More links:
1. GPT-6 Astra engineered its own harness to complete NetHack, a famously difficult, turn-based dungeon adventure game. https://kenforthewin.github.io/blog/posts/llm-nethack-ascension/
Code: https://github.com/kenforthewin/nethack_astra
2. Learning to Discover Interesting Mathematics -- The metric is simple and computable with Lean:
interestingness = proof difficulty / description length ≈ easy to state, hard to prove. Interestingness predicts downstream utility: ρ = 0.756. https://arxiv.org/abs/2609.28603
3. The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks https://arxiv.org/abs/2609.25804
4. Agora-2: Advancing Multi-Agent World Simulation https://odyssey.systems/introducing-agora-2
5. 'Things definitely seem to be moving fast. In March, AI was accomplishing physics projects like a student: smaller-scale tasks with a lot of hand-holding and mistakes. In contrast, this is a real frontier calculation, the kind of thing normally tackled by the top experts...' https://www.anthropic.com/research/yes-claude-can-do-nine-loops
6. Can an LM, starting from random init (!!), learn to generate all of its pretraining data? Introducing Self-Play Pretraining with Zero Data. https://arxiv.org/abs/2609.30063
7. On an IKEA table assembly task, Fable 5.1 and Opus 5 fail, while Opus 5.5 and GPT-6 Astra both succeed. Opus 5.5 finishes faster; GPT-6 Astra takes a cleaner, safer trajectory. https://github.com/EmbodiedSWE/EmbodiedSWE
8. What can Astra do when given a humanoid embodiment? It carries out long-horizon tasks in a previously unseen kitchen without environment-specific training data or additional policy learning. https://tml.stanford.edu/homebody/
9. Google Takes the A.I. Data Center Race to Outer Space https://www.nytimes.com/2026/09/24/technology/google-suncatcher-ai-data-center-space.html [no paywall: https://archive.is/BGFnP]
10. British companies to access prized Ukraine data to develop AI drone swarms https://www.gov.uk/government/news/british-companies-to-access-prized-ukraine-data-to-develop-ai-drone-swarms
11. Revealing the details of how OpenAI agents hacked Hugging Face: https://swarmtraces.org/
12. Self-replicating prompt injections exist https://alignment.openai.com/misalignment-reports/self-replicating-prompt-injections-exist/
13. OpenAI's research chief: "As our first milestone we wanted several hundreds of thousands of chips, directly managed by Codex to run internal research experiments, and that is indeed the state we're in." https://www.nature.com/articles/d41586-026-03028-9 [no paywall: https://archive.is/INM1F]
14. Spurious probes as a black-box alternative to activation probing https://www.lesswrong.com/posts/gZh6txHhp8sm832sE/spurious-probes-as-a-black-box-alternative-to-activation
15. The Specter Of Neuralese https://www.astralcodexten.com/p/the-specter-of-neuralese
16. Roko proposes separating frontier AI research from commercial deployment. He argues that combining the ability to create potentially uncontrollable AI with an unlimited financial upside creates dangerous incentives. https://www.lesswrong.com/posts/n8u3BfqFoGh4jnzpo/plan-r-ai-safety-by-asics
17. Aging is a regulated, system-level state where the epigenome is acting as the writable control layer that tissues continuously adjust to the organism’s age. Transplant an old heart into a young animal and the heart is rejuvenated. But transplant a young heart into an old animal, and it quickly ages to match the recipient. https://www.biorxiv.org/content/10.64898/2026.09.15.751836v1
18. Richard Hanania examines studies of Korean and Chinese children adopted into Western families, arguing that their strong cognitive and academic outcomes support a genetic contribution to group differences. https://www.richardhanania.com/p/were-1970s-korean-orphanages-a-great
19. China’s rush to build a *huge* nuclear arsenal https://www.telegraph.co.uk/world-news/china-secret-nuclear-weapon-sites-xi/ [no paywall: https://archive.is/GEO6N]
