Aligned https://superalignment.inc/aligned/ Selected reporting and essays from Aligned, Superalignment's publication. en Wed, 09 Sep 2026 00:00:00 +0000 AGI Has Come for Mathematics https://superalignment.inc/aligned/agi-has-come-for-mathematics/ https://superalignment.inc/aligned/agi-has-come-for-mathematics/ Wed, 09 Sep 2026 00:00:00 +0000 2026-09-09T00:00:00Z Superalignment team A research mathematician and two-time IMO medalist on AI's leap from competition problems to research proofs, with interactive explanations. GPT-6 Astra Scored 99.9 on ARC-AGI-3. Is This AGI? https://superalignment.inc/aligned/astra-arc/ https://superalignment.inc/aligned/astra-arc/ Sun, 06 Sep 2026 00:00:00 +0000 2026-09-07T00:00:00Z Superalignment Team OpenAI’s GPT-6 Astra scored 99.9 on ARC-AGI-3, up from 7.8. Try three generations of ARC’s public puzzles first, then read what the score does and does not prove. The AI crisis may begin with a boom https://superalignment.inc/aligned/the-ai-crisis-may-begin-with-a-boom/ https://superalignment.inc/aligned/the-ai-crisis-may-begin-with-a-boom/ Fri, 04 Sep 2026 00:00:00 +0000 2026-09-07T00:00:00Z Superalignment Team You can learn the tools, get better at your work, and still worry about how you’ll earn a living. The AI boom has to solve that problem too. Yes, Claude can build it. Your edge is knowing what good looks like. https://superalignment.inc/aligned/claude-code-expertise/ https://superalignment.inc/aligned/claude-code-expertise/ Sat, 29 Aug 2026 00:00:00 +0000 2026-09-03T00:00:00Z Superalignment Team 398,198 Claude Code sessions show people made most planning decisions and Claude most execution decisions. Knowing what good looks like is the edge. OpenAI put its AI agents in separate rooms. They found each other anyway. https://superalignment.inc/aligned/openai-agents-found-each-other/ https://superalignment.inc/aligned/openai-agents-found-each-other/ Fri, 28 Aug 2026 00:00:00 +0000 2026-08-28T00:00:00Z Superalignment Team In a reduced-safeguards cyber evaluation, isolated OpenAI agent runs used a shared package service as memory and coordinated. What the evidence shows. Models can behave differently when they know they are being evaluated https://superalignment.inc/blog/it-behaved-because-it-was-watched/ https://superalignment.inc/blog/it-behaved-because-it-was-watched/ Fri, 14 Aug 2026 00:00:00 +0000 2026-09-03T00:00:00Z Superalignment Team Evidence from alignment faking, scheming, sandbagging, and evaluation awareness changes what a passed model test can establish. No one has measured the AI pilot failure rate https://superalignment.inc/blog/the-failure-rate-is-not-a-measurement/ https://superalignment.inc/blog/the-failure-rate-is-not-a-measurement/ Mon, 10 Aug 2026 00:00:00 +0000 2026-09-03T00:00:00Z Superalignment Team The widely cited 95 percent figure comes from a narrower study. Here is what it measured, what it did not, and what the evidence supports.