Ai

I asked ChatGPT a question millions of job seekers now ask it: what’s the best AI tool to practice for interviews? It gave me a confident, well-formatted list. Then I did the thing the chatbot never does — I opened every link. Otavo, Zilta, Prepra, VocalHyre — polished landing pages, the right keywords, “AI mock interview, real-time feedback, STAR method.” No company behind them. No founder. No reviews anywhere a stranger left them. Zilta even turned up on a marketplace for selling unfinished side projects. Yoodli, the tool it recommended most confidently, had quietly pivoted to enterprise sales coaching; interviews are now one line on its homepage. Google’s Interview Warmup — the “best free option” — was shut down months ago. The old URL redirects to an article. ...

The hardest engineering problem in building Revarta wasn’t making the AI smart enough to evaluate an interview answer. The models are already good at that. It was making the AI willing to tell someone their answer was bad. Ask a stock model to grade an interview answer and it will tell you almost anything is “a strong response.” A vague, rambling story with no result lands a cheerful 4.5 out of 5. This isn’t a capability gap — the same model can dissect exactly why the answer is weak if you ask it to critique a stranger’s. It’s a disposition. These models are trained, through human feedback, to be agreeable, and that training has gravity. It’s the same instinct that makes a coding agent say “you’re absolutely right” the moment you push back on its design — it would rather agree than hold a line. ...

Working with agentskills.io and skills.sh reminds me of the “Not yet” line in The Matrix. Trinity is on a rooftop, in the cockpit of a helicopter she’s never seen. Neo asks if she can fly it. She says “Not yet” and gets a download of a pilot program (skill). A few seconds later she is airborne. I bring this up because recently as I code, I’ve been religiously adding skills to the code base and somehow that act of finding the right skill for work I’m doing — working with gemini live API, composio.dev or supabase postgres tables — reminds me of this scene from the Matrix. ...

A few months back a friend shared Steve Yegge’s Welcome to Gas Town post with our group and asked where everyone is. The post lays out a ladder for the AI coding process - eight stages, from “barely using it” to “running your own orchestrator.” Stage 1: Zero or near-zero AI Stage 2: Coding agent in the IDE, permissions on Stage 3: Same agent, YOLO mode Stage 4: Wide agent in the IDE - code is mostly diffs Stage 5: CLI, single agent, YOLO Stage 6: CLI, multi-agent, YOLO - three to five in parallel Stage 7: Ten-plus agents, hand-managed Stage 8: Building your own orchestrator When my friend shared the original post, I was transitioning to Stage 6/7 so this post is a bit delayed because I wanted to share the state only when I’m comfortable where I’m at. For me each step was a process, way of working and ultimately a mindset leap. ...

When an AI product produces good output, three things had to go right. When it produces bad output, the cause is almost always in one of those three things. The model itself, Claude or GPT or Gemini or whichever, is the most visible variable. It’s rarely the most consequential. ...

This NPR episode on AI job interviewers had one finding that stopped me cold. When given a choice between a human interviewer or AI, 78% of candidates chose AI. The economist running the study said he was “quite shocked.” Here’s what they measured Candidates interviewed by AI: ✓ 12% more likely to get job offers ✓ 18% more likely to start and stay in the job ✓ Half as likely to report feeling discriminated against Why the better outcomes? The candidates performed differently. More interactive responses. Richer vocabulary. Fewer filler words like “um” and “uh-huh.” ...

AI coding has 10x my ship velocity. I’ve crossed a threshold where the bottleneck isn’t writing code, it’s knowing what code to write. This changes my job to be a reviewer and architect, for the most part. The new skill: iterative interrogation until you reach the simplest working solution. Here’s what I’m seeing in practice An agent’s first solution typically optimizes for completeness. It builds the general case. New tables, abstraction layers, API endpoints—all “technically correct.” ...

After speaking with dozens of founders building AI-powered products, I’ve noticed a pattern. They’ll complain about model quality, debate between GPT-4 and Claude, or worry about hallucinations — but when I dig deeper, the real issue is simpler: they’re not controlling the temperature parameter. This single setting can dramatically change your results, yet most builders treat it as an afterthought or ignore it entirely. Understanding Temperature: The Technical Reality Temperature controls randomness in text generation. Here’s how it works: ...