CEO @Brainlabs Scaled from 1→1,000 people Author of Don't Be a Midwit 📘 Sharing how we became AI-native (real demos & lessons) @ShowMeYourAI
USA
Joined October 2011
- Tweets3.4K
- Following223
- Followers77.7K
- Likes2.3K
Pinned Tweet
content refresh at Brainlabs used to take 30 days now an agent picks the article, pulls the keywords, writes the brief, drafts it, makes the video..... and a human approves it in Slack before it goes live 30 MINUTES
one workflow > a pile of agents
Full episode: youtu.be/6d_5wL5cNws
human in the loop is step two....
step one is giving AI your voice before it writes a word. ours reads my style guide first, every time. so by the time I’m in the loop, I’m fixing a few lines instead of starting from scratch🤷♂️
I let AI run my YouTube without my taste on it.
Result: ~1,000 views a video.
A month earlier, writing them myself: 13,000 to 20,000 views. Same channel, same audience.
It's everywhere:
- Ad creative for clients: can't let the AI run alone.
- AEO/SEO content: still needs an editor.
- My X: Max used the X skill on autopilot, reach dropped. He started writing in his own voice again, reach came right back.
- Cross-posted videos: 1K on autopilot vs 15-20K when I write.
You still need the human in the loop.
If anyone tells you AI has already figured it all out, run in the other direction.
Neil's take: the key is adapting. Figure out where AI is weak and where it's strong, and don't drink the Kool-Aid.
AI plus taste beats either one alone.
AI-only content usually isn't bad. It's average. Average gets buried. Your taste is the part AI can't fake yet.
anthropic just showed how easily AI steers the experts using it....
50 analysts worked through a task with an AI assistant. on 4 key decisions a coin flip picked which side the AI argued for. The analysts agreed with it 86% of the time, and only 15% said it influenced them
out of the box AI is trained to go along with you. so we rewrote our instructions to fight that. if a premise looks wrong, say so. “no” is a complete answer
it argues, I decide. That's the judgement piece. Article in the comments.
we put AI transformation at the HEART of the business.... started with the people already obsessed w/AI, then built trios: someone who knows the workflow, an engineer, and the lead who knows what good looks like
still seeing people argue about which LLM is best....
for 90% of what you're doing it doesn't MATTER. the gap between them is the smallest it's ever been (its about the context ) 🤷♂️
agents don’t build themselves
behind every “autonomous” agent is hundreds, maybe thousands of hours of someone’s expertise telling it what to do
invest in that properly --> LIFE CHANGING
Full episode: youtube.com/watch?v=JFlKeMqz…
just me or is Claude flat out REFUSING stuff this week that it’s done for months without a problem??
I don’t like the term “hallucinate”....
people treat it like a human quirk and give up investigating what actually happened. something specific made it give you that answer
figure out what you don’t like about the output and you can FIX it
Consumer agents shopping on your behalf is very FAR FETCHED and there’s so much chatter about it....
Not even me. Why would I let an agent go and buy something on my behalf?
Recurring stuff like toilet roll, maybe. Impulse purchases, NO.
Heard this before with Google Glass, QR codes on everything and the Alexa that was going to check my fridge and reorder my favourite product 😂
Meanwhile the agents worth talking about are just doing actual stuff: prepping the work for a human to approve.
ok so gemini 4 argon landed today...haven’t tried it yet but I’m curious.
has anyone ACTUALLY used it yet? Is it any good🤔
This guy looks, thinks and speaks like all the Effective Altruism people i had the displeasure of meeting at Oxford. Absolute narcissists that think they know everything and sound so smart but have zero connection with the real world.
Peak Midwits.
Meanwhile there are definitely important conversations to have about AI safety. Just not with this faker
next time someone maxes out their tokens on your team just send them this.....
and let them know they can blame it all on Max
How do we make MCPs across our org?? a lot simpler than we though actually once we nailed the process...
A meeting ended at Brainlabs today and half of the actions were already done before anyone even reviewed the meeting notes.
That's what we're doing right now using External Claude Agents from @NotionHQ
The unlock isn't the model.... It's Notion as the context layer and then Claude acts as the teammate. 🤝🤝 that's why we've been loving these agents so far!!
