AI Daily — September 12, 2026
2026-09-12
openai.com
Perplexity trusts GPT-6 Astra with end-to-end production systems
Perplexity has deployed GPT-6 Astra as an autonomous agent capable of writing communications, modifying software, and monitoring live production systems with minimal human oversight. This represents a meaningful step beyond prior agentic deployments, where human check-ins were far more frequent. The case study signals that AI agents are now being trusted with end-to-end operational responsibility in production environments at scale.
openai.com
Cognition's Devin uses GPT-6 Astra to self-test software
Cognition has integrated GPT-6 Astra into Devin to handle software testing and verification, with the goal of reducing the volume of code engineers must manually review before shipping. The model's ability to reason about test coverage and correctness is being used to close the loop on agentic code generation. Combined with the Perplexity deployment, this points to GPT-6 Astra becoming a backbone for multi-step software engineering agents.
openai.com
OpenAI scales ChatGPT storage to 1 billion users and 22M requests/sec
OpenAI details how its internal storage system Habitat evolved from a Python library into a globally distributed platform handling 22 million requests per second for over 1 billion ChatGPT users. The post covers architectural decisions around consistency, sharding, and global replication required at this scale. This is a notable engineering reference for teams building distributed systems under extreme read/write loads.
www.technologyreview.com
Powering AI is an architecture problem: grid failures at Ashburn data center cluster
A July 2026 transmission line fault in Ashburn, Virginia knocked over 3 gigawatts of data center load off the grid in seconds, the latest in a series of large-scale power failures affecting the world's densest AI compute cluster. The piece argues that the core bottleneck for AI infrastructure is no longer chip supply but electrical grid architecture and the lack of resilient power delivery at scale. This is increasingly a hard constraint on frontier model training and inference expansion.
rss.arxiv.org
Quantifying grokking onset: scaling laws for the memorization-to-generalization transition
Researchers map grokking boundaries across 384 hyperparameter configurations on modular arithmetic tasks, fitting a power-law relation for generalization onset time: T_grok ∝ H^−0.27 · D^−2.04 · η^−0.50 · λ^−0.64, with R²=0.82 including interaction terms. The dominant factor is data complexity (exponent −2.04), not model capacity, overturning assumptions about how to accelerate generalization. This provides the first quantitative phase diagram for grokking and has practical implications for training dynamics research.