Next Week in The Sequence:

Our series about evals continues with function calling benchmarks. The opinion section dives into the thesis of how AI creating AI R&A can lead to super intelligence. In engineering, we discuss the first version of Llama Stack. In research, we dive into NVIDIA’s new breakthroughs in AI solving math olympiad problems.

You can subscribe to The Sequence below:

📝 Editorial: The Need for AI Interpretability

Anthropic’s CEO Dario Amodei is one of the most creative thinkers in AI. While doesn’t publish a lot; when he does it always brings some deep reflections about the profound impact of AI. This was the case of his most recent essay "The Urgency of Interpretability," in which he makes a strong case for putting interpretability at the center of modern AI development. As AI systems become core components of finance, healthcare, and national security, Amodei argues it's no longer okay to treat them like black boxes. "These systems will be absolutely central to the economy, technology, and national security," he writes, "and will be capable of so much autonomy that I consider it basically unacceptable for humanity to be totally ignorant of how they work."

Amodei draws a sharp line between traditional software—where you control outcomes through code—and today's large AI models, which are trained on massive datasets and tend to develop unexpected behaviors. These models aren't programmed step-by-step; instead, they're shaped through training, making them more like organisms than machines. Because of this, their internal logic is often hard to follow, even for their creators. And as these models get more autonomy, that lack of transparency turns into a real risk.

To tackle the issue, Amodei outlines Anthropic’s goal: by 2027, develop interpretability tools that can reliably reveal what’s going on inside a model when it makes a decision. Think of it like building an MRI for AI—something that lets us inspect internal components and see which parts are doing what. Anthropic has already made progress, like finding specific circuits in models that track geographic facts—such as which cities are in which states. It’s early days, but these examples show interpretability is within reach.

Still, Amodei is clear that this is just the beginning. AI capabilities are scaling fast, and our understanding of what’s happening under the hood needs to keep up. If we don’t make interpretability a priority, we risk ending up with systems that behave in unpredictable or even dangerous ways. Misalignment might not be obvious until it’s too late. That’s why Amodei views interpretability as a critical safety tool, not just an academic challenge.

He also stresses that this isn’t a problem any one lab can solve on its own. Amodei calls for a broader effort across the AI ecosystem, including governments, to push interpretability forward. One idea is to introduce light-touch regulations that require AI developers to share their safety and interpretability practices. That kind of transparency could build public trust and help the research community make faster progress.

Ultimately, Amodei wants interpretability to move from a niche research topic to a core part of how we build and evaluate AI. As these systems become more powerful and embedded in everyday life, we need to know how they’re making decisions. Understanding AI internals isn’t just useful—it’s essential if we want these systems to be safe, reliable, and aligned with what we care about.

🔎 AI Research

Perception Encoder

In the paper “Perception Encoder: The best visual embeddings are not at the output of the network,” researchers from Meta FAIR introduce a family of large-scale vision-language models trained via contrastive learning.
They show that the strongest visual features lie in intermediate layers rather than final outputs and develop two alignment techniques—language and spatial—to extract task-optimal embeddings for both multimodal and dense spatial tasks.

Greedy Agents

In the paper “LLMs are Greedy Agents: Effects of RL Fine-tuning on Decision-Making Abilities,” researchers from Google DeepMind and JKU Linz investigate decision-making weaknesses in language models.
They identify key failure modes—greediness, frequency bias, and the knowing-doing gap—and demonstrate that reinforcement learning with self-generated rationales significantly improves exploration and action selection in bandit and game environments.

AIMO-2

In the paper “AIMO-2 Winning Solution: Building State-of-the-Art Mathematical Reasoning Models with OpenMathReasoning Dataset,” researchers from NVIDIA detail their winning approach for the AI Mathematical Olympiad. Their solution combines a massive 540K problem dataset, tool-integrated reasoning through code execution, and a generative solution selection model to train state-of-the-art math reasoning models under open licenses.

Tina

In the paper “Tina: Tiny Reasoning Models via LoRA,” researchers from University of Southern California introduce Tina, a family of 1.5B parameter models trained with LoRA and RL to deliver high reasoning performance at low cost. Tina achieves strong results—such as 43.33% Pass@1 on AIME24—with less than $9 in post-training cost, demonstrating that lightweight RL fine-tuning can effectively instill reasoning in small models.

Paper2Code

In the paper “Paper2Code: Automating Code Generation from Scientific Papers in Machine Learning,” researchers from KAIST and DeepAuto.ai present PaperCoder, a multi-agent system that turns ML papers into executable code repositories. The framework operates through planning, analysis, and generation phases, and outperforms baselines on benchmarks like PaperBench, with 77% of outputs rated best by authors of the original papers.

UFO²

In the paper “UFO²: The Desktop AgentOS,” researchers from Microsoft and collaborating universities describe the contributions of a deeply integrated OS-level system for desktop automation. UFO² introduces a centralized HostAgent with modular AppAgents, hybrid GUI-API action orchestration, speculative execution, and a Picture-in-Picture desktop, enabling robust and concurrent automation across over 20 Windows applications​

🤖 AI Tech Releases

ChatGPT Image API

ChatGPT adds its newest image generation model to its API.

🛠 AI in Production

Language to SQL at Salesforce

Salesforce discusses implementing language to SQL capabilities in their internal agents.

📡AI Radar

• AI agent sensation Manus raised $75 million in a new round.

• Elon Musk’s XAI Holdings is in early talks to raise $20 billion.

• Google DeepMind CEO Demis Hassabis gave was interviewed at 60 Minutes.

• OpenAI expanded a lightweight version of Deep Research to Plus, Team and Pro users.

• Mechanize launched with the intriguing to build environments and benchmark for work automation.

• Dropbox adds new AI-powered search features.

• Endor Labs raised $93 million in new funding for its AI code scanning platform.

• Noxtua raised $92.2 million for its AI optimized for the German legal system.

• Manychat raised $140 million to grow its AI customer engagement platform.

• AI coding assistant Zencoder announced the acquisition of Machinet.