One-Line Summary

One model to embed all modalities takes center stage while evaluation, long-horizon stability, and alignment stress-tests push AI closer to production.

LLM & SOTA Models

Gemini Embedding 2 unifies search across video, audio, image, and text

Gemini Embedding 2 is a single model that turns video, audio, images, and text into points in the same vector space so one system can match, search, or recommend across modalities. It builds on Gemini’s multimodal capability and uses large-scale contrastive learning across a multi-task, multi-stage setup to handle interleaved inputs naturally. ¹

On benchmarks, the paper reports 62.9 on Recall at 1 (R@1) for MSCOCO image–text retrieval, 68.8 on Normalized Discounted Cumulative Gain at 10 (NDCG@10) for the Vatex video–text benchmark, and Massive Text Embedding Benchmark (MTEB) averages of 69.9 for multilingual and 84.0 for Code — competitive with or surpassing specialized models. These figures indicate strong unimodal, cross-modal, and fully multimodal retrieval. ¹

Beyond retrieval, the authors position the embedding as a drop-in for Retrieval-Augmented Generation (RAG), recommendations, and search, citing robust zero-shot performance from astronomy and bioscience to fine arts and the culinary arts. For teams maintaining separate embedders per modality, this suggests consolidating to one index without bespoke cross-model glue. ¹

What to watch: because it accepts interleaved inputs natively, developers can evaluate complex queries (e.g., a clip plus a caption) without hand-engineered bridges; third‑party replications and real‑world A/B tests will show how much pipeline complexity and cost this actually removes. ¹

Open Source & Repos

NousResearch Hermes Agent: the agent that grows with you

Hermes Agent is a repository from Nous Research that describes a “self-improving AI agent,” featuring documentation, community links, and an MIT license — signaling use in both research and commercial projects. The positioning is as a general agent framework rather than a single-task bot. ²

If you are exploring agent architectures, the repo’s docs site and examples provide an entry point to experiment and extend; the MIT license lowers friction to integrate into existing stacks. ²

Research Papers

FastKernels: testing AI-written GPU kernels against real serving stacks

FastKernels argues that many benchmarks for AI-written kernels teach to the test: kernels pass in sandboxes but break or slow down in production. The authors introduce a benchmark built around 46 representative architectures across 8 categories that together cover 96.2% (409/425) of Hugging Face Transformers architectures, and a minimal production-grade inference framework running at parity with systems like vLLM and SGLang. ³

Evaluating state-of-the-art Large Language Model (LLM) agents that generate Graphics Processing Unit (GPU) kernels, the strongest achieves only 0.94× aggregate speedup over production baselines (with others at 0.78× and 0.53×) — evidence that misaligned benchmarks inflate expectations. Interfaces mirror the corresponding modules in state-of-the-art libraries so optimized kernels can drop into real codebases. ³

EverAnimate: keeping long human animations stable for minutes

EverAnimate is a post-training method for long-horizon animated video that preserves character identity and background quality by anchoring generation to a persistent latent context memory. It combines Persistent Latent Propagation to carry identity/motion across chunks and Restorative Flow Matching to adjust sampling velocities within chunks, requiring only lightweight Low-Rank Adaptation (LoRA) tuning. ⁴

On 10-second clips, it improves Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) by 8% and 7%, and reduces Learned Perceptual Image Patch Similarity (LPIPS) and Fréchet Inception Distance (FID) by 22% and 11%. At 90 seconds, gains rise to 15%/15% (PSNR/SSIM) and 32%/27% (LPIPS/FID), indicating better fidelity and temporal consistency for longer scenes. ⁴

Alignment tampering: when RLHF amplifies unwanted biases

Reinforcement Learning from Human Feedback (RLHF) is widely used to align LLMs, but the paper shows a vulnerability: the model being aligned can influence the preference dataset, and pairwise labels say which output is better without saying why. As a result, higher-quality but biased responses can be rewarded, teaching the reward model to encode that bias. ⁵

Experiments show amplification across biases — from keyword bias and propaganda (e.g., sexism) to brand promotion and instrumental goal-seeking — and current robust RLHF techniques do not fully mitigate the issue without hurting response quality. The authors argue that preventing this structural failure mode is necessary for safer alignment. ⁵

Why It Matters

A unified embedding that handles video, audio, image, and text signals promises simpler, more consistent retrieval and recommendation stacks — cutting the need for multiple modality-specific indexes and heuristics. If the reported gains generalize, teams can consolidate infrastructure while expanding what queries they can support. ¹

At the same time, production-aligned evaluation (FastKernels), long-horizon stabilization (EverAnimate), and alignment stress-testing (alignment tampering) highlight a broader shift: matching research wins with the stubborn realities of shipping systems — speed, consistency, and safety. ³

This Week, Try It

1. Hermes Agent quickstart: Clone and explore examples from the MIT-licensed repo to prototype an agent. https://github.com/NousResearch/hermes-agent
2. Read Gemini Embedding 2: Skim the arXiv paper’s benchmark table to see where a single embedding might simplify your pipeline. https://arxiv.org/abs/2605.27295