GPT-6 Astra 확산, Atlas 공간 지능, 장기 에이전트와 추론 효율 연구

GPT-6 Astra (Max) by @OpenAI is #1 on Code Arena: WebDev reshaping the Pareto frontier! It is SOTA performance at $40/Mtoken, matching the latest Claude model pricing. In Code Arena, AI models are ranked on web development tasks, including agentic coding workflows that require multi-step reasoning and tool use.

I asked GPT-6 Astra in Codex to build me an explorable 3D Mysuru Palace. Here's what it created with Blender & Three.js ✨ I separately had it compose an audio track, then mashed it all together Loved bringing this idea to life!

World Labs co-founders Justin Johnson and Dr. Fei-Fei Li say LLMs use next-token prediction, but spatial intelligence has its own equivalent: Justin: "The soft definition of AI-completeness is there's this fundamental primitive that's an AI task. But if I could solve this AI task in its full, broadest generality, it would solve any intelligence problem." "The classic example in LLMs is that next-token prediction is AI-complete... I think from Ilya: there's a mystery novel, the thing has to read the whole novel, and the final sentence is, 'And the killer was.' Predict the next token. You could basically frame any kind of intelligence task in terms of that." "So clearly next-token prediction is something people believe is AI-complete." "New-view prediction, this primitive that we have in Atlas, especially generative new-view prediction, is also AI-complete." "I want to have a world where Martin is writing a proof of the Riemann hypothesis on the blackboard, and then the camera pans over to the next whiteboard." Fei-Fei: "Evolution had to solve new-viewpoint prediction by making animals move. Nature gave animals eyes, but nature didn't give trees eyes. Why? Because when you move, you see a new viewpoint... We do believe very strongly that next-viewpoint prediction is the equivalent of next-token prediction." @jcjohnss @drfeifei @martin_casado

As a test of our progress to advance the frontier of AI research, in June we entered the next generation of our autonomous AI research system, AIRA₃, in a live Kaggle competition run by NVIDIA to fine-tune a 30B Nemotron model. The challenge was to teach the model to reason

Rather than relying on a central controller, AIRA₃ runs many long-running agents (pairs of models + coding harnesses) in their own isolated environments and coordinates asynchronously through two shared substrates: 1️⃣ a forum for sharing hypotheses and findings 2️⃣ a shared

Highly recommended. It's not obvious, but a bottleneck in even the most powerful models, like GPT-6 Astra, is context bloat and compaction. This new config can help keep Astra persistent across long-running tasks. If it helps, here is a little visual summary courtesy of GPT-6

Harrison Chase

@hwchase17

2일 전

This is very similar to how we think about context management in Deep Agents: compaction should reduce what is sent to the model, not erase the underlying work history. In Deep Agents, the logic is: - every agent gets a filesystem to work with (can be real or virtual) - when a tool result is too large, we write the full result to the filesystem (/large_tool_results) and replace the model-visible message with a preview - when the conversation gets long, we offload older messages to the filesystem (/conversation_history) history first, then generate a summary - the agent can also decide to compact proactively via a compact_conversation tool, which uses the same offload + summarize logic Summarization middleware with all this logic: https:// github.com/langchain-ai/d eepagents/blob/main/libs/deepagents/deepagents/middleware/summarization.py …

Banger paper from Microsoft and Cornell. If you have looked at thinking tokens and decided you cannot afford the context, read this one. (bookmark it) A pause token buys the model extra compute for each next-token prediction, and it pays for that compute with a sequence position. Free pause tokens carry the same compute in a parallel prediction stream over a weight-shared backbone, riding an existing position instead of adding one. At inference it adds nothing to context length, leaves the KV cache unchanged, and costs essentially no extra latency, since the additional flops are not the throughput bottleneck. On a 1B model it improves next-token prediction by 2 to 3 centinats. The cost moves to training, where the overhead against an optimized pretraining pipeline comes down to about 1.14x while keeping most of the benefit. Paper: https:// arxiv.org/abs/2609.03807 Chat with Paper: https:// academy.dair.ai/papers/free-pa use-tokens-2609.03807 …

“Rethinking On-Policy Distillation of Large Language Models II” On-policy distillation looks extremely data hungry, but this paper shows the opposite. The paper shows that a single query can recover 72% of full-data OPD’s gain because its rollouts already visit 71.5% of the relevant state space. So the real bottleneck is not data coverage but how slowly the student absorbs the teacher’s dense token-level supervision. https:// alphaxiv.org/abs/2609.04172

Nayib Bukele

@nayibbukele

2일 전

“Results in reading, mathematics and science were above the national average and comparable to the average results in Germany and Sweden.” These are the results of our pilot program, which includes AI tutors and has been running in 171 public schools for just over a year. It was evaluated by PISA for Schools in June of this year. Since then, it has already expanded to more than 1,000 schools, and within 18 months, it will be in every public school in our country. In a few years, El Salvador will inspire the world. Security was just the beginning 🇸🇻

Ever wondered why type systems don't automatically track tensor shapes across your neural network modules? In his upcoming talk at PyTorch Conference North America 2026 (October 20–21 in San Jose, CA), Avik Chaudhuri from @Meta breaks down how PyreFly - a static type checker for Python - enables end-to-end static shape coverage for PyTorch models. Register for PyTorchCon North America today: https:// bit.ly/4qStJvL #PyTorchCon @__avik

Ever wondered why type systems don't automatically track tensor shapes across your neural network modules? In his upcoming talk at PyTorch Conference North America 2026 (October 20–21 in San Jose, CA), Avik Chaudhuri from @Meta breaks down how PowerFly - a static type checker for Python - enables end-to-end static shape coverage for PyTorch models. Register for PyTorchCon North America today: https:// bit.ly/4qStJvL #PyTorchCon @__avik