Next Week in The Sequence:

1. Our series about interpretability continues with an intro to mechanistic interpretability.

2. The AI of the week is going to cover the new Hermes 4 model pretrained mostly using synthetic data.

3. In the opinion section, we discuss some of the dynamics of NVIDIA, Huawei and Intel in the chip geo-political wars.

Subscribe Now to Not Miss Anything:

📝 Editorial: Flash, But Precise: Inside Gemini 2.5 Flash Image

Gemini 2.5 Flash Image (internally nicknamed “nano-banana”) is Google’s new native image generation and editing model, designed to combine low-latency, cost-efficient inference with materially better visual quality and controllability than the Gemini 2.0 Flash image features. The model exposes four first-class capabilities: multi-image fusion (compositing), character/asset consistency across prompts and edits, fine-grained prompt-based local edits, and edits grounded by Gemini’s world knowledge. It’s available now in Google AI Studio, the Gemini API, and Vertex AI.

Architecturally/operationally, Flash Image is positioned as a native image model rather than a multimodal text model with an image head. That allows it to support targeted transformations and template-driven generation while leveraging the broader Gemini family’s semantic priors (“world knowledge”) for more faithful, instruction-following edits. In practice, this lets a single prompt both understand a sketched diagram and apply complex edits in one step, reducing orchestration overhead in apps that previously required separate vision+image-edit pipelines.

For production asset pipelines, the biggest unlocked workflow is persistent character or product identity: developers can place the same character/object into diverse scenes while preserving appearance, or generate a catalog of consistent brand assets from a single visual spec. Google ships a Studio template demonstrating this behavior, and the model also adheres well to strict visual templates (e.g., real-estate cards, uniform badges), making it suitable for programmatic layouting and bulk creative ops.

The editing toolchain supports precise, text-addressed local edits—blurs, removals, pose changes, colorization, background swaps—without manual masks, enabling granular transformations controlled entirely by natural language. Because edits are semantics-aware, they can chain with understanding tasks (e.g., “read this hand-drawn diagram and color-code the vectors, then remove the annotation in the bottom-left”), which shortens multi-stage image processing flows.

Multi-image fusion lets the model ingest several input images and synthesize a coherent composite, such as dropping a photographed product into a new environment or restyling interiors with target textures/palettes. Google’s demo app exposes this as a drag-and-drop workflow; in code, the same is a multi-part prompt mixing text and images and requesting a single fused output. This capability is particularly useful for virtual staging, synthetic lifestyle photography, and rapid A/B creative generation.

Ecosystem-wise, Flash Image is in preview (stabilization coming “in the coming weeks”), ships with updated “build mode” in AI Studio (template apps, quick deploy or save-to-GitHub), and is also being distributed via OpenRouter and fal.ai. All generated/edited images carry an invisible SynthID watermark to support provenance and attribution. On public human-preference leaderboards, the preview model currently ranks first for both Text-to-Image and Image Edit on LMArena, indicating strong early quality and edit fidelity. Google calls out active work on long-form text rendering, even tighter identity consistency, and finer factual detail in images.

🔎 AI Research

Title: STEPWISER: Stepwise Generative Judges for Wiser Reasoning

AI Lab: FAIR at Meta & collaborators
Summary: This paper introduces STEPWISER, a reinforcement learning–trained generative judge that evaluates intermediate reasoning steps in large language models through a meta-reasoning process. It outperforms traditional supervised process reward models by providing more accurate feedback, improving inference-time search, and enhancing downstream training.

Title: UQ: Assessing Language Models on Unsolved Questions

AI Lab: Stanford University & collaborators
Summary: The authors propose UQ, a benchmark of 500 challenging, unsolved Stack Exchange questions to evaluate models on problems without ground-truth answers. They introduce validator pipelines and a community-driven platform, showing that frontier models solve only a fraction of tasks, thereby pushing evaluations toward open-ended, real-world knowledge gaps.

Title: Hermes 4 Technical Report

AI Lab: Nous Research
Summary: Hermes 4 is a family of open-weight hybrid reasoning models trained with large-scale synthetic data, rejection sampling, and specialized environments to integrate multi-step reasoning with broad instruction following. It achieves strong benchmark results across math, coding, knowledge, and alignment tasks, while exhibiting flexible persona adoption and stylistic transfer.

Title: A Scalable Framework for Evaluating Health Language Models

AI Lab: Google Research
Summary: This work introduces Adaptive Precise Boolean rubrics, a framework for evaluating health LLMs that replaces subjective Likert scales with targeted yes/no rubrics to improve reliability and scalability. Tested on metabolic health queries and real patient data, the method halves evaluation time, boosts inter-rater agreement, and enables automated assessments comparable to expert review.

Title: Autoregressive Universal Video Segmentation Model (AUSM)

AI Lab: NVIDIA, CMU, Yonsei University & NTU
Summary: AUSM reformulates video segmentation as sequential mask prediction, unifying prompted and unprompted tasks within a single autoregressive, state-space architecture. It enables parallel training for long video streams, achieves state-of-the-art performance on multiple benchmarks, and delivers up to 2.5× faster training compared to iterative baselines.

Title: rStar2-Agent: Agentic Reasoning Technical Report

AI Lab: Microsoft Research
Summary: This work presents rStar2-Agent, a 14B math reasoning model trained with agentic reinforcement learning that integrates Python tool use, reflection on code execution, and step refinement. Using an efficient infrastructure, a new GRPO-RoC algorithm, and a compute-light multi-stage RL recipe, the model achieves frontier-level performance—80.6% on AIME24 and 69.8% on AIME25—surpassing much larger models like DeepSeek-R1 while producing shorter, more efficient reasoning traces

🤖 AI Tech Releases

Gemini 2.5 Flash Image

Google released Gemini 2.5 Flash Image with editing capabilities.

gpt-realtime

OpenAI released its Realtime API for voice agents available for developers.

Claude for Chrome

Anthropic unveiled Claude as a Chrome extension.

📡AI Radar

• Maisa AI raised $25M led by Creandum and launched a model-agnostic “Maisa Studio” for supervised agentic “digital workers” aimed at making enterprise automation reliable after most gen-AI pilots flop.

• Nvidia’s latest outlook underwhelmed investors, fueling worries that AI-chip demand is decelerating—especially with China sales excluded—pressuring the stock and peers.

• Malaysia’s SkyeChip introduced the MARS1000, the nation’s first homegrown edge AI processor, underscoring its push into AI hardware amid tighter export controls.

• Assort Health reportedly raised about $50M at a ~$750M valuation to scale AI voice agents that handle patient calls, scheduling and FAQs for specialty clinics.

• Meta is forming a California super PAC and pledging tens of millions to support candidates favoring lighter AI regulation, ramping its state-level political spending.

• Accel’s new investment lifts Vercel’s valuation to roughly $9B—about triple its prior mark—as the Next.js company expands its AI-powered developer cloud.

• Google expanded NotebookLM’s auto-generated Video Overviews and enhanced Audio Overviews to 80 languages, taking its study aids well beyond English.

• MathGPT.ai, which uses a Socratic “no-direct-answers” approach with new LMS integrations, is expanding from a 30-campus pilot to 50+ institutions this fall.

• Ant Group’s quarterly profit fell about 60% as heavy spending on AI and global expansion squeezed margins despite growth ambitions.