Exploring technology, data science, AI and society through articles and analysis.
Your RAG works great in testing and falls apart in production? I've been there. The problem is almost never the LLM — it's the retrieval pipeline. Here are the 4 steps I use to fix it for good.
Demis Hassabis stepped down from day-to-day leadership at Google DeepMind. I explain why this reshuffle looks a lot less like career growth and a lot more like a strategic surrender.
Every time Claude goes down, someone makes a joke on X. I think the joke is hiding a serious architecture decision: is it worth depending on an intelligence that occasionally leaves you stranded?
A study covering 47,000 companies shows the real difference between companies that 'use AI' and companies that are AI-Native. I explain why that gap is far bigger — and far scarier — than it sounds.
While Wall Street keeps celebrating the AI race, the real financial scare happened in Seoul. I break down why the KOSPI panic is the scariest mirror of what could come next.
Hackers stole $88 million from a cold wallet by exploiting a 2010 flaw. Claude Opus found the same vulnerability in 8 minutes. Cybersecurity's arms race now runs at LLM speed.
Meta named the problem: Behavioral State Decay. The instruction is there. The AI 'sees' it. But stops acting on it. The fix: a second AI agent — the 'shoulder tap' — that injects selective reminders. +8.3 points on Terminal-Bench. Without retraining the model.
Boris Cherny, Claude Code creator and Head at Anthropic, identified 5 archetypes replacing traditional roles: Prototyper, Builder, Sweeper, Grower, Maintainer. At Anthropic, everyone holds the same title. What differentiates is how they operate.
#1 on MTEB English by 5.09 points. 5 modalities in a single 3,072-dimensional vector. Matryoshka to truncate to 768 without retraining. Text, images, 128s video, 80s audio, 6-page PDFs. Google solved the problem that forced teams to maintain 3 models and 3 indexes.