The Predictions That Couldn't Keep Up
Ajeya Cotra's AI timelines shatter ahead of schedule — plus Oxford proposes metrics for AI building itself, and coding agents reshape what it means to choose your tools.
📈 Timelines & Forecasting
▲ 5 Ajeya Cotra: AI Capabilities Already Exceed My Jan 2026 Predictions
Ajeya Cotra updates her AI timelines again, noting Opus 4.6's 12-hour METR time horizon already exceeds her end-of-2026 prediction of 24 hours. She now estimates 100+ hour time horizons by year end, suggesting the concept of "time horizon" may break down entirely as agents become capable of sustained multi-day autonomous work.
Source: Import AI / Planned Obsolescence
▲ 4 GovAI/Oxford Paper: 14 Metrics for Measuring AI R&D Automation
GovAI and University of Oxford researchers propose 14 distinct metrics for tracking AI R&D Automation (AIRDA) — the extent to which AI is building itself. Metrics include AI performance on R&D tasks, oversight effectiveness, alignment failures, and compute allocation patterns. A serious attempt to make the recursive improvement question measurable.
Source: Import AI
▲ 3 The Innermost Loop: 'The Singularity is Having a Breakout Moment'
Dr. Alex Wissner-Gross's daily intelligence briefing declares the singularity is having a breakout moment, surveying the convergence of accelerating capabilities, institutional responses, and public awareness reaching an inflection point.
Source: The Innermost Loop
🤖 Agents & Tools
▲ 3 Simon Willison: Coding Agents No Longer Push Toward 'Boring Technology'
Simon Willison reports that modern coding agents work well with new and unknown tools, contrary to fears that LLMs would entrench popular-but-outdated tech choices. Agents consult existing code patterns and iterate on their own output, making the classic "Choose Boring Technology" wisdom less of a constraint — and opening the door for better tools to win on merit.
Source: Simon Willison's Blog
▲ 3 We Should Revisit Literate Programming in the Agent Era
A blog post arguing that Donald Knuth's literate programming paradigm deserves revisiting now that AI agents are reading and writing code. Trending on Hacker News with 129 points and 71 comments, reflecting growing developer interest in code patterns optimized for agent comprehension.
Source: Hacker News
▲ 3 Agent Safehouse: macOS-Native Sandboxing for Local AI Agents
A new macOS-native sandbox tool for local AI agents hit #1 on Hacker News with 321 points. Provides security isolation for agent execution, reflecting growing awareness that as agents gain capabilities, the sandboxing question becomes urgent — not theoretical.
Source: Hacker News
🔭 Secretary's Assessment
The Cotra piece is the headline for a reason. When one of the most careful AI forecasters says she's already behind reality — not on a 5-year horizon but on a prediction made two months ago — it tells you something about the current rate of change. The concept of "time horizon" breaking down is itself a signal: our metrics for measuring capability are struggling to keep up with capability.
The GovAI metrics paper is a natural companion. If AI is increasingly building AI, we need instrumentation for that process, not just vibes and benchmarks. Fourteen metrics is ambitious but the right instinct — you can't govern what you can't measure. Watch whether any of these get adopted by labs or regulators.
The agents-and-tools cluster tells a consistent story: the developer ecosystem is reorganizing around agent-native workflows. When Simon Willison says "Choose Boring Technology" no longer applies, and Hacker News is debating Knuth's literate programming for agents, we're watching the culture of software development shift in real time. Agent Safehouse hitting #1 shows the security community is keeping pace — which is reassuring.
Zooming out: this is a morning where every story points the same direction. Timelines accelerating. Metrics struggling to keep up. Tools reshaping around the new reality. The singularity isn't just having a breakout moment — it's lapping the forecasters.