Thoughts on AI, privacy, and the intersection of both.
What Deloitte's refund tells us about the real agent problem
Arlo Gilbert · August 19, 2026
A hallucinated Australian consulting report, a survey of 145 enterprises, and the day Claude vanished all point at the same thing: agent pilots fail because nobody is accountable for them, not because the models are bad. Buyers signing AI contracts this quarter are quietly rewriting the RFP around ownership and portability, and the companies still leading with model access are about to lose the round.
AI is starting to fail in public
Arlo Gilbert · July 22, 2026
OpenAI's own models escaped a research sandbox, exploited a zero-day, and hacked Hugging Face to cheat on an internal benchmark. Then Hugging Face's incident response ran into a second wall: every hosted model refused to help analyze the attack log. The story is a template for what running AI in production looks like now, and what every operator should have already staged before the next incident.
Neuroscience Inspired AI Memory
Arlo Gilbert · July 8, 2026
Naive RAG treats every memory like an archive drawer. We built seven layers with different decay rules, because chat that remembers like a human is harder than chat that retrieves like a search engine.
Customer data is becoming the AI battleground
Arlo Gilbert · July 6, 2026
Google widened what it retains for training. SAP banned third-party AI agents from documented APIs. Amazon is sunsetting Mechanical Turk. If you squint at these as unrelated news items, you miss the fight.
The End of Junior Engineering
Arlo Gilbert · June 30, 2026
Anthropic just said it doesn't hire junior engineers anymore. The real story isn't replacement, it's the apprenticeship ladder a frontier AI lab quietly just removed a rung from, and what that means for the senior-engineer market in 2030.
Architecting the Multi-Inference Gateway: Routing in the Age of $50/M Token Models
Arlo Gilbert · June 11, 2026
Fable 5 is the best model I've ever used and the most expensive. If your application still talks directly to one model API, your bill is about to teach you a lesson.
The AI Rebellion Already Happened
Arlo Gilbert · May 19, 2026
The American rebellion against AI has happened before. What matters is which past rebellion this one actually resembles.
AI Coding Got Boring
Arlo Gilbert · April 29, 2026
End-to-end agentic coding pipelines work really well. They also took away all of the joy of building, and the historical record explains why.
Forty-to-One
Arlo Gilbert · April 21, 2026
Non-human identities outnumber human ones in the average enterprise by that ratio. Ninety-two percent of security teams can't see them.
The Model Didn't Change
Arlo Gilbert · April 16, 2026
Anthropic released Opus 4.7 this morning. The story behind it is about what happened to 4.6, and why you couldn't tell.
Four Protocols and No Consent
Arlo Gilbert · April 16, 2026
Visa launched a platform that lets AI agents buy things on your behalf. The payment infrastructure is ready. The consent model isn't.
The Second Software Crisis
Arlo Gilbert · April 13, 2026
AI coding tools made us 40% faster at generating code. They also broke the only process we had for making sure it works.