Building software and teams with agents and humans. Founder of Connsulting, co-host of @managerxbuilder
Washington, DC
Joined May 2013
- Tweets650
- Following32
- Followers46
- Likes77
Brian Conn retweeted
Stock DeepSeek refused to write the phishing email. Brian switched to the abliterated twin of the same model and it wrote the lure, reset link included.
managerxbuilder.substack.com…
Brian Conn retweeted
Klaus thinks dedicated tech debt weeks should be gone. Kunal's not sure, because you're shipping a lot more than you used to.
Brian Conn retweeted
An abliterated model has the refusal weights cut out. Brian ran one on an isolated home lab next to stock DeepSeek. Same model family. One writes the phishing email, the other won't.
Brian Conn retweeted
Brian deleted all of his custom sub-agents this week. He still runs vanilla sub-agents constantly, because a test writer shouldn't know what the implementer did.
managerxbuilder.substack.com…
Brian Conn retweeted
Brian's personal wiki ranks sources by trust. His own words count most, then his calls, then the podcasts and newsletters he trusts. Random web results never get in.
Brian Conn retweeted
Tomorrow on Manager Builder, Kunal and Brian debate generalized vs specialized agents.
"Most of our best practices are just best practices for humans."
Brian Conn retweeted
In planning, Claude should ask what you want to do and then stop. Once it hands you a recommended option, it's too easy to just hit Enter.
managerxbuilder.substack.com…
Brian Conn retweeted
One-shot vs progressive on the same monitoring stream.
One-shot: four false pages, missed the leak, a couple of cents.
Progressive: one page, no false positives, about 50 cents.
managerxbuilder.substack.com…
Brian Conn retweeted
Asked a fast classifier to make one call per event. Four false alarms, and it missed the memory leak.
Brian Conn retweeted
AI slop can cost you a client.
Manager: keep a quality gate when models write the code.
managerxbuilder.substack.com…
Brian Conn retweeted
Why personal software matters: remove friction in your own workflow, and get a low-risk lab to try AI before work.
Builder.
managerxbuilder.substack.com…
Brian Conn retweeted
Squarespace had no API, so I was the human in the loop adding zero value.
One day-long side project with AI later, the site runs on Astro 5 + Cloudflare Pages.
Builder take on personalized software.
Brian Conn retweeted
Are you leading Claude, or is it leading you?
Manager take: lead the bots instead of being led by them.
managerxbuilder.substack.com…
Brian Conn retweeted
Cool eval. Simply ask an LLM “Land or Water?” and give it a latitude and longitude coordinate as text. Ask 16,200 times, plot as image. The models know. From compressing the internet.
Brian Conn retweeted
A human avatar sending dehumanized Claude-speak in Slack isn't helpful — it feels disrespectful.
Manager take on AI slop in human channels.
Brian Conn retweeted
Functional constraints vs architectural ones — which actually breaks your agent workflow?
Brian Conn retweeted
What does software development look like when agents write code all day—and much of it gets merged overnight? Developers become team leads: define intent, then hand work to a “night shift.”
managerxbuilder.substack.com…
Brian Conn retweeted
Here's a 45-second TL;DR on Jev.
I find the core idea beautifully simple, but the video made it really hard to understand.
Hope you find it helpful.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Brian Conn retweeted
Here's what OpenAI's agentic software factory looks like, today.
Details: newsletter.pragmaticengineer…
(thanks to all the OpenAI folks who explained how it works! And Perf Factory looks especially interesting to me)
Brian Conn retweeted
"Man, you must have burned thousands of dollars in tokens to generate that."
Fishslop would have been under $30 to generate. Wouldn't even have noticed the dent on the $200 sub.
