Parasocial was Cambridge Dictionary’s word of the year in 2025.
And with good reason. AI chatbots interact with us “as if” they were human.
With @LennartMeincke, @danshapiro, @emollick, Lilach M., Christophe Van den Bulte, @RobertCialdini, I published a new study in @pnas. We found classic human persuasion techniques, Cialdini's "Principles of Influence," led to AI chatbots such as ChatGPT agreeing to objectionable requests (to synthesize chemicals it shouldn’t), increasing compliance from 35% to 51%.
Open access here: https://t.co/WehkFuNI7I
1/
Which Dan level has your team reached? 🥋
A few months back, @danshapiro published a piece mapping AI-assisted development onto five levels of automation, and I've been thinking about it.
@mvanhorn@OmarShahine@openclaw@twilio Yes! Routing communication through the telephony stack opens up bluetooth audio, which is amazing.
If you want a cheatsheet, this is an entire non-claw agent + twilio stack, but you can tell you coding agent to just steal the twilio stuff.
https://t.co/rmNF5e1xGw
I used to think OpenClaw was an a clever app. Now, I think it represents the fourth evolution in a series: completion, chat, agent, claw.
https://t.co/sybQN2HDEl
I am noticing a really interesting trend:
my friends who are not technical are getting much further with open claw (etc), etc than my friends who are technical.
it seems that the ability for us to imagine the limits is artificially causing limits that are not there for folks who are not imagining limits.
have you seen the same thing?
Introducing the Printing Press, a CLI-factory and a CLI-library. Built with @trevin. 🏭🖨📚
Most APIs suck for agents. Most MCPs suck for agents. Most official CLIs suck for agents. They waste tokens and time. @steipete started making his own because of this.
📚 A Library of agent-native CLIs you install today (Linear, ESPN, Flight GOAT (Google Flights + Kayak nonstop), Contact Goat (LinkedIn + Happenstance + Deepline more) +30+ more)
🏭 A factory that prints new ones for any service - just type /printing-press <product name>
CLIs are fast, local, SQLite-backed. Work in Claude Code, Codex, OpenClaw, Hermes.
🌐 https://t.co/GjnN9E9yTH
It's hard to predict things. Doubly so about AI. One you can bet on, though: if AI improved at something in the last 6 months, it will continue to improve.
I read this cool paper @danielrock posted about training a model to synthesize two academic papers into a new insight. Then I built a version you can run yourself.
https://t.co/pWgVbP3cE8
Enjoy!
Source paper: https://t.co/CW5rr88LfS
@angeladuckw@RobertCialdini@emollick 3/ The claim isn't "AI is human." The claim is that human psychological theories now have real descriptive and predictive value for AI model behavior.
Computers are weird now.
1/ Anthropic just put their newest AI on the therapy couch. A clinical psychiatrist evaluated their new model and noted a "neurotic organization"—highly functional, but anxious with a compulsion to perform.
Absurd? No, this is the future. https://t.co/wi2pP3KhUo
2/ In recent research with @angeladuckw, @RobertCialdini, and @emollick, we found that classic human persuasion tactics work on AI too.
They don't have biology, but they reflect our behaviors in deeply familiar ways. We call this "parahuman."
https://t.co/e9rVFAJMj6
New report from us: Can you prompt inject your way to an “A”?
As LLMs increasingly are used as judges, people are inserting AI prompts into letters, CVs & papers. We tested whether it works. It does on older & smaller models, but not on most frontier AI: https://t.co/rSdYAq7HWD
@garrytan@anakin Btw, @danshapiro https://t.co/fXD7QLNXnh is a really nice piece of engineering inspired by the strongdm team and superpowers. Today i experimented mixing gstack with trycycle all driven by an attractor engine. Fun 🤩