benchmarks. This speeds up long‑horizon reasoning without abandoning discrete token supervision.

The paper introduces Risk‑Aware Causal Gating (RACG), a framework that gauges counterfactual risk before letting an LLM‑driven agent act, defer, or abstain. By separating causal risk from predictive confidence, RACG cuts high‑cost mistakes dramatically while preserving most of the system’s utility, offering a principled least‑privilege safety primitive.

The paper shows that using data selection to prevent model collapse can itself cause collapse when the verifier’s reference distribution is biased, as happens in siloed datasets like healthcare or finance. It proves the effect theoretically and offers a collaborative Wasserstein proxy to preserve diversity.

AI model AlphaProof Nexus solved 9 open Erdős problems and proved 44 OEIS conjectures at a few hundred dollars each, showing cost-effective AI for formal math research. This demonstrates that AI can now autonomously generate machine‑checked proofs for longstanding open problems, potentially accelerating research across fields.

The Deepfake Liability Act, signed into law in May 2025, forces covered platforms to set up a 48‑hour notice‑and‑removal process for non‑consensual intimate images and AI‑generated deepfakes. Critics warn the broad language could swamp sites with frivolous claims and strain free‑speech protections. Platforms now face legal duty of care or risk liability under the amended Section 230.

The ds4 engine, a Metal‑native inference runtime from Redis creator antirez, runs DeepSeek V4 Flash on Apple Silicon laptops with 96 GB RAM by streaming KV cache to SSD. This unlocks 284‑billion‑parameter models on a MacBook Pro M3 Max, delivering ~26 tokens/sec generation with a 1‑M‑token context.