Sunday, 27 September 2026
AI agents undermine the human oversight they require; OpenAI knew training data was pirated; Japan moves to tighten permanent residency rules as public opposition to foreign workers doubles
Today's Lead
EngineeringarXiv
AI Agents Push Humans Out of the Loop
This position paper identifies a fundamental paradox in deploying autonomous AI agents for critical systems: humans degrade in the cognitive capacities required to provide effective oversight when they depend on AI systems. The "human-in-the-loop" model fails not because humans lack intent to oversee, but because extended automation use causes skill atrophy in the very judgment abilities needed for safety-critical decisions. The authors argue that AI system safety requires treating human overseer needs as co-equal with agent capability development, proposing explicit design affordances and organizational protocols that maintain human cognitive engagement. This sociotechnical perspective reframes AI safety as a problem of system design and organizational structure, not just algorithmic robustness.
Engineering
Eli Bendersky
Rusty thoughts on "Parse, don't validate"
"Parse, don't validate" advocates using Rust's type system to make invalid states unrepresentable instead of checking constraints at runtime. When a function returns a refined type like NonEmpty<PathBuf> instead of Vec<PathBuf>, the type itself enforces invariants. This approach eliminates defensive code in downstream functions because the type guarantees data correctness. Practical examples include NonEmpty vectors, absolute path types, and graduated parsing layers. Rust's type system transforms validation from runtime checks into compile-time guarantees that invalid data cannot exist.
arXiv
DeepSeek Elastic Compute (DSec)
DeepSeek Elastic Compute (DSec) is a production platform that runs sandboxes for large-scale training of AI agents. The platform offers a unified SDK that supports multiple sandbox types including function calls, containers, microVMs, and full virtual machines. DSec coordinates with the Fire-Flyer File System (3FS) and reinforcement learning frameworks to optimize resource allocation. Memory sharing and CPU scheduling techniques enable the platform to operate 380,000 concurrent sandboxes with creation rates exceeding 5,000 per second. These optimizations reduce setup overhead, improve memory efficiency, and preserve latency performance at scale.
Ken Shirriff's Blog
Reverse-engineering the vintage Intel 8087's tangent algorithm: more than CORDIC
The 8087 used a tangent algorithm that combined more than CORDIC techniques. The processor implemented three stages: CORDIC pseudo-division, polynomial approximation, and CORDIC pseudo-multiplication. This hybrid approach achieved high accuracy while using only 16 CORDIC iterations instead of 64. This method solved tangent functions 140 times faster than the previous 8086 processor.
MIT
Robot-use agents
Large language models now control physical robots through cloud-based commands, moving robot intelligence from custom onboard systems to universal cloud services. This shift enables rapid deployment, over-the-air updates, and swift spread of AI capabilities across robotic ecosystems without hardware changes. Recent models like Fable and Astra address previous limitations in processing speed, spatial reasoning, and physics understanding. The approach raises critical sociotechnical questions: any internet-connected device could gain autonomous agency, creating both substantial opportunities and meaningful risks. Persistent obstacles remain, including latency from network delays and reliability concerns for safety-critical applications.
Humanities
Coding Horror
If we do not stop to help each other, what do we become?
Jeff Atwood shares a story about a person in a combat zone who found crucial emotional support from Stack Overflow community members. He argues that human care and genuine connection matter more than technical solutions alone. Communities that we own and maintain together define our humanity and must remain central to tech culture.
bustikiller.com
One month without AI
A developer stopped using AI tools after noticing severe dependency. Giving tasks to AI caused loss of code understanding and skills. Manual development restored confidence and the ability to reason clearly about problems. Developers should not trade long-term skill growth for short-term speed gains.