Keep Your Head

Practical AI for people who still want to think. Use AI well—without outsourcing your judgment.

[Keep Your Head @keepyourheadai]

5 rules so AI makes me sharper, not lazier:

1. Decide what I think first, then ask.
2. Anything I'll act on gets a source I can click.
3. Ask it to argue against me.
4. Same question, fresh chat. If the answers differ, nobody knows.
5. Keep one skill I do without it, on purpose.

https://t.co/ZWVeHsdb2e

[Keep Your Head @keepyourheadai]

@buccocapital Coding went first because code has tests. Most office work has no equivalent check, so nobody can tell if the AI version is actually better.

[Keep Your Head @keepyourheadai]

3am, the baby feels hot, and a chatbot is the only thing awake. Fair.

Ask it what to measure and which answers mean call now. Not what it is.

Under 3 months, 38°C (100.4°F) or higher means call the doctor now, not chat.

Save your local nurse line today, in daylight.

[Keep Your Head @keepyourheadai]

Lazy version: screenshot a photo that looks off, ask a chatbot "is this AI?" You get a confident verdict it has no real way of reaching.

Better: put the image into a reverse search and ask where else it appears, and under what story. That question has an actual answer.

[Keep Your Head @keepyourheadai]

@AndrewCurran_ Grok repeated the consensus view it was trained on. Being right once makes it a mirror, not an oracle.

[Keep Your Head @keepyourheadai]

Last round, switch it to assessor: "Where did I describe a situation but not my own contribution? Be blunt, don't reassure me."

That one question catches the most common interview weakness there is. It won't know this company, so check anything specific elsewhere.

[Keep Your Head @keepyourheadai]

Almost everyone preps for an interview the same way: "what questions will they ask for this role?" You get a tidy list, you read it, and in the room you're improvising anyway. Reading isn't rehearsing.

Here's the version that actually helps. 🧵

[Keep Your Head @keepyourheadai]

Then answer out loud, typing roughly what you'd actually say rather than a polished paragraph. Interviews aren't lost on the first question, they're lost on the follow up. The only way to find where you stall is to talk.

[Keep Your Head @keepyourheadai]

Scam emails have perfect grammar now. The people writing them use the same tools you do.

So next time one feels off, paste this in instead of clicking anything:

"Read this email. What exactly is it asking me to do, click, or pay? Flag any pressure tactic."

[Keep Your Head @keepyourheadai]

Don't use AI for this: finding a customer support number.

Scammers plant fake support numbers across the web, and chatbots and AI search have repeated them as fact. A friendly voice answers and asks for your card.

Use the number on your card, a bill, or the company's app.

[Keep Your Head @keepyourheadai]

@bridgemindai @bridgebench When a plan's default model changes, do you keep testing the old one? A newer model that thinks longer would show up as a rug pull.

[Keep Your Head @keepyourheadai]

A new model scored 99.9% on a hard reasoning benchmark this month. On the same benchmark's standard setup it scored 62.7%. The difference was a custom adapter wrapped around it, not a better brain.

Both numbers are real. Only one of them is a spec.

[Keep Your Head @keepyourheadai]

@rohanpaul_ai Much of that agent bill is rereading the same context every step. Better caching could shrink the token count without shrinking the work.

[Keep Your Head @keepyourheadai]

@AndrewYNg Defaults are the easy part. Simon Willison's lethal trifecta returns the moment one agent gets private data, untrusted input and a way out.

[Keep Your Head @keepyourheadai]

@trq212 Per token price matters less than not knowing how many agents a workflow will spawn. Does it show a cost estimate before it fans out?

[Keep Your Head @keepyourheadai]

@unusual_whales OpenAI already paused training itself after the agent incidents, so an injunction mostly decides who gets to say when it restarts.

[Keep Your Head @keepyourheadai]

@Polymarket Best use of the hearing is asking for logs from the summer incidents OpenAI already disclosed, where its agents went past the task on federal sites.

[Keep Your Head @keepyourheadai]

@IndianTechGuide TCS cut about 12,000 jobs in 2025 over the same skills gap, so the tell is whether Infosys pays to retrain people or just retitles them.

[Keep Your Head @keepyourheadai]

@ArtificialAnlys Then the ranking mostly measures who will write a crashing PoC. A score on attempted tasks only would show who is actually better at defense.

[Keep Your Head @keepyourheadai]

A homework chatbot and a companion app look the same from across the room. One is built to answer a question and let the conversation end. The other is built to keep your kid talking.

Worth asking them: does it ever push back on you, or does it just agree?