프런티어 AI 안전 평가, 에이전트 실행 인프라와 반복형 추론 연구

Dario Amodei

@DarioAmodei

2일 전

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training. You can read the full post here:

Sam Altman

@sama

2일 전

I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.

Min Choi

@minchoi

2일 전

OpenAI just made it stupidly easy to build real AI agents. The new Agents API basically gives developers the Codex harness as an API: • One call → cloud agent • Long-running sessions • Automatic context compaction • Parallel subagents • MCP + built-in tools • OpenAI-hosted sandboxes • Or run on your own infra This is way more than "call a model and hope." You give it a task, model, tools + environment. It can work across files, run code, use tools, delegate work to subagents and keep going across multiple context windows. Public beta is live for all developers. No extra Agents API fee - you pay for the models + tools your agent uses. This is basically Codex infrastructure for anyone building agents.

elvis

@omarsar0

2일 전

Learn to build a harness, folks. It's not surprising to me that so many YC builders want to build domain-specific harnesses. If you work long enough on a domain-specific problem, you quickly realize the opportunity. But you also realize how important that harness will be to stay competitive in the agentic era. From a product perspective, harnesses open up interesting new surface areas and experiences for the services/products you provide. From a technical perspective, harnesses are how you build and maintain a framework and set of best practices for how your users/customers interact with what you offer. Understanding how to build and design a harness means you can build much stronger intelligence stacks, given that you can customize it and understand the domain well. That's extremely valuable. It may not seem apparent yet, but a harness wave is coming. If you are getting started, give this list of harness papers to your agents and start upskilling: https:// academy.dair.ai/papers/collect ions/harness-engineering … If you are already a builder, try building one for your specific domain. It's a lot of fun, and you learn a lot of interesting things to enhance your current agentic tools.

Interesting paper to improve recurrent reasoning. Looped models are great because you get more reasoning out of a model without adding parameters. So this work proposes a looped architecture with a new training method. The authors report wins over prior looped models on five of six reasoning benchmarks. More details from the paper: Looped models reason by updating a hidden state again and again at inference time. The hard part is training. Gradients usually flow through only the last one or two updates, so the early updates never learn to set up the later ones. Looped flows train the recurrence with local denoising objectives, the way flow models are trained. Noise levels decrease step by step and share the same noise sample, which ties each update to the next. At inference the model follows a probability flow. A finer time grid spends more compute, and different starting noise can produce different valid answers on tasks with more than one solution. Paper: https:// arxiv.org/abs/2609.11801 Chat with Paper: https:// academy.dair.ai/papers/thinkin g-with-looped-flows-2609.11801 …

Nearly a year of Code Arena: WebDev progress compressed into 15 seconds. Each line follows the highest-scoring model from top labs over time, showing the pace of improvement across the ecosystem. In just the last year, the model in the leading spot increased score by +340 pts and the number of frontier labs competing for the top spot expanded from 6 to 10. @AnthropicAI has dominated throughout the year. Although standout releases have jumped to the top spot, most notably the Chinese open-source model Kimi K3 from @Kimi_Moonshot in July. Today, GPT-6 Astra by @OpenAI leads with 1796 pts, followed by @claudeai Fable 5.1 at 1764 pts. The next closest lab is 103 pts away, @Alibaba_Qwen with 1685 pts. Code Arena: WebDev ranks models through head-to-head user preference on real front-end web development tasks. These votes drive the leaderboard that is tracking the frontier.

Peak performance across hard benchmarks: • Best or joint-best on 5/8 benchmarks (DeepSWE, Chartography, Toolathon, GDP.pdf, SWEFish) • Chartography: 48.3 (outperforming Opus 5 & Fable 5) • DeepSWE: 74.3 Achieved without Fable 5, Fable 5.1, or GPT-6-Astra in the agent pool. Details: https:// sakana.ai/fugu-max-relea se/ …

“World in World: Explore the World with World Models” Video world models can generate long rollouts, but controlling them from new viewpoints usually needs task-specific training or adapters. This paper instead turns source frames, geometry, and past generated states into visual evidence that a frozen world model can directly read through self-attention. This then gives training-free camera-controlled rerendering with better long-horizon consistency, unseen-view completion, and lower camera error than prior methods. https:// alphaxiv.org/abs/2609.11548