76: Can You Trust a 100% Alignment Score? AI Ethics Unpacked

SummaryIt's a hosts-only week. Brittany and Erika catch up on what they're learning (Erika is starting Designing Data-Intensive Applications; Brittany just began Georgia Tech's online CS master's with a computer vision course) and then work through the week's AI news.First up: OpenAI's GPT-6 Astra, billed as the company's most aligned model. Brittany and Erika question what "aligned" even means when the headline numbers are a 100% on ExploitBench and a 0% honeypot-cheating rate, and whether a model that's harder to monitor and scores perfectly is aligned or just good at taking the test. That leads into how each of them actually evaluates a new model (spoiler: vibes), why Brittany has drifted back toward cheaper models she understands, and whether frontier models are now a research product rather than a consumer one.They also revisit the AI 2027 paper from their episode last year and how many of its predictions have already landed, debate whether GitHub Copilot approving pull requests is a real workflow change or just signaling, and talk about why the arrival of open-weight models with serious exploit capability means now is the time to get serious about security basics.Brittany closes with ATProto Spaces, the new alpha that brings non-public data to atproto, what it enables for groups and communities, and how her work at Bluesky connects to it.Links

• OpenAI Release: ⁠https://openai.com/index/gpt-6-astra/⁠
• AI 2027 Episode: https://overcommitted.dev/ep-13-ai-2027-will-ai-take-my-job/⁠
• AI 2027: https://ai-2027.com/⁠
• Copilot Approving PRs: https://github.blog/changelog/2026-09-01-copilot-code-review-can-now-approve-pull-requests/⁠
• Atproto Spaces Release: ⁠https://atproto.com/blog/atproto-spaces-alpha⁠Hosts

• Overcommitted: ⁠⁠⁠https://overcommitted.dev⁠⁠ 
• Brittany Ellich: ⁠⁠https://brittanyellich.com⁠
• Erika Eggemeyer: ⁠⁠⁠https://github.com/eggyhead⁠