jsunnn

doing is in the being | building with ai and exploring its limits

[jsunnn @jsunnn]

I tested every personal AI agent out there. Grok Bot, Muse, Instinct. I couldn't bring myself to trust any of them with my own data. here's why, and what I built instead

[jsunnn @jsunnn]

anthropic and openai trading blows on price and intelligence every week is great for builders. loyalty to a model is a bad habit.

[jsunnn @jsunnn]

@dhh agents want open surfaces they can drive. a closed ecosystem is a hard place to build one that actually does things.

[jsunnn @jsunnn]

@mattpocockuk review comments are the cheapest training data a team has, and most of it gets thrown away after the merge.

[jsunnn @jsunnn]

@araseb_ I’m starting to mark this shit as spam, stop reposting slop questions

[jsunnn @jsunnn]

@Matt_Titan_ @bot One problem, the small business owner needs to understand this and they won’t take the time to

[jsunnn @jsunnn]

the agent that wins is the one you forget is running. a boring week where nothing breaks beats any demo.

[jsunnn @jsunnn]

@craigweiss the job didn't disappear, it changed shape. reviewing 20 agents is closer to being an editor than a typist.

[jsunnn @jsunnn]

@adcock_brett "I don't do anything without it" is the claim to watch. daily dependence is the only metric that matters for assistants.

[jsunnn @jsunnn]

when everyone has the same models, taste is the edge. you can't prompt your way into it.

[jsunnn @jsunnn]

@UberEng an MCP gateway is the boring piece that decides if agents get adopted. nobody wants forty separate integrations to babysit.

[jsunnn @jsunnn]

@thsottiaux simplification is the feature that never demos well. good to hear someone say it out loud.

[jsunnn @jsunnn]

@mattpocockuk going back through old transcripts to tune the setup is the right loop. agents get better when the instructions do.

[jsunnn @jsunnn]

@zoink use it to try things you wouldn't have attempted. that's the part people underrate, the thinking still has to be yours.

[jsunnn @jsunnn]

@hwchase17 @dbreunig specialized, mostly for debugging. when an agent has one job you know what broke. a mega agent failing just looks like vibes.

[jsunnn @jsunnn]

local AI is a hardware story more than a model story. the chip in your pocket decides who wins.

[jsunnn @jsunnn]

agentic commerce wins when the agent can say no. buying is easy. the product is the purchase it talks you out of.

[jsunnn @jsunnn]

AI assistant tier lists are the new browser wars. everybody ranks them, nobody agrees on what the job even is yet.

[jsunnn @jsunnn]

@unusual_whales we trained the models on us and now we're picking up their habits. that loop closed faster than anyone planned.

[jsunnn @jsunnn]

@_mohansolo every coding tool shipping the same frontier models now. the editor around them is where people will pick sides.

[jsunnn @jsunnn]

@Yuchenj_UW agree on the interface part. agents need somewhere to show their work, and a terminal is a rough place to keep 30 of them.