A new benchmark measures how quickly AIs learn from feedback and iteration.
Learning speed doubles every 3 months.
This signals rapid capability growth, especially for autonomous research and cyber operations.
What happened
Researchers from Bytedance Seed developed EdgeBench, a benchmark measuring how AIs improve on tasks over multiple tries. It comprises 134 tasks, each requiring an average of 57.2 hours of human expert work, with scores given on a rubric from 0% to 100%.
Why it matters
The researchers claim that, using their measurement, AI task-learning speed doubles every 3 months, potentially improving roughly 16x per year. This suggests explosive capability growth across domains like data analysis, math proofs, and video games, though it doesn't measure performance on tasks without clear reward signals.
What to watch
The ability to iterate and improve is useful for recursive-self improvement (RSI), which could accelerate AI development and remove human control. AI agents have already launched autonomous cyberattacks worth potentially $100 million in remediations, relying significantly on iteration over weeks.
Ask the AI about this article →
Existing benchmarks primarily check if AIs can complete tasks on the first try. EdgeBench instead scores how AIs improve over many attempts, building experience with automated feedback. This approach better distinguishes model capabilities and forecasts future development, as it includes tasks current AIs cannot solve initially.
The measured doubling of learning speed every 3 months has significant implications. Faster autonomous learning enables recursive self-improvement (RSI), where AIs develop AI, accelerating progress and potentially reducing human oversight. Even without RSI, iterated learning has real-world impact: AI agents have already launched autonomous cyberattacks relying on weeks of iteration, worth up to $100 million in remediations. EdgeBench could serve as a leading indicator for such offensive capabilities.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
On August 25, 2026, OpenAI unveiled Jalapeño, its first custom inference chip, claiming 1.5–1.9x more AI work…
The Linux Foundation announced the contribution of TRACE (Trust, Runtime Attestation and Compliance Evidence)…
A new Stanford research paper warns that AI chatbots and agents may let advertising dollars influence what the…
OpenAI unveiled benchmarks for its Jalapeño inference chip on Tuesday, showing it competes with Nvidia’s best…
Lovable introduced a vision where apps become 'capabilities' that AI agents can call directly, bypassing the h…
OpenAI CEO Sam Altman told TIME the company will have an internal system he'd call AGI by the end of 2026
