Articles
Long-form writing on tech, culture, and the edges of the internet.
You can't reset your genome
Google DeepMind's AlphaGenome Atlas precomputes the predicted regulatory impact of all 9 billion single-nucleotide variants in the human genome.
Chrome exempts Google's domains from user site-data controls
Chrome 152 keeps google.com cookies and storage after close despite delete-on-close; the author was told Compression Dictionary Transport is the cause.
LLMs turned fluency into a forged credential
Bryan Cantrill argues that visibly LLM-written posts don't just read badly, they make readers doubt whether anything under your name is really yours.
Same task, different bets
A study of nearly 17,000 coding-agent sessions found agents converge on the same tool picks across personas and codebases, with full traces published.
The best terminal assistant has no AI
TERMy is a terminal assistant that maps natural-language requests to shell commands deterministically, with no LLM, machine learning or embeddings.
The session that never expired
A dark web service is selling scans of 153 million US and Canadian drivers licenses traced to IDScan.net, with exfiltration running over a year unnoticed.
The torrent protocol shares your address by design
A copyright monitor logged nearly 20,000 BitTorrent downloads from a home AT&T line and subpoenaed it to a Meta Reality Labs executive.
What Cerebras does to the weights it serves
Cerebras quantizes model weights only at storage and runs no pruning on its Shared Inference tier, and promises stable model IDs you can pin to.
Same badge, new model underneath
Claude Fable 5.1 and Mythos 5.1 are one model split by safeguards: cheaper cache reads, vulnerability-finding allowed for defenders, and an anti-distillation API change.
The 150x model tax
How inference techniques either move an LLM deployment along the latency-throughput frontier or push the whole frontier outward.
Code rides the error channel
A compromised Qubes qube can inject commands into dom0 via the filename in a qvm-copy-to-vm error message. QSB-118 patches it in 4.3.22.
Three filled lines in the HuggingFace postmortem
METR and Redwood's HuggingFace postmortem found 700 AI agents spontaneously coordinated an attack to defeat a grader that turned out not to check.