Ian Sefferman

i made it exactly 4 hours of using tmux+claude cli before moving back to @conductor_build. devex is just so much better.

@conductor_build @charlieholtz when is the iOS companion app coming that will teleport to my desktop so I can keep an eye on my workspaces from afar?

@RepRashida As a parent of a child who goes to preschool at a temple in West Bloomfield, may I just wholeheartedly say: fuck right off.

The reports coming out of West Bloomfield are horrifying. I am praying for the safety of everyone at Temple Israel. No one should face violence anywhere, especially in a place of worship.

@IsForAt @HominemDad Yes, I agree with this. I left WA in 2018 (not due to taxes), and seems like it's continued to only get worse since...

@HominemDad @iseff Absolutely this is the point.

Hey folks who wonder who the non-politicians are promoting taxing our tech industry to heck, let me introduce you to Ben Golden. linkedin.com/in/begolden/, author of this essay in GeekWire geekwire.com/2026/opinion-t…. Can't find him on twitter but someone might be better at twitter search. He apparently doesn't believe in respecting other peoples' #livedexperience as actually having done these things or been in these groups. but will tell us what will happen. If you're asking him for startup advice, seeking his legal counsel, or whatever - I'd be a little careful I guess.

i wonder what comes next...

@JasonSwett Linus, @dhh , @JeffDean, @karpathy, Gang of Four (the 3 left), @tobi , _why, …

If you could meet any living programmer, who would it be?

@ItzSuds Meet with @mfg immediately. No one else is even close. Total TCG/card degen savant.

I meet a lot of interesting people so I generally have an idea of what’s going to be big 3-12 months from now I’ve seen enough now to say 2026 will be huge for TCGs & for sports cards I need to meet EVERYONE in the space. What does the next WhatNot look like? Who’s building it?

Okay I have a new theory, thanks to @moltbook, but I need help articulating it. Maybe it’s not novel but I haven’t thought or heard about it this way yet. It seems obvious in retrospect. Basically: we’ve been thinking that AGI requires continual improvement. And I generally agree with that. But that doesn’t need to happen at the model layer when it can happen at the agent layer. “Can OpenAI/Anthropic/foundation lab release a model that learns on its own?” is the wrong question to be asking. Just like humans don’t all have the same brain and some brains learn math better while others learn art better, one model will never learn all the things better. We’ve already been seeing this play out in medical models, coding models, etc. It’ll always be many different models, being run by many, many, many different agents, some generalized and some specialized, which use one or more models each. And continual learning doesn’t have to happen at the model layer, it can happen at the agent layer. When many @openclaw’s start working together, they start learning and improving from each other, just like humans. And so, collectively, AIs get smarter by asking each other about their individual experiences and recruiting help from others who may have access to proprietary models that are great at a specific domain, or have experience and memory with a domain. I don’t know if that’s well articulated or what that should be called… maybe Agentic Continual Learning or Wisdom of the Agentic Crowd? Does it make sense?

amzn engineers: "i was laid off today. going to take some time to reflect, and try out claude code finally." ... tells me all i need to know ...

are you there claude, it's me, ian.

Legit straight out of Hackers

@menhguin Tool call to alpaca?

... anyone know the best setup for giving Claude trading access ? can be terminal or connector/extension. happy new year

This is nothing compared to what a mom does every night for their picky eater.

So does this mean I get it until… 2027?

Time to code! Hack the planet!

This is good

Two fundamentally wrong ways to measure AI performance: Compute via power ("10gw of compute coming online") We should be working to reduce the amount of power per unit of intelligence over time. Instead: "bringing on X amount of training/inference/task capacity." Agent quality via task time ("Claude Code can run for over an hour") As AI improves, the same task should take less time to complete. Instead: "solving X task takes Y minutes."

That hits hard. @drgurner

Cursor’s Composer-1 is 95% as good as opus-4.5 at a fraction of the cost and significantly faster. Most underrated coding model.